Carburetor Rebuild Basics (Part 1) -EricTheCarGuy

greetings viewers eric the car guy here back again with another repair video for you yes it is a repair video this time my 1979 Ford Fairmont straight-6 200 cubic inch displacement 3.3 litres runs great runs like a champ except for when I got to start it in the mornings there’s a lot of start stalls start stall kind of stuff the other day I’m driving and I kind of smell gas and I’m wondering what the heck is going on so I popped the hood and lifted the air cleaner only to find that the fuel filter is leaking a bit of fuel under the intake manifold kind of dangerous and not really good for the gas mileage I could remedy these problems just by replacing the fuel filter and doing probably just a minor choke adjustment however where’s the fun in that because I have always wanted to do a carburetor rebuild video for the longest time and this presents an ideal opportunity because this is one of the simpler carburetors it’s just a one barrel Motorcraft I believe made by Holley that we can go through that shows all the basics of what you’d find normally in a carburetor and in a carbureted rebuild situation like I said I’m going to take this opportunity well rebuild this carburetor I’m going to do in a separate video the actual setup I’m doing that because there are times where as I said you don’t necessarily have to rebuild a carburetor all you need to do is a few tweaks a few adjustments so I’m going to make that into a separate video in and of itself but this video is going to involve the actual rebuild and then the next video following this one will be the one on adjusting it and I’ll put a link in the description to that when it becomes available without further ado let’s get into this let’s get this carburetor off this car next it’s time to actually remove the carburetor from the engine this can be sort of a mixed bag some of these are going to be relatively easy like what I have here and some more difficult some will have many many many vacuum lines and some of them not so much a lot of times I find even with the vehicles with lots of vacuum lines that when you just let the vacuum line lay where it’s where it was hooked up before many times you can identify where that vacuum line is supposed to be placed now if that’s not the case and you want an extra level of insurance just take a piece of masking tape or just some piece of tape or even label makers I’ve seen people use those wrap that around the vacuum line and also you know some sort of identifying marking on the port itself I know some vehicles might have on the under hood stick or a vacuum diagram to show you where all the vacuum lines go you might use that as a guide or perhaps a service manual if you get lost or if you’re in a situation where maybe the carburetor was already awesome so that that can present a challenge there are not that many vacuum lines on this carburetor it’s not that difficult to see where everything goes so I’m probably not going to be marking anything another thing to watch out for is the vacuum lines that go from one part of the carburetor to another just leave those on it’s one less thing for you to lose and you can deal with that on the bench as opposed to inside the vehicle but when you go to take it off just work carefully and methodically try to mark things as you take it off that way you can get everything back where it was when you took it apart you know come to think of it you can also just take a picture of everything before you take it apart and that can also help you with reassembling use line wrenches whenever removing lines that are like this to avoid stripping things and here’s my prize now know the carburetors still have gas in them even after you remove them so don’t tip it upside down or else you’ll get a bunch of gas everywhere but now let’s take this over the bench we’ll disassemble it and get it ready for cleaning now that we have this off the car and hope you remember to mention that the fuel filter was actually what I found that was leaking gasoline so it was just for whatever reason around the seam it almost looks a little bit bent so that’s the cause of the fuel leak I actually like rebuilding carburetors and doing this kind of work I think it’s fun so I bought myself a kit they’re not that expensive I think the kit for this was somewhere around 20 dollars I also got a new float for it and the float was also about 20 dollars I’ve probably got about $40 into this rebuild also really important is this tag hopefully your carburetor still has one of these this identifies the type of carburetor you have in fact that’s the carburetor number right there that’s how I was able to find what kit I

needed so on your carburetor somewhere there’s going to be a at AG there should be a tag that identifies what type of carburetor it is so that you can purchase the correct kit I’m going to start my disassembly with some of the stuff on the outside I’m just going to take that guy off because we don’t want to get him in our carburetor depth and I like to keep all my screws and parts and everything organized so I will keep all this stuff together in other words the screws everything I’m just going to keep it all as one piece and another thing work in a cleaner area these are a lot of small parts you can very easily lose stuff this is the choke assembly that I’m going to remove now we don’t want that to go in our chemical dip either and with some of these screws if they round off on you you can use a pair of vise grips to get them off should we lift this up off of here and this is a thermostatic choke the way this works is this is called it’s a bimetallic material means there’s two types of metal each type of metal has a different expansion rate when exposed to heat and cold because of that one will pull on the other during a given temperature so that will cause this choke to coil or uncoil depending upon how hot or cold it is and you can see that it’s directly connected to the choke plate here so as the choke moves it’s moved by that spring right there it’s very important for the operation of the carburetor to work correctly is to have this so we need to make sure that we get this set up correctly when we’re done but I also want to before I get it further apart just sort of play with it a little bit and make sure that it it can move nicely and over here are your fast idle steps so as the idle gets faster until the point where it’s completely down to base idle so that’s what all these little steps are for like when you first start it up it’s also connected to the choke you see the more closed the plate is the further down this will go of that you have different levels of idols so you need to set this screw this is the fast idle screw when the engine is cold you need to set this up at that time which I believe mine is just a little bit too high I’d like to see mine come down just a slight amount but I also believe that this choke is warmed by exhaust gas so they do a lot of things to make your choke work faster to get your engine to a place to where it’s more efficient but I believe this is run some exhaust gas up into here to warm this up see here’s our nice little gasket a little one-eyed smiley face and I’m going to keep all these things grouped together three main parts we’ve got our top cover here we’ve got our main body and then we’ve got the throttle assembly down here at the bottom sometimes screws will go all the way down through and hold the entire assembly together so you will often find different length screws when you do this the things that you want to watch out for like there’s that gas I spoke of I don’t see anything sometimes there’s throttle linkage that’s connected to the upper levels here like this one for the choke is but I don’t think that’s going to present a problem I’m also going to remove this fuel filter that I’m going to replace anyway it all wants to come out as one piece I anticipated that and it’s bent that might explain why it’s leaking but we’ve got a new one so we can just toss this one this whole cover is rotating down here so the seal is gone here’s one of those this is for the accelerator pump linkage right here here’s a piece of linkage that we’re going to have to watch out for it looks like we can just remove this screw on the back here and possibly just pull that out of there not sure carburetors are neat I always thought they were yeah I don’t think that’s going to help us any as far as removing this piece of linkage that’s going to happen when I actually get the upper assembly off of here it’s probably how that’s going to go down I’m just curious what’s underneath this I’m doing this completely blind I’ve never rebuilt this carburetor before but I’ve rebuilt other carburetors so I’m hoping that my skills will take me through we got here that is I believe some form of emissions deal here could be looks adjustable I honestly don’t know what that is we kind of got to get most of this torn

down because I’m going to put it into a chemical dip that helps to clean it yeah that’s just the retaining screw as all that is hmm interesting a little rubber diaphragm in there keeping everything together not a bad idea to maybe take pictures as you go to remember where everything connects also you may also find that helpful I’m just going to unhook this now that it’s off of here and away we go and this to test just push it down put my finger on the end of it if it stays closed like that it’s good watch what happens when I release so I know that diaphragm is good let’s help I remember how all this goes back together I’m just going to start taking the screws off on the top and this is what I also often do I will lay the screws out in a pattern that tells me how it all came apart that way I kind of get them in the same place you can use like old egg cartons or something too and you can get crazy label everything and all that you can do all that I’m not doing all that I may be burning myself as a result but you know where’s life without the risk there’s our long one we remember that this one’s up by the choke but the way I’ve got these laid out here is pretty much how their they were on the carburetor and as long as I don’t bump the table or move things around I should be good I think we might be able to top of this off now and we can and we’ve got this little piece of linkage back here that we got to deal with so once I lift this up off I’m going to lift as straight as possible there’s also some choke linkage back here that’s hanging up I think what I’ll do is I’ll do this side since it’s easier let me show you how these clips work sort of get in under them and twist work them out like that you get your clip and then you can pull this off I hope there we go now I can get that out of the way just going to put that with its c-clip so I remember how it came apart now that that’s undone I can lift up over here it’s accelerator pump there we go that was kind of a pain in the butt this is the accelerator pump every time you step on the accelerator it forces down into this area and squirts a given amount of fuel into the engine to help rich in the mixture so that you can accelerate we’re going to be replacing that there’s a new one that came in the kit actually they just gave us the cup so we’ll just be replacing that part of it remove any gasket material before you put it in any kind of chemical dip that’s what you’re doing and I strongly recommend that you do because it does clean things up very nicely and try not to damage the gasket and the reason I say that is because many times carburetor kits come with more than one set of gaskets and it’s good to know the gasket that came off so I always try to save my gaskets just so that I can match them up geez tiny holes or everything to a carburetor this is the float think of a toilet bowl your toilet bowl fills up and shuts off this works the exact same way so your carburetor is like a toilet this is the fuel Inlet here remember where the fuel filter was so that’s where the fuel comes in so as the float comes up it shuts off the fuel coming in but as the fuel gets used float level goes down allows more fuel to come in it’s a fairly beautiful easy system always like carburetors and their engineering and efficiency and how they’re put together I just have that little retaining clip

it looks like set that guy over there and it looks like it’s like this is going to come out as one piece they may have a new one of these in the kit because it doesn’t look like any of these parts are serviceable many times what you’ll find is this is a serviceable part and that there’s a little sort of a needle valve in here that opens and closes and allows the fuel to come in and looks like I can just lift the float right out of here now there’s our float assembly now I ordered a new float just for good measure over time these can get heavy with fuel so they’ll sit down a little bit lower and there there is going to be a point where we’re going to have to adjust the float level float levels critical in the carburetor so you want to make sure that it’s correct so if these get heavy and and lay down inside the bowl a little bit further it’ll be letting in a little bit too much fuel so we want to keep our float level as close to correct as possible now these are our neoprene floats they also offer these in brass which are sealed and those never go bad because like they never get a chance to soak anything however they can get like maybe a little pinhole or something and they can get gas in there I suppose they can also fail but they don’t seem to fail as much as something like this might and we’re down to just about nothing now all these other components can sort of stay in here it looks actually fairly clean I am going to find out what’s behind here and yes I can’t remove a throttle body I’m also going to remove the throttle body looks like we’ve got three screws for that yeah that was simple enough the carburetors are when they say one barrel that’s what it means there’s only one hole one throttle plate two barrels got two of these four has four but that’s all it is but that is when that’s wide open as much era can go to the engine is going into the engine at that time but this is directly connected to the gas pedal so when you step on the gas it opens and closes this and in essence throttles the engine something you need to watch on throttle plates is sometimes these shafts will get worn out and as a result it will cause a vacuum leak right here so if you’re spraying around the carburetor you just rebuilt it you know you had a vacuum leak and you’re trying to address that the particular problem on Quadra jets so look for that one last thing I’m going to remove here is the idle mixture screw this guy is super important to the base idle mixture and I want to make sure that this area is really clean because this is this is the idle this really helps the idle we’ll show that in the adjustment video how we make that all happen but it should have a pointed end on it yeah and that pointed end allows air to escape past the throttle body right there and you control the amount of fuel that goes into the engine at idle through this group there’s usually more than one if there’s more than one throttle plate but since there’s only one here there’s only one throttle plate but once again I’m going to save the gasket that’s completely torn down and ready to be cleaned this has a couple more things I want to see just what’s behind here and what I’m looking for like rubber parts and I want to make sure that I get my cleaner into all the areas that I can and replace any gaskets that I can because really that’s kind of what it’s all about it’s it’s sort of resealing everything and making sure that it’s all in a good and happy place I sure hope I’ve got that gasket it’s a valve of some type what it actually controls I am Not sure it’s bimetallic it looks like a valve that opens and closes like this you can sort of see it opening and closing like that assuming there’s some sort of port here it looks like it’s got some exhaust gas on it so looks like another gasket down in here almost another thing to watch out for is tiny little balls I know that’s funny yes you must be aware of tiny little balls sometimes there are Czech balls inside of the Assembly they’re kind of critical to the operation of the carburetor so you don’t want to lose those and so far I don’t see any that have fallen out so I’m really surprised because many times as I said there might be some engine but I don’t see any

that’s about as far as I’m going to tear this thing down and we’re just going to soak this guy in some solution this upper part also much of this as we can yeah I wanted to try to get that rubber piece out of there alright just to get all my bits out I’m going to remove this fastener for the accelerator pump hopefully I’ll be able to take the assembly out as you can see it’s on there sort of spring-loaded sort of needed to twist it around to get it off of there but that’s how it how it went together once again I want to try to keep all my stuff in place and this well probably one of those things that only comes apart one way and it does keep that also with that and I’ve got my fingers crossed that I get a new one of these you never know sometimes the carburetor kit so try to try to keep it yeah wind even got a little bit of tear in it so hopefully get a replacement for that otherwise I’m not going to bother taking the shaft for the choke or anything out I just I don’t see the point well let’s see if these items actually fit into my container I have chosen an empty paint can I’ll see if these parts all fit down in here it just fits in my little bucket this is my caustic carburetor solution you might send yours off to have a dip soap whatever but this this is a special solution just for doing this kind of stuff it’s well let’s put it this way it’s eaten through two metal cans that I’ve kept it in over the years so be careful with it but it’s really good at cleaning all this stuff up and I probably could have done a better job of getting some of the gasket material off of there but you know I can do that after the fact cuz it’ll be all softened up and loose that submerged most of it enough of it where I’m happy right well we’ve let our parts sit in the caustic solution for a little bit there’s still some stuff on here I’m just going to go in here with a wire brush and clean them up a little bit to get them a little bit cleaner than what you see here and get rid of some of this excess gasket material this is kind of why I waited because this is really good at dissolving this stuff and you just a little rinse so I’m going to go back through and do a little bit of agitation a little bit of a little bit more work to get these parts a little bit cleaner and we’ll jump back in when everything is all cleaned up now that we have all our parts relatively clean I’m going to say pretty clean I’m going to dry them off with some compressed air and then we’ll get into the reassembly right well we have cleaned all of our parts everything’s disassembled we have our carburetor kit and we also have a new float one more thing I want to mention about these throttle bodies I know I mentioned that sometimes the shaft itself will get loose inside the board try and feel for that feel if there’s any play down here any back-and-forth or any up-and-down in this because if there is you may have to replace this part because a throttle operation can affect everything as far as the rest of the carburetor goes so it’s a good idea to make sure you check that out thoroughly before you dive into your reassembly but I believe all my parts are in order I have all those parts there and I’m ready ready to do it and fortunately I haven’t seen a carburetor kit yet not come with a set of instructions so if you if you get into a jam let’s say and you don’t know where something goes often time like building a model car or doing Legos they give you some pretty good instructions for how this whole thing here’s an exploded view of everything and also an explanation of all the stuff and what it is and what it does Oh so let’s see if we can find out what that piece was that I was unfamiliar with it’s a hot idle valve okay so it is a thermal valve but it’s there to I guess put a little more air in to keep you from running rich this guy here’s the choke pull off looks like we’re going to start with the float level adjustment okay with

reassembly I am going to start with the float but something I’ll just cover briefly here it appears that on the exploded view you come over here real quick there is parts 48 and 49 right there and what those are are the check valve for the accelerator pump now when I disassembled this and I’ll look over the footage again after I do my edits I didn’t see anything come out I didn’t see a ball or anything in fact I even mentioned how it seemed odd that there wasn’t any check balls or anything in here when I did the disassembly because that’s not normally something you have to watch out for because the check but the the accelerator pump does use a check valve in that fashion so what I’ve done is the carburetor kit has come with a new bulb the accelerator pump so I have that that actually goes into here this is this is the accelerator pump chamber here and when it fills up with fuel this accelerator pump pushes down when you step on the gas forces fuel up through this passage and out this tiny hole into the venturi assembly here but you don’t want it to bleed back into this assembly so to take care of that you have this check ball which will go in here and put this in towards the end and then I’ve made sort of a counterweight of an old bolt I’ve basically shaved it down to the correct size a little bit of machining on it I looked at the drawing and I’ve got the proportions I believe similar to what we have here so just put that down in there and from what I can hear it’s moving freely which is what it needs to do it needs to be able to move up and down freely but that’s that’s my fix for that this thing may not have had a check valve at all when I took it apart so I’m gonna put these in a safe place here and put them inside this plastic bag so they don’t go anywhere but as I said I’m going to start with the float adjustment which is one of the first things to do during reassembly here’s when you float put it next to the old float so we need to swap this over from one float to the other for the new one this little tab here is what you bend to get things correct but in addition to that I’ve also got to put in the inlet valve but the inlet valve on this some of them are different this one’s all one assembly here’s the old one this is the fuel that comes in so this is like where the water from the toilet bowl for the toilet bowl would come into the toilet and this meters the incoming fuel based on this little tab right here touching the end of the valve so as the float goes down it opens up Laos fuel come in as a float goes up it shuts it off quite simply that’s how that works let’s see before I put the float in after I put the float in it’s going to drop it in then I believe there was like a little spring that went over the top of it I think I’m also going to put this in here at the same time there is a gasket I’m not mistake there it is there’s this gasket that needs to go around the outside yeah this way we know how far it’s going to to drop snug that up just a little that’s good with carburetors are usually made out of a softer metal so be really conscious of that so that you don’t over torque things usually in our instructions Yeah right here we have our float adjustment and it looks like it should be bang on even with the top of this float Bowl it will put my spring back in here’s that sort of arched spring that seems to hold everything into place all right that keeps it from going anywhere and then we tip it upside down and it looks like we’re hanging down a bit too far you can see that there because what we need to do is we need to back it up to where it’s hanging down like that on that front part so we need to we need to bend that tab a little bit so that when this comes all the way up that’s where it stops so because that hangs down too far I want to bend that tab back a little bit like this I just want to bend it back towards this put the retainer back in flip it again

and as you can see I’m getting closer I’ve got to bend it back let’s hanging down a bit too far before but I’m going to be bending it back up or back down cuz I went just a wee bit too far almost there so close a few millimeters more I think that nailed it that’s about as close as I think I can get so float levels set sometimes the kit will come with like a little ruler or something there now we’ve got the float taken care of that’s assembled

【Lecture 04】 CS246, Mining Massive Data Sets

All right, welcome everyone to the first lecture of Cs two, four, six I am yukari, I work with you here I’m going to be your core instructor throughout the quarter If we were wondering if I was a bath on Piazza, just replying to her answer as I exist as a person so you are going to be seeing me around until the end of the quarter, and before we started this lecture just let me give you two important announcements Her as one is right after the lecture You just have ten minutes to run two gates, Beichuan, we’re going to have a refresher on linear algebra is going to be extremely useful for your homework and the final exam so recommended to be there We already put the handouts on the website, and we’re also going to be putting the video line by tonight or tomorrow So don’t miss their professional in your algebra Second thing next week is going to be quite intense for your guys, so we have a bunch of their lines We’re going to be having a more zero or more Kwan A scene on Grace Court Many of you have sent me to the radio moral zero, so good job Make sure that you’re working fast on American because his way longer than a more zeros you might have realized we’re also going to be having two creases that lines on gradient So we decided to postpone the first one costs Some of you just realize that you have to work on quizzes or so you are still deciding If they won a dropout or not, so we made sure that they would be the first one would be postponed one week you can use as a reminder you can use the late period for the homework You cannot use late periods for the quizzes, so by Thursday at midnight they have to be submitted Ok, all right, so let’s start to this lecture topic is going to be the fury or locality sensitive machines On Tuesday, Euro gave you an overview of our message, and today we’re going to try to understand exactly why he works in the way we want, and also we’re going to try to generalize it to different similarity functions Ok, so let’s first of all, let me give a recap so we’re sure everyone is on the same page on a message and then we can build on top of those foundations So what is the task? What do we care about? Basically, when, whenever we were given a large collection of documents, I said when the order of is in the millions are in the billions and we want to find similar ones, we want to find near duplicates This is an extremely important task if there are search engines The winery index again again pages are similar It’s an interesting task if you are running something like I see a cybercrime showing you want to match the fingerprints, so that’s another very common scenario where you are going to do well, a stage or anything very busy, one to find singularity in very large datasets Why is this a challenging problem? Because if you wanted to pair wise comparisons, you just have too many pairs These are quadratic problem, and here in Italy, whenever you want to make something scale, we hate when the complexities quadratic We wanted to be linear or even better sublinear at times So basically what we do with our messages we bring a quadratic problem into linear problem I was okay, alright So what’s the solution that we are going to apply? What we want to do is we want to ask these documents so that the similar documents and ducked into the same bucket and Ashley is a very fast operation We can do this on each single document, so it’s a linear one, and then once the action has been performed, why want to find is we want to just work on some of the candidate pairs that we were able to identify and we’re gonna perform the expensive similarity comparisons only on those pairs OK That’s how we bring down the computational cost of running similarity on such a large dataset How do we do this? So it’s a straightforward pipeline The first step is, we do this shingling representation of our documents, and I’m going to give you a while I reminder of how it works But basically, the output of the first step of our pipeline is a set of strings of landscapes that are appearing recurrently in the document Once we have this representation, we want to make it even more compact, so we are going to be performing me, gnashing there and the output of manipulating you’re just a set of signature scanning from the document, and this signature, in the case of rushing or just short integer vectors and the important property of mean rushing is that the signature that is how putting are maintaining our reflecting the similarity between the documents So we have a more compact representation, which still retains the similarity property among the signature, and the last step is we are going to be performing at the stage So I’m going to be reminding you today about roars and bands and so forth, and the goal here is that we want to generate only few candidate pears, And these are the kind of repairs where we are going to perform the similarity check OK, so let’s go one step at a time and have a refresher on each element of this pipeline so first one shingles So we said that a case shingles, sometimes also called kilogram,

is a sequence of cake tokens that appear in the document I recall the question there we go lots on Tuesday What is the granularity of K of what this integral right here, these tokens, they could be characters, they could be words they could be sentences It really depends on What is your task if are performing? Spell checking than you want to find common K shingles In terms of characters, if you are just trying to find similarity between documents than usually your tokens are gonna be warns, OK, So the granularity depends on your domain application Now let’s let’s have an example, so we said, OK, you’ve got to do so we’re going to have to shingles This is our document and the output set of two shingles that we obtain is just a BBC and CNN OK, so we’re going to do for this thing to only keep the shingles that are unique, and this is the way in which we obtain our representation Basically, it’s a set of our own values hobbyists case shingles Now you will realise that hope to hear these parts of our pipeline is the main dependent because now we are working with documents From now on, whenever we were going to fly down here doesn’t depend on what is your input data OK, because at this point that we reach a representation that is in form of sets OK, so now, whenever we can turn into a set that can be fed into our similarity measure and one very national similarity measure that we have been using throughout the past lecture is the jaguar similarity very common, used in a lot of context we’re gonna give a reminder So basically, what it does is we have our two documents see one in situ, and what we do is first of all, we take the intersection of the common shingles and then we divided by the Union of all the shingles OK and this intuitively tells you how many shingles two documents having calm OK? This is the similarity measure that we’re going to be using The Dzhokhar similarity Once we have these a representation in place than will move to the next step, which is performing the main hashing and what is the main property that we care about in me, GNASHING, so we want to go from these very large sets of shingle soil You can imagine that in a very large corpora with millions of documents, we will be funding hundreds of thousands of the shingles, depending on the size that we sat and the growth of the token So basically The beehive of these metrics is going to be a very, very big, so we want to convert it into a more compact representation, still preserving the similarity OK, so basically, what we want to do is that the probability that two documents are two sets are going to be having the same hash is equal to their similarity So how do we do this? We take our input matrix columns are the documents roles are going to be our shingles and then we perform our permutations This was the trick that we are applying Tuesday So let me give a reminder, Let’s say ah, why this value here is four, so we’re working with document tree Then we’re working with a psycho Nash, the yellow one What we do is we keep the first instance of a one appearing in our collection of shingles to see wealth number, one which I hear it’s zero than which had a number to zero again, which had a number three yet another zero Then we go to four and we sign a one, and we put here the index of our permutation OK, that’s how we perform this conversion So we are going from an from a wider and more expensive representation into more compact won The beauty or the magic of this tree is the following Is that the similarity of the columns and remember that each column here is representing a document is proportional to the similarity that you’re going to be standing on our input matrix? Okay, so we can go to read just as an example to check if it’s actually working So let’s say we want to compare document One in document too has receded on that anybody is in common I invite you to check the same on the bigger metrics, but say, you want to check one and three So we’re going to be having these two types of values in common, the bottom two in common, so our signature similarity is going to be two thirds While if you go and check on the column one in three, you will find that three fourth of the values are in common Now you might say This is an approximation is not as close as we will wish is absolutely true, and the tweak that we’re going to be seeing today is that we can make the similarities we can make those two values getting closer to each other each other arbitrarily and a tweak to make that happen is to have more hash functions, as it is that today I’m going to teach you how to choose exactly those functions, how to choose the size of the bands, the number of rows and so forth But? As pretty much anything that’s in computer science, what we are going to be doing is a tradeoff The Moorish functions we introduced in the more computations we have to perform

OK, so there is a tradeoff between how much work we are going to be doing and how well these two are mapping to each other OK, so that’s first right After that, I want you to keep in mind for today And last step out, the recoveries were moving to the locality sensitive bashing so we have our singers representational Now we asked the columns of our metric, Sam, and what we’re going to do is that we want to put similar columns into the same buckets and once you have a bucket So keep in mind that this step is basically linear because we are computing a saddled ashes per document We placed them in different markets, and then we are going to be performing the similarity check Call it within a bucket OK, That’s where we go from quadratic tool to a linear complexity So as we were saying last time, rather than just doing the comparison between two documents so two columns, we can subdivide our space into bans OK, so this is our common representation that we’ve seen before and we consider a candidate pair if is matching on two bands only OK, so let’s have an example for me or, for instance, say we have documented seven, I think and document for you see they are rushing into the same bucket, so these two guys are going to be candidates for our similarity check Ah, is who takes six and seven rushing to different buckets, so we are never going to be performing a comparison among them all right? So, ideally, what we would have to happen is the following is what we see here on the X axis we have the similarity, the extra similarity between the two documents and on the y axis we will have the probability of sharing a least one bucket, which means they are gonna be at least matching on one band right So why do we want this beer or say we want all the documents, all the states that have more than zero point six similarity We want to be able to retrieve all of them and be part of the same buckets so that we were going to be comparing to each other But anything as it is below the treasure zero point six We don’t want it to be part of the markets where we were going to be performing the similarity comparisons OK So the whole goal of today is how to turn this plot from a line into a nice curve that is as close as possible to our traditional line All right, and I’m going to be explaining the tradeoffs of what happens when we introduce false positives and false negatives I’m going to be repeating this a couple of times so let’s start from now What do you think is more worrisome for this kind of task? Having false positives or false negatives and also do is not strictly correct answer so if anyone wants to give it a shot, I’m happy to talk about this for a minute Anyone wants to try Yeah, why human intuition why? I think it’s called? Exactly so, whenever we introduce false positives in the second step, we’re actually computing the similarities You will be able to discard them eventually because they just come at the price of performing more competition all right while if it’s true, if we are false negatives in our pipeline, what actually happens is that we are never going to consider them as candidates So this is the data that we are going to be missing And so my goal for the lecture today is basically to give you one of those eureka moments where you know that it was worth together, master or underground, in computer science and a rather than going to a stack overflow and copy pasting codes I want to give a reason why is worth to study so much math for so many years, because then you will be mastering this, this tool, this technique, and we are going to be able to understand why you’re choosing certain parameters OK All right, so let’s go one step at a time and reach that And the other goal for today is to generalize mean gnashing and make sure that Alice age is not going to be used only for documents where we can perform these shingles extraction But it can also be applied to basically by data points, all right and is extremely important because, as you know, data comes in many different forms and shapes or rhyme, especially, you have seen all the craze that we have today, with deep learning, the fact that we can do great with images and multimedia data and so forth son data cannot always be represented The shingles and this technique is an equally powerful also on different type of data so we’re going to try to generalize it in that way All right So what do we need to do to make that happen? Is we have to design a locality of sensitive as function that is going to be working for a specific distance metric right? So this has magic that we’re missing last Tuesday and up to now is our jacquard singularity,

which we’re going to see later, how it can be converted into a distance But we want something that can be applied to different distance magics OK, we are going to be seeing euclidean, kazan’s similarity and so forth Then everything else stays the same We a pair or sing as her has short, introducing issues that are reflected in the point similarity And then we are going to be applying our elder sage technique, doing some of the magic’s with a band and role cells we’re going to be seeing later, and eventually at the end, we began our set of candidate pairs, where we’re going to be performing the similarity check all right OK, any questions up to now, or shall we dig into the technical part? All right, as it is So as I I spoiled my lecture at the beginning, so I told you that most of the magic that we’re going to be seeing today lies in this s curve OK, that’s where the magic happens And why is this the case? so say, we selected only one single ash function Okay, that is able to retain the similarity between the documents What we will be seeing is exactly that a straight line, as in this plot So again we have X axis The similarity to the two sets Y axis the probability that our list sharing one bucket So what happens here is if you have only one ash that is able to represent the similarity between these two documents is going to be exactly proportional, with actual similarity between them OK, so doesn’t really make us any service, because we are not going to be able to filter out the bad candidate pairs we just want to try So the trick is where we want to introduce more ashes, and we want to end up having a shape of this form So whenever our similarities below the treasure, We don’t want any chance of those candidates appearing in our in our buckets to the protein, the similarities above the treasure that we want the probability to be as close as possible to one So ultimately, we want to turn that line into step function And everything that we are gonna be seeing today is an expectation like many other algorithms to make your pipeline scale with large data you can really expect to work with billions of elements and making sure that you’re going to be doing a deterministic, perfect job were on the percent of the time Okay, So what we want to do today is to reach the status where we can do very well one to nine per cent of the times and that’s really great result question forecasters said the Y axis for that is technically sharing one bucket, so this one is just a bit in that way Yeah, yeah, so it’s a toy example is just to show you the two extremes that is useless This one will never happen We are going to try to approximate this step function as much as we can Thanks for a passing grade Okay, so how do we shape in the way we like our s core, So we have two tunable two parameters we have the number of bands in which we’re going to subdivide our input matrix, and then we have the number of roles which are gonna be composing Each single bank kept so These are the two parameters that we play with So let’s start with having our similarity between our two sets equal to us, and what is the probability of at least one band is equal, so we’re going to be basically getting eventually I’m going to be able to give you a formula on how to compute exactly in the probability, so we pick some band and then the probability that the elements in a single row of these two columns are equal is equal to s OK now if you want, that is that all the rows in a band or equal we need to take has to the power of our cat, so we have our rows so that’s what we get with the probability and if you only started the probability that some role in a band is not equal, That we’re going to be taking the inverse of that so one minus X to the power of art, the bear with me who are going to another couple of steps And then we get into the fun of formula Now the probability that old bands are not equal, it’s going to be a falling We are going to be getting this to the power of big, so the number of bands because we want all of them not to be equal, and ultimately, what we care about is the following results, so we want the probability of at least one band is equal OK, remember, this was the condition that told us these two pairs are a good candidate to be checked Okay, only when they match a list on one band So does the probability that we just found which is going to be one minus this times this step before the cool thing is that this is the probability that she wants it to the candidate pear and we can simply plot it so that’s what we’re gonna be doing in their necks, Chinese lives or so you can plug this curve and then find the sweet spot that exactly matches your use case

OK So here is one example first say that we want to pick our R and B to create the best possible escort for our scenario We want to offer the budget of computing fifty hash functions All right, so this is one possible configuration We can have our equal to five so five rows per band and ten bands in total, but as you can imagine, there are way more are combinations that you can find here So this is the s curve video obtained as an output again, X axis, Oxford Similarity Wax is probably the thing that are sharing a least one bucket and therefore we’re going to be computing the similarity between them and our threshold in this case is zero point six What do you think about this headscarf? Do we like it? Is there something that we dislike or we could do better? Exactly so, this is the air that we don’t like right, OK, so how could we still use this s curve in other contexts, where it will still make sense? Right, what can we do here Exactly so? we can take the treasure from zero point six Let’s say we move it to zero point that we reach this intersection, which is gone and probably nine nine, nine, nine, nine, nine, nine, nine, nine, nine, nine percent And this s curve is gonna be great when you want to find all the candidates that similarity of zero point eight or above all, right so that’s why I was saying to you before our and be our tunable You change them in such a way that once you know what is the treasure that you care about, you shape the s curve to make it sure that he will not return to many false positives But you’re still OK with the false negatives Sorry, the other way around right Yes, I make them okay Here I just showed you the answer that we were talking about before So with a treasure of zero point eight, these are scoreboard in a sudden becomes a very good feet So how can we shape our cordula’s? Let’s fix one parameter change the other, and we’re going to see how the s curve is modified Thanks to that OK, so first of all, we fix our threshold, so we can take our zero point six from before and then What we want to ask him until you make happen is that we choose RMB in such a way that the candidate pair as a step right around us Okay So yes, Carbon in a southern as these hawkish RT artistic growth, exactly on the fragile So let’s talk First of all, the top left axes are exactly the same as before acts similarity why the probability of being a candidate pair, so we fix the number of bands equals to one and then we were in the number of rows Okay, so we go from Ah, we go from the straight line with one role to this curve down here When we have ten rows Can you give me the intuition why this is happening? Why do you think that keeping? Why do you think that considering more roles at the same time makes the probability a core of change in that way? But is the intuition behind it? Exactly exactly so, basically, what we are doing is we are considering more rows at the same time, and we’re gonna be funding if their ashes are matching or not If you have only one row at a time, the probability is going to be higher, but as soon as you consider more rows together than when a sudden our condition becomes stricter OK, so that’s why we would see here that you will start to return potential candidates Hollywood have like zero point eight similarity The prominent with a curve that is stretch in this way is that is unfortunately not going to be returning a lot of interesting candidates OK, they are just about the intersection line, so is not enough basically to tool only hard Let’s do another example Here we are fixing are equal to one, and then we’re we’re playing with a number of bands So once again, one one who got a straight line, the more bands we introduce, the more we stretch it on the top left here, we’re doing exactly the opposite of what we did with the roles, The more bands of introduce like the more we subdivide our metrics into bans, the higher is the probability that at least one of them will be matching OK, so one band is he goes to one It means we are we are We are comparing the two columns for each single value The more bands we have the higher is the likelihood that at least one of them will be matching Okay, So we have these two very nice construct, contrasting tradeoffs, and when we start to tune both of them, then we reach this nicely as shaped curves

OK, and as you can see here in these two plots in the riots were fixed in the number of rows is either equal to five or ten, and then we are changing the number of bands and, as you can see, for instance, if cargoes from five to ten, everything is going to be shifted on the right, because the more rows you get, the stricter look on the condition becomes so you will be returning Candidates wholly of the similarity is higher question Oh, yes, yes, so the question is, once we have our own treasure, can we reverse engineer are and be in such a way that C matches, he gives us the best possible escort The answer is yes, I’m not going to be spending too much time today telling you exactly lethal, how to compute it But it cinders lie, that I’m going to be giving you all the intuitions all How that can happen are good libraries for our message Do that for you Likely so? OK, now before I jump into the theory of LS Age, any questions The question is, how do I decide how many hash functions? I want to use right? Oh, yeah, that’s quite so busy you Your question is what if I choose a certain amount of hash functions and all you know I suddenly realized there are not enough to have the performance of the care about, and I will have to recall impute to give another full paths on the data That’s the reason why we’re doing this lecture today, because basically you can tune or you can find out what are the bounds of an R and D, and those two values give you the number of hash functions that you are going to need C you can approximate what you need before you have to give a pass all their ways to make that happen are also whenever we have a very large dataset, you work with samples That’s a very good recipe in January life, you know, if you know that the data distribution is not too skewed, it takes a small sample you work with, that makes some assumption, and then when you give the full path, you’re going to find out If our assumption holds or not That’s a great question Okay, now let’s jump to the two year of our lesage really works So again we wanted we use it We use a message up to now to find similar documents, or in our case we are working with shingles or similar sets Can we done use our lesage for order? This has measures all right, so you’re clear on distances, cruzan distances and so forth So let’s try to generalize what we learn quick refresher what is a decent measure, so it’s a function that given a pair of points X and Y, you will return real numbers with the following properties So the distance the distance is always greater than or equal to zero It’s equal to zero if Annaly Here X and Y are the same element it’s symmetric So the distance between X and Y is the same as why an axe and the triangle inequality, poverty holes So given any other point with other than X and Y, then you know that the distance between X and Y is always less radical than the distance between X and Y and The Z point the turf point, plus the distance between Z and Y point OK So there is a refresher from the past, And how do we turn, for instance, our Dzhokhar similarity that we use for the earth, for documents into a distance metric, as as simple as doing one minus the Joker, Sinjar, as you remember, the jacket, singularity was between zero and one went to Documents are exactly the same even turns one in our case Instead, we want two documents to be the same when the distance is equal to zero, so we just take one minus hijacker Similarity odor distances they were going to see today, for example, are the cosine This means for vectors, where we are analyzing the angle between those two vectors or another, a distance that is are daring to defray one is the nuclear one so usually recall the as to alarm and what we do here is just we take the square root of the sum of the squares of the differences between X and Y, and this is the most common notion of distance is used in many different contexts, also, images and a P and so forth Another interesting one is the L, one arm, which is the sum of the absolute value of the differences in each dimension Sometimes called the Manhattan distance because it’s basically the decency will have to travel in our city that looks like a grid, so one block at a time and you’re going to be making the sum of all the steps you do along the coordinates This works well in the? U S very bad in Europe, as you can imagine, because our cities are not That’s quick Question Ah,

ah, yes, we are gonna be seeing how yeah, he’s going to be a relatively different hash function from whatever we have seen up to now is based on projections But I’ll give you the intuition why than can workers in all dimensions? Great Okay, so Another key concept for today that I really want to be happy in your mind is what are the families of hash functions So let’s say we have our main rushing signatures We we have, uh, minami functions for each possible permutations of the roles as you can remember like this, where all this was the space that will be possible ash functions we could be using for me Lashing OK, so basically as a reminder of the hash function is just any function that tells us whether if two elements are equal or not so what is done, a family or lash functions is any sort of any set of hash functions from which we can pick one at random efficiently, and here are the key word is efficiently Okay, it’s relatively easy to code, something that will give you proper family of as functions, but if it’s extremely expensive to compute that than you are kind of defining the purpose of doing, let’s say it’s OK So we want to be able to create those families very efficiently and so we saw the example for the main ash is all the possible permutations of roles They are very easy to generate We’re going to be seeing similar tricks for the other as functions that we’re going to be seeing today Now let’s define local into sensitive family OK, so suppose we had our space of points as where we define our distant measure Our distance measure, so feminine bash functions is defined in the following way is said to be the one day to prp too sensitive if for any accident while by using us so for every candidate pair that we work on, we’re going to be having the following a part of this So if the distance between X and Y is less than or equal to the one than the probability over all, the possible ashes in our family is going to be at least be one and conversely, distance between the two points larger than detail than the probability is going to be a most people So here are basically defining boundaries right? We’re saying if our two elements are close to each other within a certain treasure than we want the probability of the ashes to be equal to B above a certain probability, because these are the candidates that we care about their close to each other, we won’t ever be want to be as big as possible All right Conversely, If X and Y are far from each other than we want, our probability p two to become as small as possible right and what we’re going to be seeing today is going from the standard means cash that we saw last Tuesday that as not very exciting values for P one and P two, we want to stretch them in such a way that everyone is going to become almost close to one and Peter He’s gonna go straight down to zero OK And that’s why this is the most important week for realization because it really makes the difference on how much computation we have to perform Henry is a visual clue what was going on? First of all, first thing that I want you to pay attention to This is a distance right, so we move from similarity to distance Keep always in mind that when you have juggled, for example, similarity is nothing but one minus the distance OK, and that’s the reason why our s curve now here is flipped Right now Don’t get, I don’t get confused by this Okay, We just mirror the curve and what we want to happen is the following Whenever there is a small distance, we’re going to be having a high probability that it washes are gonna be the same When the distance is I once again We want the probability of hashing to the same value to be very low How do you think we want to squeeze this plot, like we have this for moving dash lines? What would you do with the one in the to, and what would you do with P one and P two? What should happen to them Once you give it a shot? So the peace should be very easy Right, we said, won’t be want us to become as big as possible and be to us to become as well as possible So we want we want to go on the top be due to stretch to the bottom now What do we want between the one and the two? Do we want these to dash lines to get close to each other to get further from each other? Closer, alright, perfect! so why, why do we want to get them closer? Ideally our treasures we lie between the yuan and it all right So we want to make sure that this the shape will look in such a way that everything that as a distance less than the one we have a high probability Harrington,

they will have a distance larger than the two will have a very low probability and that’s exactly what we’re going to be saying okay, So these are distant treasures here, and this is how basically we approximate the beer row as curve are going to be playing with the parameters so that the one in the two are squeezed He wanted People gets four far from each other All right, hold here Great, OK, so let’s apply now this lesson to minish so once again, as space of all the possible sets, these are Jack our distance we really use It is a family menashe functions for all the possible permutations of rows Then Frank Asche function H, belonging to the vocabulary is family We want this property to hold so the probability that they too will be ashes to the same value is going to be one minus the distance we just put We just went from similarity to distance here we haven’t done anything different from what we have seen on Tuesday Ok, perfect So what is our claim? Well, Well, what do we want to? What do we want to analyze from the behaviour of mere lashing? Is that this is a one third tutor tutor one turn sensitive family since the sensitive family for us, our space of options and for the jackal, distance d OK, so this is the behavioral meaning rushing without playing a role with bands and rose And that’s why I was saying before these values are not exciting Once again, it’s way better than doing quadratic comparison OK, Because here what this is this is telling us is the following you are gonna be if the distance is one third, which means if the similarity is two thirds, because we have to take one minus one term than the probability that the candidate will be returned is two thirds So not nothing better than doing a standard as function, so we want to do better than this How do we make it happen? OK And you go faster, otherwise we are out of time So can we then reproduce? Whereas curve that we saw before for any locality of sensitive hashing family OK, so first week that we’re going to apply is how our bans technique OK that we’ve seen also, alas, as Tuesday And we’re going to be having two different constructions, as we said before, one or both of them are tradeoffs, for as one is the end construction Where is we’re going to be working with a number of rows in a band? The second one is going to be called the or construction, which is tuning how many bands we are going to be extracting from four from our input matrix All right, OK, so how do we amplify our hash functions? How do we amplify our s curve? We can play in the in the following way OK, so first of all, we start from our large family or hash function H and then we want We want to get a subset of those with which we call each prime, and here we’re talking about the end constructions OK, So we are trying to reproduce the the effect of the band size, so the more rose we have in a band, the stricter this condition becomes Okay, So essentially, what we want is that all the ash functions between Element X and Y will be matching for all the elements in that ban OK for all the roles they are contained in the inn that ban now in a sudden what we do is we go from this function, sensitive just to be one in Peter to something that is going to be senses beyond two one, to the power of R and people to the power of R OK, so we’re making those probabilities smaller, right And? To make sure that this happens, you have to make assumptions that our hash functions are independent OK, we’re going to be seeing in a while why this is important So Do you think that both these changes are beneficial or one of them is beneficial, and the other one is detrimental So I want to give it a shot Together we make that probably small and that’s the probabilities of their similar arrangements website and then you’re missing, but it’s good for you On the other hand, exactly what you write is exhausting, so we lower the probability for the large distance Remember D two is when two elements are far apart So we were lowering that probability We are doing something good, while at the same time we’re lowering the probability for the small distances, and this is bad right So that’s why none of the two constructions taken in isolation is useful We are going to be having to play with both of them At the same time Now, before you ask me a question of why do we make this assumption that the hash functions are independent?

Let’s jump to this slide You can read more with more attention on your own later But here’s the gist of the idea, especially for me, Nash There will always be some permutations that are very much correlated to each other OK, but we were not talking only about the worst case here, like in general among wholly possible members of our hash functions The average case is not going to be the worst case, so the two hash functions will be pretty much independent So this is another reason why we’re introducing a bill of error into our research and that’s why I told you before this is a problem that we’re studying expectations We don’t expect it to have one hundred percent perfect performance Okay, but this is just to reassure you that the assumption were made for the independence makes everything more tractable today to talk about it, and he doesn’t introduce a lot of additional lives All right so this is for the and construction we jump to the poor construction very similar The only thing that is changing here is that we want our hash functions to be equal for at least one of the elements OK, so for one of the bands, so here, what we’re saying is we take one is as functional per band say our bees equals to tan, for instance, only one of them has to match if they match then we return those two elements as a candidate So once again we took our original function and then we modified it In this way, distance are retained the same, but then we are moving to P one and P two, expressed as one minus One must be one to the power of B, and it follows the same for people You shouldn’t come as a surprise now that we are doing something good and we’re doing something bad who once again, okay So we’re raising the probability for the small distances here seem to the Y because we’re doing one minus a quantity that is getting smaller, and at the same time we were raising the probability for the large distances, which is bad Okay, so once again That’s why we need the two constructions happening at the same time Otherwise we all optimize one aspect of the problem OK Sorry is the intuition the end are making the probity shrink, but which, when we choose our correctly, we can make the lower probability approach zero, while the higher does not OK, so this is the first property that we care about The probability of having candidates among elements is a barrel of similarity Values should be good, so good as close as possible to zero, and that’s what we tool by making are larger and at the same time, as we said before, the your construction stretches these curves on the top left so we are making higher the probability that Whenever the similarity dykstra similarity value is higher than we return it as a candidate, with probability very close to one OK, So we went from the intuition now to the mathematical instrument to make this happen in our s curve expression Any questions look here Cool, great, OK, so so we combine We can combine the two constructions together Ah, I’m going to be going very quickly to to the next slides the day that I want to give you is that they can be composed in many different ways you don’t have to apply Strict refers to the end and then Dior You can apply them as a chain you can do, and or and or the outcome that you care about At the end of the day are tiguan and people OK So let’s say, for instance, here are our composing construction Let me just go to an example with the curves So here we chose four bands and four rows per bank Okay, so you remember, this was the formula that we were using before we’d be an hour on top So what happens, for instance, when we have point who has a similarity? OK, so the chance that a candidate we will be returned if the similarity is equals to point to is this very small probability There Conversely, with point, it is gonna be appointed sound probability of returning that has a candied, okay So we went from a family of this with these properties, so not as powerful as we hope to something way stronger, especially for the candidates that have very low similarity values Ok, so we’re starting to stretch this s curve as we need, and as you can imagine, you can make those number bigger, and this shape is gonna Keep on changing So now we’re going to be addressed your question How do we choose hiring me in a way that does exactly what we care about? OK, so we want to pick hiring me to get our desire performance

So say with Hicks once again fifty hash functions so we have R equals to five, five rows per ban and then ten bands in total What we, what we can accept I’ll discover once again is our green area here because it represents the false positives As we said, we want to make it as small as possible, but we can still afford having it because we can compute the actual similarity in the sequence in the next step and then discard the bad candidates What we tend to worry about his studies is blue here a year, because it contains are false negatives that will never become candidates OK All right now, here we start to see some curves that gave in a totally different way, even if we have fifty ashes in total So the computational cost for running on the stage with fifty ashes is going to be exactly the same But just changing these two parameters The behavior is drastically changing All right So, for example, if you take the blue curve, which is Article Five B equals Tan is it’s probably a sub Barker for a similar to zero point six because it will be having the blue hair there that we don’t like and something like the red curve with Tang and five is going to be given to strict OK because it’s going to be returning with probability equals to one only similarity values that are very high OK, So by changing those parameters and not too much, you can change the behavior of the system dramatically OK Okay, so we’ll go quickly through this business here What I want to show you is that you can also do the or construction before and then the end, and you get different numbers You just have to plug those two parameters in our formula, plot the curve and see how it behaves OK and you can change them As I said before, so now I just give you a real world example with a number that is not shocking, so let’s say you’re using a total of two hundred and fifty six ashes in total, so you do are for and construction of four or an end on a lower Once again you get this number in total, and this is the barrier, all our family, hashing timing a candies predict, will be a remains that pretty much from it It goes to one for anything that is above zero point of similarity, and this number is extremely small, frightened as below zero to OK, so you can see that the performance explodes very quickly as soon as you are adding a few more hashish OK, You don’t need to compute a million ashes per datapoint because you will be defined the old point of doing alization medical care So it’s yes, it’s still very cheap but it gets extremely more powerful by doing that Okay OK At the time Right, he’s again check later so let’s go just to the summary so we do with the key points that we saw before So you pick any two distances between our our two points, and then you start with a standard, Alice family that is added as the following sensitivity, and then you apply different constructions to amplify the behavior of this family OK and you want to make P one almost close to one and Peter almost close to zero OK All right Questions before we jump to other dcs metrics It seemed a lot of class file serving ones You mean something like this, Because here we are always plotting curves that are not very extreme You see, we have like wherefore ashes for thee We have this a foreign construction for construction, so you still see this kind of behavior As soon as you plug in one order of magnitude more ashes than these curves, you will be almost looking like a hockey stick OK, this was just for the sake of having values in this column In these roses are different Otherwise we’ll be seeing everything equal here and I was thinking on the top tax rate question Let’s go quickly back to take a look at that So in an ant, construction is is the following You want to make sure that you are considering a band Remember what we were saying before, So we have all the rows in a band and you want all of them to be asking exactly to the same value So this is the young construction it means If you want, you’re bound to be formed by ten rows, then you’re going to be needing ten different ashes If each one of them is matching, then you would say this band is matching OK And that’s why it amplifies the probabilities in that way

Conversely, do your construction is when you want, You have to be bands in which subdivide your metrics and then you will be applying an ash for each band and you are happy if just one of them is matching Okay, so we are just putting in numbers They turn it off That we talked about at the beginning right makes it more clear, no great Anything else, or we jump to the else Yeah, you can do bands within dance Yeah, let’s does thee This construction here okay, so aplenty for foreign construction, follow body for four and horror construction you, you can do pretty much, any kind of composition and they will change the numbers they will turn the curve in different ways So as I was saying at the beginning, Depending on your context, let’s say you are trying to do matching of fingerprints than it will be extremely bad for you if you don’t consider certain candidates but if you’re using this as a basis for a recommended systems, for instance to recommend shoes and clothes, if you’re missing ten good candidates, nothing happens OK, so that’s how given your requirements you play with those numbers Okay, Thanks very much Yeah Exactly because I’m still working on these examples by letting you think about your car similarity does is never going to be seeing from the following his lights, works with a different is a distance metric, and then we’re just gonna be applying that in a straightforward manner Thanks for your clarification Okay, oh, question Ask them since matched the number of roses Does the number of such functions matches the number of rows? No, he matches the number of rows in the inner band so you are gonna be having as functions for how many rows you consider in a band, and then we’re going to be having ash functions for our many bands you have in your metrics But there was if you have a varnish function Peru in your dataset, you would be working with a billion ashes right, and that’s not something that doesn’t work It doesn’t scale Make sense At times, the number of other functions Yes isn’t as the road, the reverence in which that’s something that’s a space of possible as functions that you can have is like with what we saw with mean asking, you can get all the possible permutations of the rose The number is extremely big, but you don’t need to use all the hash functions in your space OK, so that’s why we went from The space of possible has functions to the family that we care about, which has the properties that we like for PR and people We’re going to be seeing exactly the same in the following examples I’m gonna be giving you up It’s basically a base, all the possible as functions that you can generate a few remains of linear algebra, so you have this base You can generate plenty of them, but we are gonna be selecting only the subset that we care about makes it clear no great Anything else Okay, let’s jump to the two, the last parts So now we are going to be seeing how we flagging difference this as matrix into a stage So this should look familiar by now we’ve seen it probably ten times, so we are feeling now points in general data points in our as functions we want to generate shortly into their signature Now we learn how to design our They wanted it to be one people sensitive final hash functions But what we want to do here is we need to come up with a way to generate this family from other particulars distance metric Up to now we have seen it with a joker, similarity and the Jackal distance Now we are going to see it with different examples Everything else that we learned up to now is going to be applied exactly the same way, so that’s another beautiful thing about the message at once you perform this conversion and you obtain the signatures than whatever is the input data? The rest is going to be working exactly in the same way All right, so in the lingo machine learning, you can think about this as some sort of future engineering, and this is the model where we are touring the other parameters Okay, I guess kind of you are taking two to nine to the fore in this period, so I tried to make your life easier

OK, now We have seen how this works with a mean ash, so we have our bans technique, and we go from these representation with the implant metrics to something more compact The first example that we’re going to be seeing led around it has to do with Cassandra stance and, for instance, is started using the gnashing We’re going to be using the random airplanes, and this is the kind of representation we get Everything else is going to be working exactly in the same way OK, so let’s jump to the cuisine distance Why do we care about it? Why why is it important and it’s using a lot of different contexts It has been used a lot in information retrieval when you want to retrieve documents that are similar to each other, for example, so you do you have a pictorial representation of a document and you want to measure the angle between those two documents The smaller is the angle, the higher the likelihood that these documents are similar and what we want to do without a sage is once again returning only few candidates where the document back to this should be done checked OK, we can express the equal sign similarity you would find it in papers and textbook in a lot of different ways What we care about is that created as a range between zero and pi, and you divide our hunger thither by a pi so that we normalize our distance in the range between zero to one and as before, you can obtain the consent similarity just by doing one minus the distance, so the two things are connected in thee In that way, You will also find it define at times in the following way, with the dot product between the vector, A and B divided by the product of the norms, everything can be normalized to the other representation easily so I’m giving you basically this advice in case you see it implemented in libraries or you find it in papers The two expressions are pretty much the same What we care about, though, is the falling of capsize We were saying before two vectors that are very similar to each other are gone out of very small angle So the concern goes to one vector is our orthogonal Therefore clearly not similar to each other are gonna have a cosine equals to zero and interesting enough if they have hop was its course the cassano we’ll go to minus one, and that’s also very interesting in case you want to fine in case, for instance, you want to do going off serendipity in Europe from their system, such as trying to recommend items that are as different as possible to each other So this is why it’s also interesting to have this measure that tells you if two documents are at this, similar as it gets OK, so now that we have this distance define, how do we create our ash as family? We have this thing, we call random, I replace similar to me and I think so A lot of ideas will be in common, but we need to apply a few tweaks here and there to make it work OK, so we start from our standard family, which is the one you need to design or distances, and is our probability so one minus the distance once again normalized by my pipe And this is just a reminder of what this means We have seen it twenty minutes ago, so it’s not worth going to see it again So this is our definition of the random rubber plants ash family How does this work? How are we going to use it and why does he generate a slightly different signature company compared to mean ashes? So we do the following for each vector B we determine our Nash function H We just as two buckets So the return values are going to be either plus one or minus one And let’s see what this very simple hash function does so it takes the dot product and it returns, plus one if the dot product is equal to or larger than zero or minus one if the dot product is less than zero, what are we doing here? exactly so? What is this telling us? Anyone wants to try if you are fresh in your algebra in your brain Parts of the upper flame, the two vectors are lying Right OK, so if they’re lying in the same parts of the upper plane, we’re gonna be getting a plus one If it does not the case, we are going to be getting a minus one OK, so it’s very easy ash functions and how do we build a family? Basically, each possible vector that we can have in our vector space is a potential good candidate for an ash function, so we just take an arbitrary vector, and then we use it to do the dot product between the vector of our element And we use our claim is that the probability that the ash between the two elements is going to be the same should be equal to one minus the distance between those two points Okay, so one minus the angle between our two points Okay, So the closer is the angle The higher is the probability that our hash function will return a plus one

OK, now let’s try to make this initiative We are going to see it visually on those lives who was going to be okay These are our two data points action why is an angle, theater between them and we select our first vector? B? We don’t care that much about the vector We rather care about the upper plane, that is this normal orthogonal to our vector B, exactly for what we were saying before What we care about in our as functions is to understand on which side our points are lying fits on the right side or on the left side of the hyper plane And so, and is important in the fact that the upper plain is outside the angle, Not that the vector is inside OK, if the vector is inside, we don’t care because it’s not going to change the way the output of the hash function But if the upper plane cuts between the two points than the ash function, the ashes to the wax will be different from the ash output Hawaii OK because they are lying in different size of the upper plane so intuitively What are we trying to build here? We subdivide the space with different different airplanes and this set of hash functions We’ll just define an upper polygon in our space so we cut this space when multiple planes, and then with all the outputs of the different as functions, we’re gonna be understanding if the point lies within or outside of the upper plane Okay, so as it is that? And let’s try to see it here now Why? the the output probability is proportional to what I was showing to you before So again, let’s go step by step We have I just lost my Potter Sorry, so I’m gonna be waving We have our fine acts up there, why here then we have two hyper planes, one defined by the red vector So it’s the red dashed line, another one defined by the blue vector, which is the blue dashed line So what happens here? X and Y are gonna be hashing the same within the blew up a plane We’re gonna have a different hash for the function defined on the red upper plate Okay, because one factories on one side and the other one is on the other side So Y is the probability proportional to the angle The smaller is this angle The less is the likelihood that there would be another plane cutting between those two vectors OK, that’s the only intuition that you care about The closer they are to each other, the fewer hours play hyper Planes will be cutting between those two vectors Therefore, most of the hash functions will be returning a plus one Conversely, if this is your your initial angle, then there is a high likelihood that it will be an upper plane cutting between them Therefore, a lot of values will allow the rash functions will return minus one OK, that’s the only intuition about these random outer planes Technique is working Everything clear questions Great question, so you ask, are we going to sample our vectors that are defying the ash functions in a uniform way or not, and I can also add on top of that another question that could come up to your mind by the way, how expensive it is to generate random factors of different norm and with real values It turns out that we don’t need to do that OK, so the only thing that that we have to do is not an angle I’ll go back to this line, is we just wanna? Generate vectors of that component is equal to plus minus plus one or minus one so they’re all going to be in the unit sphere, and in this way we also make sure that all the hash functions we generate are not skewed, because we can just take Random Bayless plasma plus one and minus one, and then we can have a uniform distribution of our vectors that are creating the ashes when it OK So that’s why it’s cheap because we don’t have to generate random real values, and also that’s why we can make it uniform just by promoting the plus minus ones On the components of the vectors, OK, great Dimensional data, and, most specifically this customized wanted to start again Yes, correct the question of grade the question was If you’re if you’re datas and dimension doesn’t mean that you cannot have more than two to the power of van hash functions The answer is yes, and the answer answer is it’s way more than enough right because you usually we

We care about our message when our data is as a very high dimensionality Otherwise this technique is overkill, which, and two to the power of one hundred is a real astrophysical number So you can generate more than enough hash functions In that case, okay, right thanks for your clarification Okay, let’s go quickly through the air through the example Now we recovered our message Now, let’s try to get the intuition of how this will be working with the euclidean distance Now we have another family of hash functions that we are gonna be using, and those are all based on projections Okay, so what we do is our hash functions are just corresponding two lines we take lines in our euclidean space OK, and then we partition our line into buckets of size K We take uniform buckets, we partitioned is lying and what we do is we ask each point to the bucket containing its projection onto the line are going to be seeing a visual examples have become super easy well after that, but it’s all based on this continent Jared Lines very cheap operation to do in a euclidean space, subdivided in buckets Take the projection whenever you are a member of a certain bucket Does the output har ash functions occur and the key intuition here is that you’re nearby Points are always close to each other right, And so there is a high chance that they are gonna be belonging to the same bucket Conversely, distant points will not belong to the same bucket, so the ash output will be different Now let’s see the visual explanation, so we start from the lucky case OK, so we have one is functioning, we only have our life Some devaluing buckets than pauses are close, hush in the same buckets of our green and blue are gonna be machine together and read his in another bucket perfect scenario It’s not always the case So now we see the two unlucky cases The first one is when we are not lucky with a composition, so with our many buckets we create on our line So sometimes we could just be a lucky and deed The boundary is between the two markets we lie within two points therefore they in their split a simple workaround for this is simply to have more buckets But then it becomes more expensive So it’s the same tradeoff that we’ve seen before We’ve been asking, for example, and another A lucky case is when we are unlucky with the projection so here we’re always talking about two dimensions I think I’m answering your question now Can we generalize this to multiple dimensions of the idea is clearly yes here you, you can now see why, because you can project for nine dimensions easily, but the reason why you need multiple as functions is because if you are in a direction, that is a lucky like in this one, you might have hold The projections have ended up in the same bucket OK, even if you did a projection in the orthogonal direction, they could be belonging to all different buckets K So that’s why we need more than one line clearly to make it work well Great Okay, so we have our animation That’s That’s how you can work in case we have multiple directions, so you see, in some of these lines we are getting lucky and the projections are good in some others are bad, so similar to what we’ve seen before with a cousin Similarity is some of the ashes are not going to be ideal, but the majority of them will be putting their points to gather in case their clothes and we get the output that we want One interesting thing to see here is if the distance between the two points is weighs a lot smaller than a than the chance that the points are going to be in the same market is at least one minus this divided by a OK so closer they are to each other, the more chances you are to end up exactly in the same bucket Conversely, we have another case where, if the distance is studies larger than I than the point, the angle between the two, the two points, basically the production line and the two points should be close to ninety degrees OK, so in this case, When the angle is not close to ninety degrees, even the projections will end up in different buckets Now you can imagine those two points having almost an angle of ninety degrees compared to the production line, and that’s the unlucky case, when there will be projected exactly in the same bucket But once again, the three curious stuff Matthew Brush functions is not going to be working In this case, you will be working with another one and we can easily done make them members or different buckets Perfect now, here, once again we are going to be able to play our tweaks with bands and with the rose So the the standards family that we obtained from this is the first instance is eight divided by two hot way,

and these are the two probabilities that we were seeing before playing We’re using our end and whore constructions than we can amplify the behavioral as curve as we have seen before OK, so basically as a summary, and then I’m going to be taking some questions so that we can clarify everything What we’ve seen today We have data that comes in very different shapes could be tax could be vector, a schoolboy images, whatever you want, as long as you can represent it in a way that allows us to build our family of hash functions with the properties that we care about than this can become the input of our allies age pipeline so we produce our signatures Then we go through the early stage pipeline, obtain our candidates and we can play with the behavior once again with our Andorra constructions If it makes it easier for you to remember, just remember with our band and roads construction because asking me that it, those are the two parameters that we are modifying and we’ve seen that this can work It will mean ash, or you can work with a random number of planes or just with the last hash set of ashes that we’ve seen What’s in the euclidian space OK Westerns The earliest said hash functions correspond to the lines of the Quaker Hal It’s how they would classify as functions, correspond to Rosen did uncle respond their apply too, to the roles, we said, we have hash functions for the rows in a band and then we have aas functions for each single band The rows of the ruins of the design role, so it really depends on the representation that we have of the data in the main ash is going to be the output of the mean ashes From the promotions from the cosine similarity we have seen this sketch representation with a plus and minus ones OK, so this result is a signature A sketch where each component of our thinking she was going to be a plus or minus one in D In Euclidean version, we have seen that it’s based on the distance, So regardless of what is your output at the end of the day, you’re going to be having this a same shape of an input matrix and there you can subdivided by bands you can decide how many rows you have in a band Apply your rash galeras curve and the magic happens Thanks OK, let’s go to the recaps light so that we are in time and then I’m happy to get any final questions So the two important points that I want you to take out from this room today are the following First, one is we talk about the property where if the probability that two sets she wants you to our ash in the same value, then this is proportional to the similarity among these two documents OK, so we can build hash functions that have that kind of property OK, and this is the essential core or the lesage Without that we can’t really do anything at all Okay, if we were unable to build those as functions that everything as we talk about would be impossible and the additional trick that makes a lesser age from a very elegant solution to wander the scalable, an extremely powerful with large datasets is the fact that we can play with our bans technique, so we can decide the size of the bands and the number of bands with which we subdivide lower input matrix And once we applied this, we can make our message behave in drastically different way, Depending on our domain, OK, so I wrap up the lecture here I’m happy to take questions Qualifications Chat This is Tom You’re not necessarily using a rush function After you’ve got the candidates you want to check their similarity it could be, It could be anything knowing me You could work with softer plagiarism, and then you’re going to be having functions that tell you If they have similar constructs are similar blocks of code, you can have images, and then you’re gonna be analyzing objects in the images, so it really depends on the on the use case So the beautiful thing of our research is that first part you go from your input into something that is a set of signatures you get out of there He had the members of your buckets, and then you can refer back to your data Guess it’s a transformation, like when you do a linear algebra, when we go to a different domain, we do the dimensionality reduction that domain And then poof, we go back to our real data Make sense question how are we going to partition the lines, or indeed in the euclidian space example?

OK, yeah, great question thanks, this is a problem in case you have outliers for instance, like if you know that you have points far away from the others, then you will have to stretch your lines and your buckets are going to become wider just because of those who outliers Elysees, like plenty of other techniques work with dimensionality reduction, they do require some preprocessing on the data First of all, so if you really were outliers there this will underperform if you remove them You kind of define our nearer and d dimensional space where the points are lying within and then the lines are going to just be going from one extreme to the other Thanks for the great question Times Red Cross worker Yeah, I repeat the question for the sake of the old people in the classroom I’ll do we overcome the curse of dimensionality without lesage Ah, the answer is a lesage doesn’t help you to address that you need to figure out your own way to get rid of it so, and basically what it’s telling you is that if you’re using croissants similarity and you are aware that your date is highly dimensional and you’re going to be returning a lot of sketches with my as once so one day and the doctors are taller than you know that either kazan’s similarity is the wrong similarity measure for your use case, or you want to process your data in a way that removes the curse Great question thanks Anything else I think we are running out of time Thanks, We’ll see you next week Thank you

EXIM Webinar: Money Saving Programs for Exporters

good afternoon and welcome to our presentation money saving programs for exporters I cdisc and export credit insurance my name is Adrian cellco and I’m a senior editor for industry week which is producing this event this event is sponsored by the export-import bank of United States in industry our mission is to advance the business of manufacturing the industry week staff is dedicated to working with you to identify and share the trends strategies technologies and best practices that will help you build more innovative more competitive and more profitable businesses I’d like to introduce today’s speakers first we have Jennifer Simpson regional directors ex-im bank she has over 20 years of experience and credit insurance tax and financial statement preparation commercial insurance financing and finance management prior to joining XM Bank she held positions of the private broker for the ex-im bank for 14 years we also have with us Paul ferraro who is president of export tax management Paul started the firm in 2008 after decades of working in international tax he saw there was a need in the export community for an expert in international text advisory focusing on interest charged domestic international sales corporations I see DIFC services I’d like to turn the call over to Paul good afternoon everyone and thank you for joining us today what I want to go to the IC disc in how it relates to exporting companies and make this as simple as possible the beauty of the IC disc from a technical perspective not that we’re going to speak technically today it’s probably most difficult aspects of the Internal Revenue Code however for exporting companies here in the United States it’s very simple to administer and the beauty of it it is non disruptive to your normal daily operations and I want to stress that word non-disruptive absolutely nothing changes in your daily operations and will speak more that shortly but first we’re going to define wouldn’t i see disc is to acronym which stands for an interest charged domestic international sales corporation of a long-winded name but basically statutory incentive in the Internal Revenue Code was originally enacted in 1971 and there were some modifications in 1984 and where that said the disc has been around for about 45 years and I want to define statutory what that means I’m going to find that with an example when you think of your mortgage interest deduction your real estate taxes for those other two homes we deduct our mortgage interest on our individual income tax return because of a statutory provision within the Internal Revenue Code the purpose of such was to enable in assists us taxpayers with purchasing and owning a home the IC disc is in place it is a statutory incentive to foster and promote exports from the United States and the reason I say that the stress that is the disc in its current form is what Congress intended it to be now how would it provides us with this wonderful incentive in a simple sense as all of you know the ordinary income tax rate on our normal sales highest rate is thirty nine point six percent what the icy disc allows is if you sell products outside the US that qualify for the disk instead of paying thirty nine point six percent on those sales that were speaking of the export sales you will not be paying on most or all that income a 23 point eight percent tax rate which is the qualified given rate so how the disc achieves its purpose to assist us companies to sell their products abroad is by reducing your federal tax rate as many of you know that is a wonderful way to a preserve cash to reinvest in your company to pay employees to be more creditworthy the bank’s more cash flow much healthier company much healthier economy with that said uh when does this apply one part of this it’s only for privately held companies and not to be overly technical only individuals ie owners of S corporations partnerships sole proprietors are subject to the qualified dividend rate packs of twenty three point eight percent which is why in essence it’s only for privately held companies large multinationals do not get any benefit from a qualified given right so it’s inapplicable to them this is for small and medium-sized exporters so what does it apply for words that are critical manufactured and we use examples of sight this textiles in United States shoes jewelry clothing

produced films that are produced software that has code in the United States architectural designs are drawings that are produced in the United States grown agriculture and horticultural products processed timber extracted seafood the only one place in the world where we can buy Maine lobster and that’s the northeast one of my biggest clients actually nonetheless scrap metal those are the four basic components are products that will qualify for the disc there are some rather esoteric exceptions one this is literally mention the Internal Revenue Code enriched uranium does not qualify which is probably a good thing of anything that is mine from the ground whether it be cold bauxite raw aluminum any direct oils or petroleum products natural gas oil crude oil anything that comes out of the distillation process ie kerosene gasoline jet fuel heating oil those don’t qualify outside of that everything else qualifies now the next question is when does it apply and they set an example a US company let’s say processes pistachio nuts we’ll pick on California today processing company processor nuts with in California notes for grown in California and processing company exports those nuts to Japan that qualifies now we need to go a step deeper with the IC desc and another type of transaction that will qualify is what we call an indirect sale and this is very important is a very very underutilized in Miss part of the the IC desc in my example say we have a u.s. manufacturer that builds a large piece of capital equipment industrial machine in New Jersey that company manufacturer sells it to a distributor up in Pennsylvania and then within one year the distributor in Pennsylvania cells that same machine to a end-user in Canada the beauty of the IC disc both the u.s. manufacturer located in New Jersey as well as the distributor located the Pennsylvania can claim I see dis benefits now something to remember interject here the IC discs benefit is a based on federal tax incentive in order to take the benefit you have to be paying tax within the United States if you’re not paying any tax there’s no benefit to it it’s a reduction of your federal income tax it’s ultimately based upon the taxable income of the export sales so in this case in essence it’s not double-dipping the distributor will most likely have a much smaller margin on their sale as compared to the manufacturer and the limitation of the benefit is taxable income now that’s a whole nother 428 our class which I won’t go into today but basically always remember the indirect sales to qualify and next slide okay one of the other requirements of the IC desc we spoke of destination anywhere outside the United States qualifies with the exception of the blacklisted country’s north korea and so forth and also US territories won’t qualify puerto rico guam US Virgin Islands and such the other qualification we spoke of manufactured produce grown extracted the destination another quality or quantitative test is the u.s. content rule now when Jennifer speaks of us content for the XM Bank it’s a bit different but I want to focus on our rules for the IC discs at least fifty percent of the sources the fair market value must be within the United States now to illustrate that V in a mathematical example basically the denominator in our fraction is the sales price in other words or so in our capital equipment four hundred dollars per se going to our client in Canada no more than fifty percent of that total sales pricing come from outside the US so in other words as long as no more than fifty dollars of input is into that product other words no more than fifty percent of it is products from outside the US we could be buying Pistons from Germany electronic components from China and so forth no more than fifty percent of the sales price can be foreign components it’s a very liberal guidance that the IRS gave us on this is very it opens up a lot of

a lot of exports will begin in this day and age a lot of companies by components overseas where they’re cheaper but the value is driven here within the United States now remember the disc again is driven by taxable income with that said Congress intended for us to say if you have your gross margin or your markup is in the United States that’s what we want you’re paying more tax United States we’re going to give you this benefit you’re employing more of our workers in the US you’re selling your product and bringing the balance of payments back to the US a very very important part of the IC disqus is critical the question is does your product contain at least fifty percent us content the choices are yes and no and while you’re voting I just want to remind everyone to please go ahead and ask any questions you have throughout the webinar and we’ll answer all the questions at the end yes okay there’s our results so 96.4 percent say yes go ahead Thank You Adrienne and it took a little more to the benefits as we spoke about this earlier why do we have an IC desc what is it going to do for the exporter one it’s going to serve the purpose of Congress intended was to foster and promote sales from the United States very critical to our economy but again how does it work for us basically the ordinary income drive for the normal course of your enterprise ie your normal export sales absent a desk in place stacks at thirty nine point six percent when the disk is in place again the disk is non-invasive to your daily operations you do not invoice your clients separately you do not your clients your customers will never know a disk exists your vendors or suppliers will never know what disc exists it’s very transparent to the outside world in absolutely nothing will change with your daily operations that is very critical to remember what the iso disc as well but with the disc in place instead of paying the thirty nine point six percent on your export sales somewhere all that income will be subject to a twenty three point eight percent tax rate which is a significant margin in a significant savings that’s how the disc it sees its ends that’s why it’s here now how that is calculated in the detail is a whole nother eight hour class of happy to bring them to teach you guys that one but even though they went to today the process the first step we’ll use our exporting company and pennsylvania that makes the capital equipment capital equipment manufacturer forms an IC desc separate legal entity it’s actually a walking talking breathing living corporation as a checking account has corporate corporate document bugs as seal has certificates it’s a real corporation we formed this corporation with in any state in the United States or the District of Columbia once its forms we calculated commission based upon the Internal Revenue Code the Treasury regulations and so forth and other guidance from the Internal Revenue Service we calculate the commish which represents basically the savings on these export sales what’s that’s calculated the expert in company will then pay that Commission to the IC desc giving the exporting company a 39-point 6% deduction the IC desc then immediately pays that same em all back to the exporting company and the first to think well where’s the savings remember the commission paid to the disk is the deduction of thirty nine point six percent once that is will call it remitted or paid as a dividend back to the exporting company it’s only taxable twenty three point eight percent of the shareholders hence the tax savings hence the incentive to export and moreover foster the US government the Congress is helping us to foster and promote export sales what this also does and he distress this for a lot of you guys out there that are in competitive businesses you’re competing with another country or another competitor if you have a disc in place and say you’re you’re selling in a margin of and that Marge of eight percent because your after-tax profit is now going to be much higher on these export sales you’re really need to consider the fact that wait a minute our after-tax margin is much higher if need be to be more competitive to secure a new customer we can reduce our sales price we do that analysis for our clients quite often it’s a very very powerful tool to compete out there in the marketplace say commodity industries that the prices are fixed by the market this is a god-centered a lifesaver with a lot of our clients now remember something again this does not disrupt your business operations whatsoever your

CPA firm will know about it the IRS will know about it cuz we file it election the ISIS had this own return to file which we do as well but it does not disrupt your operations whatsoever it is as I tell all of our clients easiest money you’ll ever make now with that said this I realized this is probably most complex subject have any questions after the seminar peels please feel free to contact us email us we will be happy to help an agent alternate over you okay thank you I’d like to turn it back over to Thank You Adrienne and thank you to the participants from 42 different states that we have on this webinar today for the balance of the slides we’d like to tell you a little bit about who act them is and to go over one of the products that’s already helping thousands of exporting companies around the u.s increase their exports sale the export-import bank of the united states or XM if the independent agency of the US federal government established in 1934 headquartered in washington DC with 12 regional offices around the united states these regional offices meet with exporters in person and over the phone on a daily basis to help educate companies about the insurance and financing products that XM has to offer to help increase your exports in fiscal year 2015 XM supported more than 3.1 billion worth of export dollars from us small businesses exporting you might have heard of XM in the news as participant in large projects and developing countries but we do want to let you know that small business is XM’s business and nearly ninety percent of the individual transactions that were underwritten it Washington DC or force us small businesses we the XM follows small business administration’s definition of a small business so I wanted to just mention that if you’re a manufacturer and you have 500 or less employees or if your distributor with a hundred or less employees that’s the definition of a small business or exome the regional offices field a lot of calls from exporter’s not quite sure if they’re eligible to work with the XM products so we want to list some of the first questions that our regional office typically asked an exporter and that is have you been in business for at least one year do you have financial statements and tax returns or tax returns that you can produce and do you have it done in Bradstreet Duns number I want to share with you that there is an online portal government portal to the dun & bradstreet system whereby if you don’t have a Dun & Bradstreet number we can help you sign up for a DMV account at no charge if you’re not sure that you already have one we can certainly look that up for you as well in addition two years in business exims also looking at the products and services like Paul mentioned the fifty percent us content and I’ll explain our content rule I gave an example down here of a u.s. manufacturer that is importing raw material and the processing of finished good here in the US so they may have soft costs like the labor that goes into the production of that final product they may also purchase other raw material from US manufacturers that go into the cost of producing this end product as long as the us-based costs are at least fifty percent of what it took to make that finish good that’s qualifying as fifty percent you as the content product for XM if you’re a service organization a good example of this is the architectural firm you’re providing drawings or engineering services to a customer overseas as long as you’re us-based employees or us domiciled employees are providing those services to the foreign customer even if that means traveling abroad that would qualify as US content for XM Adrian I think we have a poll question now the question is what types of payments do provide your international buyers and here are the choices cash in advance open account letter of credit progress payment or other and the answers are cash in advance with 25.2 percent open account 21.4 percent letter of credit 30.1 percent progress payment 12.6 percent and other 10.7 percent that’s quite a mixer on this next slide I want to share with you one of the products that’s requested from XM the most one of the flagship products that XM offers us exporters

to help grow their exports the export credit insurance also referred to sometimes as accounts receivable insurance because it’s protecting you the u.s. exporter when you’re selling to a customer in another country on open terms of net 30 net 60 that 45 should this boring customer not pay you XM offers a guarantee of 90 or 95 percent of your invoice amount so including your price markup including freight to the customer if you’re billing them for that accents protection will offer you a 90 or ninety percent payout should that customer not be able to make payment some of the top risks excuse me top reasons that exporter’s use this kind of insurance protection include commercial and political risks that a lot of us companies identify they’re working with why by selling overseas let me go into some some examples of those specifically on the next slide one of the other reasons that US exporters use insurance for the receivables is as a sales tool we have quite a number of exporter’s tell us that they’re competing against other manufacturers maybe a manufacturer in China is willing to give your name customer in Spain 30 days to pay in fact dozens of other countries have their own version of the XM they may call it something different but there is government assistance for manufacturers in other countries who are selling around the world so by using the insurance as a sales tool you may be willing to extend this customer in Spain 60-day terms because you know you have been 90 or 95 percent guarantee of your invoice with XM what are the other common scenarios that we hear is a u.s exporter that’s been selling to the same customer for five years in Brazil and they’ve given them a fifty-thousand-dollar credit limit because that’s what the u.s. exporter is comfortable with other own risk they’d really like to sell more to the Brazil company and they know that the Brazilians can move more product so now by having the insurance the u.s exporter may go to the Brazilian come and offer them a hundred thousand dollar credit limit twice as much as they’d been purchasing so that they can stalk their product and make it more readily available country building brand recognition for their product and increasing sales but all the while the exporter has a protection from XM 49 year ninety-five percent of that balance typically the XM does not tell the foreign buyer about the insurance so it’s really up to the exporter if they let the the international customer know that they have this invoice protection another reason that exporter’s will use insurance their invoices is there’s a financing tool for themselves if you’ve sat down with your bank and spoken to them about what size line of credit they can offer you to support your export business and they’ve gone line item by line item on your balance sheet you’ve probably heard from them that they’ll include the inventory and some of the other assets here in the united states to determine how much they can lend q and a line of credit but they’ll often explode your porn receivable because that collateral now shifted overseas by using the insurance to collateralize those receivables a lot of lenders will include those foreign receivables in your borrowing calculation increasing the amounts that you can lend against some of those reasons that i mentioned that are commercial and political in nature when selling abroad and i want to mention here these are probably the top reasons that x mcs claims so when a u.s exporter is not getting paid by their foreign customer these are typically the top reasons that that claims coming in the door but definitely make sure we pay attention to these on the commercial side insolvency or bankruptcy of your foreign customer some countries call it something different but the equivalency of an insolvent or bankrupt customer of yours abroad they’ve gone out of business altogether protracted default is a fancy way of saying your customers having cash flow problems but they haven’t officially gone out of business yet so this is your slope a customer’s your protracted default accounts and I put an example here because it’s the classic one that we see it’s you’re selling to a distributor they have an end user their reliance upon payment from whether it’s for a project or ongoing sales that end user is not paying your distributor and now you’re not getting paid because it’s a trickle effect of non-payment that’s occurring it’s extremely common and it’s not typically anything relationship between you and your distributor you have a great long relationship but it’s really that they don’t have the cash flow to pay you because their end user hasn’t paid on the political side of risk we see currency transfer a lot of

us exporters are billing and US dollars they’re expecting their foreign customer to go to their central bank with their local currency exchange it in u.s dollars and to pay you the u.s. supplier historically we’ve seen countries freeze the currency exchange altogether you might have remembered where US dollars can’t leave the country at all last year with all the Nigerian elections caused a heavy devaluation currency and so a lot of Nigerian companies were waiting for that election to end in hopes that the currency would rebound and if you’re selling to Trinidad we hear a lot of reports about a limit to how much US dollars can be converted at one time we have a lot of US companies being paid in dribs and drabs against their invoice and not your full invoice due amount and then of course most recently of the Euro devaluations cause a lot of European companies to not be able to pay their bills in full or to delay payment these are the sorts of currency transfer risks that were more commonly seeing now war hostilities in the region that you’re selling to that causing you not to get paid cancellation of your exporter of an export or import license if you’re in the beef sector you’ll remember in the early 2000s Japan closed its borders to the import of us beef there’s a lot of product on the water a lot of products doing at the port waiting to be delivered to Japanese companies and that was not possible if you look at these risks there a lot of unforeseen events not something that’s really predictable and there’s not the aren’t risks that are under your control but they’re all risks that are covered by the XM insurance there’s a footnote in the bottom left here that I want to point out and it’s XM military restriction I encourage you to reach out to XM’s regional offices if you are selling a military item or if you’re selling your goods to a military agency in another country just to see is that the product that accent is able to support aside from that XM insurance for your invoices is available in 189 countries of the world as of today on the XM valve website there is a link to the country limitation schedule or CLS which is an alphabetical listing of countries that XM is able to support sales too it’s updated a handful of times a year you can even sign up for those updates to come by email there are footnotes on this document and so again if you have any questions when you read you over the page whether XM supports sales there or not please reach out to the regional offices a lot of questions so I’m going to go right to the questions the first question is does risk insurance cover at mount payment due to a disrupted a freight damage jennifer insurance is the AR insurance like i mentioned it as one of the terms is to cover the actual invoice remaining unpaid and so if products being ruined in transit that’s picked up by a cargo insurance policy which is sold through freight forwarders as they’re loading your cargo or through independent which I’ve heard is a little less expensive independent Freight insurance companies so the AR insurance or this credit insurance is for invoice payment it’s not for a product damage another question Jeff or how is commercial risk to find a commercial risk is a risk our apparel that’s associated with your customer directly so it’s not that they were impacted typically an event like i mentioned war hostilities in their region that would be more political event but again the most common things are your customer going out of business or they are having cash flow concern and they’re not able to pay being the most common commercial risk okay Jennifer next question if a company already has a subsidiary for its export sales can that subsidiary be declared a disk to get the tech saving if so what is that process that’s actually a great question a little get overly technical when a disk is formed it starts its life off as a typical everyday corporation and it elects to be treated as an icee disk with the Internal Revenue Service if it’s a newly formed corporation the disk is effective as of the date of incorporation if it’s in a corporation that’s already in existence the effective date is three months from the date of filing so it’s very inefficient to use a company that’s

already in existence moreover one thing to remember an IC desc cannot be the manufacturer so our professional recommendation notwithstanding other facts of the situation simply if you’re a good candidate for an icy disc form a new corporation the day is incorporated you can start taking disk benefits as of that day it’s very very inefficient and from a tax technical standpoint we’ll just call it very messy to use an existing corporation what great question thank you so much and Paul here’s another question for you can the ownership of an IC just differ from the underlying company and if yes how are the tower changes in ownership of IC disc handled another great question the disc can have a different owner than the actual exporting company there’s some very very technical issues in there and so much as gifting shares to your grandchildren or who should own the disc that warrants much more legal and tax technical discussion but yeah we have a lot of clients at the disc is not owned by the operating entity it’s owned by could be shareholders of the IC desk or the operating company or the exporting company also are the shareholders at the IC desc that’s very common yeah great question thank you okay Jennifer this question would be for you is the XM export credit insurance free if not what’s the cost that’s a great question most of the X and programs almost all have no upfront cost to the exporter to get going most of the programs are set up to where as the exporter ship to a customer that XM has approved insurance for they declare in an online portal what the actual sales were per month and a premium calculation appears on the screen could be based on the country that they’re selling to the payment terms that they’re extending their international customer which dictates the cost of that insurance and then with a credit card or a bank ACH debit the exporter pays to ensure the sales to that pre-approved XM customer typically okay Jennifer this question is after how many days does a protracted default kick in and assume your terms were in 62 your customer great good question there’s an XM claim filing window of five months so that window open once that 60-day end voice becomes 90 days past due so we’re looking at 150 days from invoice date on a 60-day termed invoice and again the exporter has eight months after that due date to file a claim the three to eight months after the due date of 60 days there’s that’s the open claim filing window for an exporter to style of XM okay Paul this question is for you what about musical instruments can I see discs because I’ll be available once again great question and absolutely as long as they’re manufactured here in the United States and we have at least be fifty percent content if they could be either directly sold to the foreign customer or they could be indirect sold ie through a distributor here in the US the then export the musical instrument absolutely great question thank you another question pause can content requirement for the IC disc includes labor as well yes one of the statutory in or sorry regulatory terms to quantify us content is what we call conversion costs and then there it includes labor absolutely always when it deal with especially international taxation the easiest way to clarity and understandings through intent and congress enacted the disc in 71 to also keep jobs in united states so yeah that will clearly qualified that will clearly add to the US content and once again great question thank you a question does an LLC qualify and does construction materials that do construction materials qualify great question and absolutely and more so very absolutely we have a lot of clients in the construction industry we have a lot of class for LLC’s there’s actually not to get too technical there’s an added benefit to active participants in an LLC due to the self-employment tax so yeah LLC’s are wonderful candidates for the aussie discs pay another i suggest qualifying questions are the c corporation able to take advantage of it yes that’s a very good question i was going to get a little technical and the folks at asses question more than welcome to contact me after the event but yes what would happen is if it’s a closely-held seeker you would have these shareholders of the

C Corp be the shareholders of the IC desc and what will happen is instead of the Sikh or paying dividends directly to the shareholders the IC disc will pay dividends and that the is actually more of a benefit with a c-corp so yeah it’ll definitely qualify destructuring changes a bit but yeah absolutely here Paul this is another one that’s a little long such as stay with me it says we distribute fasteners in the aerospace industry some of our vendors or manufacturers that we buy parts from and then sell to our customers are overseas so we are importing these parts initially does this make a difference with regards to the disc that’s a very technical question that require a lot of all the velveeta great question more understanding of the particular situation but if you’re importing parts to the US and adding no value to them and then just shipping them right out unfortunately no because again think of the intent we would be doing there is fostering and promoting export sales for another country if you’re bringing the parts in adding value to them then yeah then it could be a different circumstance but on the onset unfortunately no okay so Jennifer this question is for you how long does it take to put an export insurance policy in place good question the XM application is as short as two pages for some of the program and our regional offices of XM will help exporters put that together submit it to Washington DC to an underwriter typically the exporter themselves is underwritten in a week or less and you xme typically will attach go ahead and attach a customer an international client that you would like insurance for to that request so while the exporter is underwritten a week or less there’s a due diligence period for that international customer as long as XM knows that international customer it can be coming back to the xword with their own approval so it could be again a week or less if XM has to do brand new due diligence on an international buyer that they don’t already know or that they have stale information for it could be a week to two weeks tops it’s sort of like ordering an international dun and Bradstreet report but from a local Asian country are sometimes the X and will get assistance from the US commercial service to provide information our intelligence on a customer in another country okay let me do a follow-up one on that the follow-up is once a policy is in place can I use it to cover sales to existing and new foreign customers absolutely x10 both options so if you’re comfortable with the ten customers that you’ve been selling to for the last 20 years and you just want to use the insurance to grow your export add new customers and again you to the sales tool or marketing tool you have the option of not ensuring these existing customers that you’ve already been giving credit to you on your own or selling to that you’re comfortable with so XM has the option to ensure as few as one customer and your whole portfolio that you give credit to if you decide that and there’s middle ground policies as well and I would you want to emphasize that no transaction is too small for XM if you’re selling two thousand dollars to a customer Romania on 30 day turn XM still wants to hear from you okay Paul question for you i’m currently selling components to manufacturers who export overseas how can i track when i export products with my components so i can take advantage of IC disc great question Thank You Adrian the easiest way for that a lot of our customers are involved in that very same transaction and the customer who is the exporter Auggie the distributor can provide a letter stating which products were exported some of our clients there are many factors and very specialized products they’ll know they look at a schematic say it’s an engineering somatic forgiving type of specialty piece of equipment and they know it’s being used in a foreign country based upon its use at the voltage and so forth there’s many different ways the we call them the secretary or the Commissioner the IRS accepts the proving destination very very open to that we like the letter from clients but there’s always more ways to figure it out almost anything okay Jennifer this questions for you what are the deductibles on the insurance is a deductible per claim and is there a minimum claim is there also a minimal claim per customer or a limit on the sales permitted exims policies for small business and again we defined that isn’t the manufacturer with 500 or less employees a distributor with 100 or less employees

most of the products that XM offers carry a zero deductible and again it’s a 90 or 95 percent coverage you’re looking at what we would refer to as a coinsurance of five or ten percent but most small business products don’t also have a deductible on that as well for the policies to our us exporters that are not small business or who are extremely experienced exporters there are policies that will have the deductible on them but as a trade off the cost of the insurance might be less so XM does consider the whole portfolio that exporter’s are selling to on your deductible policies and trying to minimize the amount of deductible that does exist unlike to auto insurance policy the deductible is not proclaimed it’s a one-time annual deductible if there is one at all okay Paul question for you does it matter what US states the IC disk is incorporated in in other words is subject to state corporate income tax Thank You Adrian that’s once again a very great question as well as an important one most states favor the IC disc you’re not going to get any benefit at the state level but I’ll set some examples the three largest economies in the united states California Texas in New York George state all have favorable laws to these IC desc in other words they don’t tax the IC desc there are a couple states out there that are not as favorable to it but the majority of the states are very favorable to the IC desk and think of it two states are just as competitive as countries are to get businesses to their states so we who’s them to be very frontal to the IC desc now that question does require more obviously understanding of the actual client itself or the company itself but for the most part you’re pretty safe okay Jennifer this question our experts are non-cyclical a client may purchase very randomly can we purchase a our insurance with spot orders provided to the customers pre-approved with XM absolutely we have a lot of customers that are an evil nature and they may ask one to two months out of the year or they may buy the insurance for a particular sale to a customer but they’re only going to make a sale to one because they have some capital equipment there’s not a need for that end user to purchase again so soon when XM approves an international buyer for the u.s exporter that approval is good for one year so even if you’re not ready to ship to the foreign buyer but you want to go ahead and make sure that you can obtain that current approval you can receive that insurance certificate from XM it’s good for 12 months but at the point that you actually make the export you’re logging on to your XM account declaring that that sale happened paying to interpret to ensure the sale but other than that there’s no cost incurred and it can just be a one-time sale shirt okay Jennifer this question for you is the insurance sold directly by XM or is it through a bank or another financial institution that’s a good question X M is the insurance company is the what’s on the paper of the policy ultimately back by the US Treasury accident is a self-sustaining agency by the proceeds from these insurance policies XM supports their own expenses and does not draw on taxpayer money and so it is written on XM Bank paper yes okay sorry there’s not a lot of questions here this question is for Paul does IC disc apply to all my exports were only exports to certain countries a great question thank you your exports to all countries with the exception of the bad quest countries North Korea Iran if you’re selling to those countries you have much larger issues than an IC desc but for the most yeah every country qualifies except the the blacklisted countries a great question no thank you okay Jennifer the question is it looks like that their companies registered in the US purchases from Canada and sells international would they be qualified to cover for a insurance I’m sorry because I believe you said that they purchased from Canada with that correct Adrian yes correct understood thank you yet the u.s exporter is purchasing a finished product in Canada and like Paul mentioned earlier not adding a value to it why as it comes through the United States XM would not be able to support it because the content were made the foreign content a non-us content exceeds 50 percent and so XM is held to a

standard of supporting products that are fifty percent us content or services provided by us employees if this is another question senator this is they attend a trade shows and interview various manufacturers on latest technology they’re mainly an exporter not a manufacturer how would a 50-percent role work with them I’m so sorry Adrian could you repeat that question it was coming fair now yeah okay sorry it’s a little confusing it says they go to trade shows and interview very manufacturers about the latest technology presented they want to know about determining the fifty percent rule because they’re an exporter not a manufacturer okay so if you’re a US distributor of products manufactured by other companies there is typically a discussion with the manufacturer as to where those products are made the XM would mean some sort of certification whether that come from the manufacturer themselves to justify that they are fifty percent us content products we have seen that a lot with us distributors buying and they’re not quite sure where their suppliers manufacturing but that is a requirement of using the XM products to drill down and make sure that they are fifty percent US manufactured products thank you okay thank you Paul those questions for you what are the key questions to ask a tax advisor to determine their ability to provide guidance I see disc hey let me explain it this way the icy disc is analogous to go into your general care doctor and he sends you to a heart surgeon who then sends you to a very specialized heart surgeon in order to the disc properly it takes one a very deep understanding of federal tax to a very deep understanding of international tax and three a very very deep understanding of the IC disc rules and one thing I stress for the IC disc when you do it properly in accordance to the code in the regulations you get a better answer most of our work actually comes from other CPA firms letra for us their clients to do this if you can find someone has all three for that does take a very sophisticated software to do the calculations there’s a lot to it so I’d say that’d be very difficult but we’re always happy to help Thank You Adrienne Jennifer the question is you spoke about dividends what about use of producer loans are they still liable major I think that’s a question for me actually that’s producers loan as part of the IC desc and I assume the folks that asked this question are very tactical that’s a great question they’re still very viable we do have clients that utilize those it does add some sophistication to the up to the calculation and there is some rather strenuous cases aren’t there that you need to look at but yes it’s only for those who manufacture the products but a very good question in the article liable thank you sorry about that Paul Jennifer the question is our premiums paid monthly yes they are and all the policies that XM Bank offers the exporter may or may not pay a deposit up front typically that’s that’s no no front commitment and then on the monthly basis the exporter will declare the sales that occur to these XM approved customers and that’s how premium fit on actual fill that went out the door to XM proof customers with no requirement that there be X amount of premiums reported so there’s no minimum reporting requirement for premiums paid i should say thank you okay and then we’ll do this last question Paul the person wanted to ask if you can refer them to another small business that is taking advantage of IC disc absolutely we have a myriad over copia of clients that can speak of this I have clients all the way small as a million dollars of exports all the way to about 500 million dollars of that sports it’s all over the gamut so yeah absolutely okay with that I want to thank our audience

Haskell 2014: Keynote, Patricia Johann, Interleaving Data and Effects

so thank you to speak so I’m going to tell you a little bit of ever leaving data and effects as this is that don’t work with block acting economy and particles so pervy lectures provide a wide array of contracts for storing and manipulating data of course our built-in data types Thanks like those trees are as other kinds of algebraic data types and often not always but often these are pure so the numbers they don’t cooperate any kind of effects but sometimes data types not only incorporate um effects but also actually interweave those effects with pure data so unfortunately this isn’t always reflected in the types of the data themselves so to see how this can happen let’s consider two different scenarios one is that a effects are implicitly built into every house cool data type right so every hospital type allows the possibility of non coordination while it’s better to pure a value of that type so not only is an automation present in a type like for example this day but it’s actually interleague right because there’s a possibility of illumination at the type a as well as f at the broader list level so there’s actually this fundamental evening of the data and the effects throughout the non termination those particular effects but in general effects throughout the entire time and in particular because of how those lazy semantics house constructors can also be infinite or non terminating as well as finite so if you think about the list the topless a that’s actually the data type of finite list as well as infinite list those infinite list but if you look at the time you don’t see that anywhere in the type annotation itself so that is maybe not entirely ideal another situation in which you might don’t look that closely interleaved effects is a solution like this if you think about something like a ticket contests um function this takes a fine handle and reads from that file and then returns um well not believe there would it returns but at least the type suggest that it’s when it’s returning is of a pure list into a pub and wrapped up in an IO action and that’s precisely my point but that’s not really what happens right so what really happens is that you have um you have the the effects kind of happening in this interleague play because um you read from the program of you read from the handles from as the dullest elements are needed so here we have a slightly better situation because um the at least the type here reflects the fact that there is an effect right that was better than in the list situation but it doesn’t really tell exactly what’s happening in terms of these standard implementation and this is not again ideal there are lots of things that compassion for example you get a error messages recorded somehow the wrong time like so it can look like I’m errors come at it and that the errors come when a pure data’s being ridden or handled and that isn’t exactly right and can’t actually close the file handle um yourself you don’t have access today I’m going to add and um of course sometimes the well but you may not have a cliff at closing either so the first question I’m asked is how can you make din relieving of data and effects explicit in x so we don’t end up with the situation’s out me ended up with both the list example and this handle example I have a flower handle example so what we can see is that it is possible actually to represent the interleaving of data in effects so here i’m working at the first example this intervening of non termination and we’re going to have is a type on the right here of this lazy list type and what i have is it possibly not terminating a computation that produces other an empty list or a list that i get from a value angular possibly nonterminating list so this at least represents a situation others talking about before similarly we can interleave the list david talk now not with non termination but here with an io action okay so you have a non interaction that other returns an empty list or a list in this case I use done char because the the farm in the example would return chocolate of course there’s nothing special about our and another unpossible interleaved io list and so that’s um that’s a little bit better every sweet know that we can represent the pun interleaving ever interested in what

once you do that can you reason effectively about these effective data types so i’m going to use the term the fact that data types to talk about this kind of interleaving and i’m doing what in part because that’s a term that foods can stove i used when they first considered these kinds of data types so to answer these questions what I’m first going to do is a recall standard initial algebra techniques so paired data has benefited greatly from the idea of initial algebra semantics and I’m going to remind us of that and how that works because what i want to do is build on that to show you how same kinds of ideas and I am an effect interleaved setting so um having reminded you of that i will then argue that the straightforward application of those techniques is an interleaved affection setting doesn’t up doesn’t really give me what you want so one possibility that you could try to straightforwardly applying those techniques and that might seem like a good idea but it runs into trouble one possibility is that you could try to ignore the effect and of course that’s going to lead to add some conclusions another possibility is that if you do try to be cognizant of the effects they end up with reasoning that oh well that’s a lot more complicated than what you would like so what we’d like to do is have something from where we end up with really nice nice reasoning principles for a factual data is just in the same way that we have very nice reasoning principles for pure data so the way we’re going to achieve that is we’re going to separate the pure parts and the effectual parts of these games these data types using them what are called fnm algebras so effective and basically the palms our lives because you’re trying to reason about the pure parts of the data and the effect reports of the data at the same time and if we have a way to separate them out Leslie then that will be lead to this kind of nice one larity thought i thought i’m interested in here so we’re going to use a methane and algebras the idea I’m going to explain that that is during the top so you’re not it’s not good to know right now hopefully you’ll know at the end oh if I do my job right the idea is that algebra are going to describe the pre Parsons and effective data type just as they I’m just as to do in a pure setting in then we’re going to use on my own for more articles for a lunette em to describe of the effects ok so these FN ml tours are going to allow you to handle this evening because what we’re going to do is represent our effectual via text as initial fnm octopus and what I hope to show you by the end is that doing that allows us to handle these data in effect at the right level abstraction and what I mean by that is not just that you get some reason Invincibles because of course you can use your pure reasoning principles if you’re willing to kind of go from a lot of trouble and I will show you how how that can look but the idea here is that we’re going to get reasoning a reasonable Center epic right level of abstraction so in particular will be able to reuse the Pure Reason principles for the Pierre part so not just do something kind of similar but actually we use them and that will be nice and then we’ll have to do a little bit of extra work to handle the effects but these things will be completely separated so that we can see what obligations company / from the pure parts of the data structure on which proved obligations come from the affected parts of the data structure and then finally I’ll go back and revisit those two examples that we had in the beginning that one’s about the lazy lazy lesser possibly non terminating this and the one about the pot handles I’ll try and look at those from the perspective of effin ml troopers so I get started let’s remember how things look in the pure setting so the idea behind it a shot with semantics in the in the particular case of pure data types does it have a functor app and it’s going to model the individual layers of the data structure and I’m here when I say I’m sure I really mean a hunter so I’m you know thinking about a categorical front row but of course you can also think of business just an instance of the hunter class in high school and I’m thinking that the functor of the f not much in here so I’ve got a functor fmf Gotham a promotion and that that function is going to preserve out in these in composition although of course there’s no way to enforce this right and then I’m going to describe how every player and an inductive data type of data structure um can be can be processed using an F algebra and then F algebra is just a carrier a here and the function that tells us how to take the layer down to a value of type thing and then I want to characterize the data type as a carrier of the initial algebra and we write that as a UF and we tend to

call the the nap this this method goes with that structure now we tend to call that in and it’s going to go from F of that carrier to the carrier itself just as just as it does in the general situation oops to do oh okay apparently those life wants to change so we’ll let it change so I wanted to make sense of at that part about the initial F algebra you need to know if in FL 7f algebra homomorphism is and that happens to be right on the next slide that really really wants to come to you so an economic super homomorphism is just it’s a map um from one carrier here to the other one that I’m interested i sling with those two structure maps okay so it’s just the normal thing and the thing that I’m allows you to have an initial F algebra is that oh well you have if you look at this from this algebra with UF is the carrier indies in structure map then if i take any other algebra over here there has to be a unique map from the initial algebra to the other algebra and we call the the map that sort of mediates that mediates day we call that fold so i’m going to use the bananas notation but this would be fold k if k is the structure map of the tamo morphism I sorry the target algebra so this is just the standard on Pierce pure situation okay so an example of this would be if you look at the functor list fa then this will describe the individual layers of a list with data type a so the functor is described over here on the left and not giving it set up over here and yet not a function just does exactly what you think it’s going to do it it preserves a structure let it um it changes the data a little bit not the a data but the first data and then for the moment if you think of the type list a as the type of finite list and i know this does not have to somehow school but if you think of it as the type of finite list then then that will be the carrier of the initial F algebra and here I’ve actually given an explicit description of that structure now okay and then the fold is just our standard bolt our function like so very familiar to everybody so the fold is mediating that this unique back that i get from madhu shalini is the thing that is the thing that we know and love was just full are okay but there’s nothing special about this year I can do exactly the same thing in general so I could have an arbitrary founder at those just describing the layers of the day to time and um we can implement again ignoring on possibility of documentation for the moment we can implement it on the carrier for the F algebra that has that’s going to be our initial of algebra and so here I’m using household record syntax to make an inverse I’m in to the constructor in and this is I think I’m completely standard and then not to F is going to be the carrier of the initial algebra and as you can see the structure map is exactly how this constructor here and the fold we can justify using come standard recursion this is basically just chasing them that dynamic on my hand on the previous dog previous slide but instead of something linked down and across from that and saying that that’s the same as if I go over and down all I’ve done is said well if I go across like that I’ve got a new perspective chase around them the square it this way okay so it sounds like I’m saying is that we can make a fold we can talk about the opinion or the possibility of determination of the money we can describe the carrier of initial F algebra and we can talk about its fold as soon as neatly enough okay so great so that’s a situation the pure seven so what if we gained well this gives us some nice um nice definitional principles for describing of data types and functions over them so we have fold operators and we can talk about I’m definition my pattern matching we have some nice principles for a reason about such functions we have things like they’ve got two rules we have full fusion rules and many other kinds of things that are similar and um from the National for semantics you actually get quite a lot of stuff um you get quite a lot of tools for structured programming ad for an ad for reasoning about those those programs and I’ve listed some of them but you get lots and lots of tools I don’t know I don’t all set for free but I kind of think of them as is referring but mostly what you get is um you get this very principle of generic approach to dealing with pure data types okay and that is exactly the

the thing that we want to have enough in the effectual sent as well okay so before I start talking about the effectual said i’m going to show you the kind of thing that you want in into pure settings its kind of thing we want to port to the effectual setting so i said you get these kind of proof principle well here’s a here’s a free principle that we get them for free from the initially so suppose we have an f algebra that’s an AK so a is the carrier case is structure now and i have a function of G going from UF to a then I know that that’s the full of some okay if well if basically if it sends it if G satisfies this this requirement that I want to fold this this diagram that defines the whole right because the director has informed just like this except to have full came right here right with little handles and all I’m saying is i forgot function that does exactly the same thing that whole does and it fooled us a unique thing that does that then g have better p not fold okay so that’s the proof principle we use it in though to which direction i will want to show things like that full k is g and so what i will do is show that on this condition here is satisfied ok so i’m going to use the principle in that way this is about conditional but i’m only going to get one direction ok so let’s use it for example 20 to prove that depend as associative again this isn’t the pure case nobody surprised by the fact that about dependence associative in the pure case but on this will allow me to give a when we start talking about effects so let’s assume that we have we have this carrier finish algebra so in other words assume that we have this data talk that we’re trying to describe by the functor with Stefan then you can define a parent um just I think I’m in a pretty standard one right just think of a countess being the fold of some okay and what k well the one that I basically just walks down your list reconstructing it except that when you get to the end you stick the thing that you’re trying to upend onto the onto the first argument of append in cornell and if you unfold this definition of a pent you pretty quickly out get to these equational properties of a pad these are personal properties just telling how pen works on empty list and online idealist they work just like you think so if you want to show associativity of the pent well let’s just write down and what it is we’re trying to prove that this is the statement that we’re trying to prove that’s completely not surprising and the way that we’re going to do it is by using up in the principal so what we’ll do is well Stan Smith the proof principal oh but we’ll think of these people think of the left hand side as the application of the phone they were trying to show that that fold is equal to the right hand side and the way that we’re going to use this principle as we’re going to see that the left hand side of is full k where k is exactly the k that was inflicted the I was like that’s defined event so in that situation remember that what what the K did was it to flipping at weird depend on to the first argument to event and it replaced a lil by that and you see that happening right here right this is the second this is the thing I’m trying to upend onto the acts and here I’m replacing the nail exactly why that thing and otherwise I’m just completely they’re constructing list ok so to do that according to my for principal it suffices to prove um just this lowly quality basically what I again would everybody do the show that um that for me I’m trying to show that the function i’m interested in satisfies the property full and so this ends up being the equation i need to show i’m not actually going to show it i’m just going to tell you that on the way that you do this is by a very simple case analysis according to whether or not the list is empty or not and you’re going to use those properties of the pen that i just showed you on the previous slide that describe the behavior of appended are empty or not and it’s really easy and straightforward takes about nine lines so uh yeah very short it’s me okay so this is the kind of thing that want to port over to the effect while setting well if you want to do that we better think about how to talk about effects at all and i’m going to talk about then easy one ads so I’m gonna add is um you can tell from the fact that i’ve got the this end and I’ve got a deaf Mac you can’t LOL factor and i also have a little bit more structure I’ve got overturned a return operation and a joint of course the Mona I’ve also have to be satisfied I can’t just go around picking

any any functor and two other things in common return join some naturality laws for attorney join also be satisfied so all things need to behave one half way and the idea here is of course that is going to should maybe go pepper turns going to show me how to take a peer these are doing it and see it as of an addict data piece of data and joy is going to show me how to collapse these layers of effects right so that’s the step that is the standard situation let’s just go back and some examples are the ones we’ve already seen the North terminations on that and also the iron out those are the ones from my two examples earlier but there are lots of other one two Airmen are continuations nomads in Sylvan for the purposes of this talk you need to do it about anarchism is so that’s just going to be a knot between between to monett so assume a cada mañana m1 I’ll just reverse them other carriers and another momentum m2 then a modem or person is just an app from m1 of some from data type a to n2 of that same data type that preserves um he’s f cannot return to join so I just need to preserve the structure so if I once I have a way of talking about 1 F then the time does not using with that then I can go back and those effectual this so those are thinking right now about the yourself or I’m interleaved with um with on termination and I’m also thinking about that the list that were interviewed with the file and I see that I have a common generalization of those I think there’s nothing special about I oh there’s nothing special about the documentation runner here I think basically have exactly the same structure but with a pending in the hairline so make a new day to talk that um has a an hour level possibility of an effect wrapped around the list that’s how you’re acting or if i get if i am find Sofia value i have on to another session list another potentially effectual list so this is this um completely abstracts away from the particular moments audio and a non termination limit that we saw the examples earlier and in fact there’s nothing special here at all about this so i can actually abstract over the data tongue as well so instead of using the hook term list fa i called it before 44 list i can use any vector F that describes a data type and now you see that if I have an effectual data type again I have a potentially a potential effect and I have that wrapped around a data structure that users F to describe the layers and then constructs me data from other to let other potentially effectual structures of the same kind so this data top that I’m describing about that it describes a pure of inductive time that whose layers are determined by F I’m interleaved with the facts that i gave them i am so this is the way i’m going to talk about interview data and effects okay so let’s think about on what what would happen if we charges prove that event is associative for these effect full list well i have to have a way up to describe the day / time so as before i’m going to assume that um that a certain a certain initial algebra exists right now I’m the one I’m going to be interested in is the composition of em and my list one so i’m interested in factual this and that’s why I’m opposite homeless again but I’m not particular about the pomona so lift ma is up is exactly um looking kind of write it out it’s nice a Norfolk to exactly this that will ever act around this area of initial for that i’m done i’m asking you to send these exist and then we can write our effect Philip ansonia pen for effective role in hand we can write that up well it looked at the actually if you write the code that looks really pretty someone told me before for a hand except that I’ve got this new stuff enough written all this new stuff in green and all of this new stuff that is there because append itself doesn’t know I mean are all dependent to throw anything about facts but now I’ve got to have a bunch of effects you know inter alia Bhatt people properly handle the effects and the way I do that is by using return join and that’s nap and I use them as described here too to make my sexual event so it looks really similar to the old append

but I have to take care to properly handle all the effects okay um let’s say somebody said that so okay so um as before i’m just going to unfold the definition of our effectual event and we will get properties of a pen that will use um well that would love to use in a very similar way it won’t exactly be possible but i would like to use them in exactly the same way as before so here again i put the new stuff in green and I any what you can see is that at least in the statements of these properties you end up with a lot of returns hanging around but there’s actually more going on underneath because remember that in the definition of ESN there are f valves joins an turns here it only looks like I’ve got returned but I’ve really got a lot of other stuff going on because I’ve got the definition of the event here so in any case I end up with some properties that describe how the effect i hope you have behaves more or less on lists that are either empty or not and I get a little bit more um I had a little bit more um here but um I have to do I have to do some work because I have to shuffle all these returns and joins enough maps and all these things around simply because the the principal doesn’t think about how to handle the facts and this is the something strange about the example I’ve chosen I mean I’ve chosen a fairly simple example there’s nothing there’s nothing here at all that’s specific to the particular functor there’s nothing particular to well I haven’t even chosen I mean even specified in effect so it can’t be particular to the effects but we end up having to repeat this kind of work over and over again so it isn’t it isn’t as though I just want certain properties and I need to do a little bit of work for these properties I end up in a situation where I have to do a lot of work a lot of shuffling over and over and over again just to deal with the fact that the my profuse will doesn’t know anything about facts okay the situation is actually a little bit worse than what I just said because um we won’t be able to use these properties directly just as we did in your case and as you’ll see a young very very shortly I think on the next slide um what we have to do is unfold the definition of effective ahead in a way that we do we have to unfold the definition of the purif and in order to even begin to be able to use the proof principle so that really is not what we would like at all okay so I’ll let see how this is going to go i’m just going to write up here the statement of the the thing i’m trying to prove it’s exactly the angle of a statement that i have for ten it’s just a saucer tiffany so we’re going to begin in a very similar manner we’re going to try to use on from principal one the one that we have before that says if you want to show that something’s a fault you just have to show that it satisfies the property that that fold satisfies but here I have to do a little bit of work because if you look go back and look at the definition of of our respectful attend you’ve got this joining who’s math and I’m so you have to actually fold that definition of a respectable event and somewhere inside there okay and green you see that there is on that there’s a fold so that’s good that suggests we might be able to use our group principle but I have to do a little bit of work first and here I’ve introduce some notation I’ve started subscript in my um my function K with the thing that I’m trying to upend the thing I’ll prior to replace the empty list with I’m trying to attend on to the other thing and that just keeps the formulas from getting enormous okay so then I have to use the definition of effectual append three times I using naturally of join I’ve to use the third bone at law I have to use the fact that if that preserves composition and then finally i can rewrite the right-hand side to something that has oh it looks a lot like the left-hand side except it has this part in green and except for the if I sort of them focus only on the green parts then I see I really do have something too much I can apply my original prints wall fell joins enough mass and and things hanging around but if I day around and flat big enough of wanna then imma find something to which I can apply my my first Buddha principle okay so that’s good so this it seems like it’s applicable so we’re going to pick on those we’re going to pull up those green parts and we’re just going to try and prove this equation and this is this is again trying to show that something is all so we look at the pot the condition that we needed to show according to improve principle in order to prove this some equation about fault and it’s not you know too horrible it’s not it’s not too long but we have to do we have to do a little bit more work than in the other case so as the more

we’re going to use a case analysis according to endure all this is empty or not and um well what you see is that you end up having to use a lot of extra reasoning so in the pure case I um I showed you that the proof which I didn’t show you the whole proof but I kind of them walked into it a little bit and I said it’s very short it takes about nine lines it’s really sure it’s really simple absolutely difficult at all but here I do a lot of extra things right I’ve got these effects lying around and that means that I need to use um I niches you know nationality and I need to use profit properties about that that I need to use on loan and laws so I need to do a lot more work just to be able to show the same kinds of things ok so here the proof is upward of 25 a lifelong um they’re complicated you know it’s not just follow your nose so you have to do a little bit it’s not terribly complicated but it’s not as nice as what you would like so in the interest of full disclosure at this point I’ll say that um well the problems are that some non-trivial me writing is required energies purposeful one that’s why should you two slides ago and you have a lot of unfolding unfolding the definition of a pad and you need to have all these this kind of um repeated work this work that’s just done over and over and over again again not because of my particular example I didn’t choose a hairy one but just because of my example I mean just because of the situation here where we’ve got this principle that isn’t some is a cognizant of effexor doesn’t know how to handle them very well at least but but it is true that you could you could maybe make a slightly shorter through them but when I than the one of my 25 line proof the reason I said that is because you could be a little bit more clever and you could realize that in fact full event is actually some kind of argument / muted bind operation and that would let you two more quickly get to the situation where you can use the principle that I had and you could also say oath or recognize this this is just instance a full fusion so um if I if I useful fusion that I will deal that will eliminate some of my lines and that’s true but you have to go around and prove that fault fusion works in the effect whole situation and that would be another exile and by the time you got done with all of this um you might have a shorter proof of the appended you’d actually have a few reusable tools but you still wouldn’t have in my separation that I want so what I really want is to be able to separate the reasoning about the pure parts from the reasoning about the attacker parts and even with these alternatives those things are still intermingles so what I’m going to instead it she sees them F&M algebra that I mentioned before I’d imagine that they were due to Flynn skin stove rank um originally in an sfp 2007 paper they were only in the category CP oh so they work with the mega complete of Omega scipios and they they worked in the setting of induction and they were trying to prove certain properties of a functions and they could only have all your certain kinds of thunders not arbitrary functor sackboy with handle hair and they were only able to work in CP oh and we’re working here in arbitrary categories for these internal do is exist and things had a very specific emotion of property so I’m not using a lot of this generality in this talk but the situation here is very general in a way if there’s one but but still they introduced this idea in methanol algebras um that we’re going to pick up on so nfm algebra is an F algebra so these are these helpers that describe the data structures and but we don’t just want any of that Valtor where we want that sample taneously what’s called an m11 more algebra so and it’s all going out here and I’m not going to tell you my remedy under more algebra is but before I do that let me give you a little bit of intuition so the idea that I want world was going to describe how to properly incorporate effects of type M into into sometimes of time of a phase the type of technology and one or more algebra is and the f numbers go ahead to the pure parts related the separation because they have toppers can handle impure farts on our data structures and the end of the war on the rest are going to handle in vehicle parts and they’re not only my garage occurs are not only going to handle is that progress with a guy handed them in a nice way so we’re going to account for the possibility of a lack of effects through preservation of return of the Monad and they’re going to account for the potential merging of offense through the preservation of joining so they going to handle is the mice white so what are these things so and then under more algebra is just an algebra it’s just a pairwise RF after

this work right we have a carrier yeah and we just oh and I’ll go now that we don’t just want anyone we want to be able to preserve a return enjoying any one sense of described in these community diagrams so we just want him to we want our edge which interact I stood with the Monad operations and then just as before we needed the idea of enough algebra homomorphism here we get the idea of an empire go for algebra homomorphism but that’s just say that algebra homomorphism so there’s no underwear pneus here you could say okay so just have a an MLG Rahul morphism if I think of the more that Anderson help drive all right anyway bricks yes where’s this iconic form at the bottom of this oh sorry that is my l so at F&N of algebra it’s just going to be a triple it’s going to be a carrier and it’s going to be the f algebra and it’s going to be the end I with boolean algebra but again not as usual hassle i don’t have a way to enforce these on this these extra requirements but you can see that i just got a triple I’ve got the algebra I’ve got an antigen that I wanted to be in the minds of my and then I’ll talk about we’re home oops that a homomorphism is between these and it up just as before it will be a function that’s going to interact my sister I need to have something that’s simultaneously AF algebra homomorphism and in and out of a homomorphism okay so i just want to sort of do do the effect will stuff I’d be pierced i hope i don’t want to do it in parallel right kind of want to do them separately and it turns out that this is enough so what will do is I will think about having an initial F in a controversy that’s going to be i’m going to write that as we you and then f /em but i’m going to have to have three pieces i’m going to have to have the structure piece for that for the functor the fear factor i’m going to have those peptides a structure piece for the noggin and it’s going to be initial i’m going to have an initial fnm algebra if its initial with respect to F and also initial with respect to M and all i’ve done here is drawn these um Asia ality 4f and the 14 m and I just need to be able to satisfy both of these things simultaneously ok so I just want one structure that’s simultaneously an initial F algebra and initial amount of laundry and as before we’re going to have of something that is full we’re going to have to see knee cap by an issue ality and I’m going to denote that as bananas KL so if my algebra is a que el the K part is the algebra the elves the amount of morale dobrze i’m just going to write the nanoscale that ok so that’s what i have going across the bottom here this is life no you need no and of course we get a proof principle um just my initial a litigious as we did in a pure case so in this case um what I might want to do um if I want to handle with X is be able to show that something again something is that it’s not initiative that I’m sort of simultaneous old i guess i think i could say so i wanna show in a function G simultaneously um is this whole this kind of simultaneous full than what I need to do is show that a simultaneously enough algebra homework ISM and an M algebra at homomorphism because that’s the that’s the requirement the hold is supposed to be the unique one going out of the out of the engine and if that means I have to satisfy both of the equations of the one birthday I fell to run one from the mi amor algebra and this gives me a proof principle that’s very similar to the one that I had in the pier case ok so again this is a bi-directional principle and i’m only going to use it in the direction in in this talk where I want to show that some other function that I have is this whole and I’m going to show that it satisfies the quantity to being full but the interesting thing here is that what happens is you can see that I pro publications are split so the top this line here when the hand is this is um this is just describing what happens to live here in the in the pure setting right this is just a pure part of the effectual data structure so um ok so that’s pure so that’s nice this second line here this is the front the handle state attacks so what I’ve got now is I’ve got two publications rather than one that’s not too surprising because I have a fat tours before I didn’t so but

um but I have a nice completely split the food congregation so you can see what’s what’s extra that’s needed ok so if we go back to this is just a recap here at this top line we just go down and think about our effect for Liz let’s say then we’re going to represent that effect for this as the carrier of an initial well the pure part is the soul is def a and M on algebra so the pier course going to be described by list fa we’re going to be talking about this that have a data in them and my dad is just given by M I’m not even my to specify what it is so I can think about my federal this as being carrier of that effin and algebra and um here I’ve just it just to be more can continue I’ve set with the structure naps are so the structure amount interface is just going to take a pure structure and return its sort of pneumatic analog and the structure map for the the amount of algebra is just going Walter fits essentially join if you kind of get with the list constructors it just is just a salt like actually looks exactly like them so I’m going to represent the data structure in that way and then I get a fold that’s my unique nap on the initial fnm algebra 2 to any other FN I’d algebra given violets manl so if I if I think about that foal this gold you take as input a map k that knows how to handle the pure parts of the data structure and it’s going to take a math down that knows how to handle the balloon a departs and then it’s just going to them in turn is going to use KN 0 to just sort of interview throughout the data structure I’ve been using exactly the same pattern of recursion that we have in the definition of the data structure okay so let’s think about the success will depend again now that we have now we have another preference of old learning that is able to deal with maybe a little bit then through the facts that’s what I’d like to show you so I’m going to assume again as usual that I have be in charge over and I’d like to define in fact full intent and you know my effects will append is going to be defined in terms of this move all this kind of combined fold and it’s really interesting because um if the definition of the federal pen looks just like my definition of appending the pure case except set of fact that I’ve got an extra argument here for um where is it here at this extra argument again but if i look at the Pierre part this is exactly the same except for the time okay so things are starting to look very similar to how able to make your case and you don’t see this kind of infection of the this infection of the effects in with the victor parts of the data so these are the only two differences there agree so um like I said I’m except for the times this the function are taking the folder is for the fear part it’s identical to this to the the one that we used in in the entirely pure situation and that gives us a little bit of hope that we might be able to reuse the reasoning so again if you want to use this principle will behave will proceed in the same sort of way we did before we’ll end up with some equational properties um for our acceptable event and accept the types these are active equal to the work that we have in the pierced and so death that’s nice right soand then remember before we had um we had F naps and joins and so on when I was trying to use the pure tools and the effectual setting so here we don’t have that anymore in addition I know that um I have um that for any fixed why I have that depend well something like a pen to basically is um isn’t in my number for algebra homomorphism so I get a third property which I get the syrup i’m going to use in my proof so i have my two properties that are the same as in the pure case and i also have this third property that i can use and up focus just a short note about an interpretation of this appendant says that if i pretend always evaluates the effects that are placed before the elements of the first argument that’s what that line says so here I stage as activity of the federal pen again that looks just like it did before of course a noun step using purposeful 15 in principle to so i’ll try to approve of

this equation so this says that my um my effective oh my papa showed her from tens and disorder is going to be the same as my competition with natural transition times in the other order and what i need to do is show that um if i’m trying to show that the thing on the right hand side is a hole that i just have to show that it satisfies the property them hold satisfies so okay so i’m just exactly in the same situation it was before oops except that now i have to perform I gatien’s and I’ve written up there’s the pyramid that’s the one that comes from the F algebra and I’ve got the one that comes below from the more algebra so I’ve got to prove obligations if you look at the top one it’s exactly the same as the equation that we have to share sometimes it’s exactly the same as the one we have to show in a pure case so now I can just reuse my reasoning I don’t have to do something that’s kind of similar to what I did before and I don’t have to manipulate joins and maps and returns and so on I can look just just exactly the same reasoning I can reuse my perfect so that’s really nice that’s part of the modular again after the proof of the second light of the proof the second want the paper publication for the dynamic situation well that’s just for mirth long so now I’ve got a much shorter fruit um I think I was able to reuse the pure part the reason for the deer park that’s i think is um is good for the monadic with navigation well i have to do a little bit of work but i don’t have to do very much markets only four lines of work so so end up in a situation we can actually reuse or reasoning in the pure case in the neglectful setting you just have to be careful about how to how to do that the result is that you end up with a simpler shorter and i think this is actually very important a more intuitive approval than the then we got when we tried to leave use the pure tools and interpreting i don’t think it may be surprising that we end up in a better situation but i think it’s nice that we do and we end up with something that’s foreign to my baby okay it’s probably good to say that there are some limitations on to this i’m not gonna be able to free things that aren’t true for example if I’m trying to share that a couple of notes you have tried to your ship things that are true i’m going to have trouble so okay that’s that’s good right um so in this kind of situation I would be able to prove for example that if I appendix to Y and try to reverse the nuts the same is a five reversed into a visual lesson listen then append the results and this makes sense intuitively because of course the left hand side is going to try to do the handle the effects in X before swine on the right hand side has been trying to give in the opposite order and well the defects may not be communing and that way but what’s interesting is that um the techniques I’ve shown you allow you to see exactly where the problem lives by technically and if the problem arises that when you’re trying to show that um you try to show these this fun run out of obligation you’ll get stuff okay you won’t be able to won’t be able to formalize it so there’s a pleasure you can actually see formally what’s what’s going on and how you can formalize the intuition that we have in this type of below plug there um just done with three more slides I think to go back to the beginning of the talk um we looked at these effects will list write these books a bit early with on termination and I hope I’ve convinced you that you can you can do this by just using this sum this list were you explicitly interleave the the non termination effects right using this data type that I called the slicing and what we do receive that Miss lazy is just this initial um with F a and M algebra I’m Vicki my closest containing dative tarde going back to the other example of the one with the file handle well here i can think of having an initial list fa and I Oh algebra so it’s just infection or different hairstyle this and you can use that to actually rewrite age get contents in such a way that the intervening is explicit now and then this solves a lot of the problems that we that I mentioned at the beginning of the slide so now the eyewear is a report within the scope of our actions rather than some seeing interest if that appear in pure could we have access to the IO monad lyrics which is opposed to file and so on so I think that’s an improvement um to convince you that I’m not only looking at things like this I think I lied about the response may be in before um I need to do I want to talk a little bit about how it made it a study seniors as an example with this so this is love or like this way

off I don’t know he’s here but here we can think about the leader type also as an initial on fnn algebra here I have I am a reader type got some effect described by Eminem and it results in either the value of time fee or it’s a request for more data of typing and I just see this as an initial on fnn algebra where F is the sort of obvious of this filter that corresponds to the date time and here you could use purposeful to to reason about that mate if you want it you can eat that up a little bit and talk about pipes instead of just it or a tease so here we have the centralization of the pipes library so so what we have is just a proxy type and that’s just dirt entry of requests of bone here Todd a prime breeding values of time na i have values of type of the beam intervals three values of type B Prime these things are interleaved with the facts of type M returning pure values so I’ve got a much more much greater much more complicated on data type so this builds on the leader data type because it’s a possibility of my direction with requests and responses but the idea here is basically notice that this is also an infant in an inch off in an algebra and that means that we have a way of reasoning about them i believe that gabriel gonzalez was reasoning about them is some kind of loose form of co induction that this gives us another way to reason about them ok so to sum up um the idea of the talk is that an ethnographers are just at the wrong level of abstraction for a into of reasoning about data and early birth effects you can do it but it doesn’t it’s not as nice as what you would like it doesn’t have a nice operation of pure parts of data and effects of flutes against over examine of algebra generalize um again too I want here just two categories other than CPR that’s true but we also hand over general functors again there were restrictions on their doctors and they were trying to properties to have a very specific national property so the work the work in this top is actually based on a more theoretical and categorical a version of a paper that other lasses all three of those restrictions and also allows you to talk specifically about induction principles okay so we’re able to generalize in many ways FML chores are the effectual analog of RF algebra so everything that went on in the pure case with that algebra is going on here with with fnm algebra in the situation and again you can reuse some of your reasoning so that’s nice separate pierre de facto concerns so you can just transfer your core principles on your proof techniques other effects with data clubs not just lists can be expressed as initial F in and out reviews there by letting us use these techniques and time for a few questions jumped yeah a quick pair that i’ll ask together that’s okay so the first is when you were trying to prove that huge tivity will reverse and if you also had a property about safe huge tivity of the Mona could you combine it so the person then my second question was about the example that you came back to the beginning with filehandle normally that start by the effect is done based simply on demand I couldn’t see how demands got reflected in the second one wondered whether just evaluation demand over there or whether it hooks up some other way no okay so right so if you have other properties like I’m in the in the in the situation where you’re trying to prove this property that I said you can’t okay so here of course if you have other properties your disposal well you know that might do but you would actually need to show that you still have no more more more algebra so if you had other assumptions that rely to prove that that would be fine and yes I was just that I understand today so if you look at the types of the the in and the out of your effective data types and then the two witnesses to be either morphism and we see now both a propulsive earth gotye alone at the time but it’s the is the out that has the

Monad in what seems to be the obvious place so if you take one of these effectual data structures and observe it to see whether it’s an ill or cottons you have to first do the be the end action to get that far so that’s where I which is applies the arrow that should’ve known that end that in is not in behaves a bit oddly I don’t have a good intuition to what’s going on there if you’re just looking at the type of it I wondered if you could say something about that and if if there’s a duel with this final car with your story and which way around is the sensible half of the isomorphism for that one right um so there may be the thermally I think there should be in contrast I and with connection but I actually didn’t even talk about induction explicitly here of course so I think there should be one but I come any details um what was the beginning the inman apps leaving that yeah so you’re right that out in I mean bleep off yep and that is where you know that is really where you see the mode I’ve reflected yes I and what did you want so so what is there in doing then computational intuition for what’s going on there I mean I don’t have anything more to say other than it’s just reducing the mountains you know I don’t see that there’s anything there isn’t anything especially no natick going on I mean that’s what you’re saying it’s just it’s just an algebra right and so it’s not it’s not taking an effects into account at all I think that’s exactly what these go wrong sort of when you just try and use the under standard techniques and in the obvious way because there isn’t anything else precisely because we’re seeking one last question so you give examples of inner peace and pipes and everything work out very beautiful in that’s where that is been um you know any examples that don’t work out so nicely and you know why they don’t work out or you just not have any tables well you know there are once in effects I can’t be described my dad so here on I’m only restricted to a very restricting to the very specific signing not kinds i mean general specific setting if you know what i mean but anything that fits exactly in yeah so this is completely generic over the bunker and scullery generic the moment and it’s actually very generic as well over the the notion of property so I didn’t really talk about this in in this talk for just American and all those dimensions but it’s not enough of everything in the world okay but 30 minute break til 1030

FIRESEAT 2011 Presentation 7 by Elizabeth Blanchard

so thank you for having invited me today on my presentation is with use of martial master so foreign languages k choose another fire safety application so it’s quite difficult to stick just after your knees and stiffer because they think so many things about you will see that we are other objectives we are quite default approach fine I hope it will be interesting for you under we are not present a massage and it appears by CS DBS of which were within normal dragon siege by laptop like virtually of the french national software or scientific research and clothes smell pee pee participe so my presentation is divided in four parts first I introduce new complex visual directives also simple mouth parts days with the Wizards obtain a different scale very last item b2 few perspectives so first the world detectives since couple of years since 2004 cseb as being us to dig up a fire in cooperation with special so much in some domestic you languages because of those eternal 25 years of having cure in eternity for one BK we can cite France from the most tragically favors number Simone optional equals the 39 des and suburban energy infrastructure ask the fan what’s million say paternal fire as very severe consequences young and financial consequences because of its geometry the smoke is maintained in the channel to smoke and the Tony Awards also participate to the combustion and they also constitute a sweat for people of firefighters subaru is a tournament elation the tanami is always modulation due to either natural gas move more in the channel induce Constance either so the anchor lumens of due to the mechanical ventilation and last photo show returns wait when if they are chosen chiller theater if it concerns the passenger car attract or or even illegal to be a cure the entrance and policies China of the items are very high calorific potential so if you look at MK requirement for world to nurse they are on here a vote for one decade but also it is looking particularly French authorities and friendship operators are still looking for new ways and new system for ensuring your house of tea room and the exact context and educational systems I lost searching and Alonzo or watch on the system and so it’s difficult we say or what time is this gonna start to white spaces more pleasant walk outside and you should look at Alicia the water spray is considered as in water miss the National person some water easing books with favor Douglas once cells and make commissions so if you look at the phenomena which are involved when the water is playing a player moment first jealous of face creams analog of each is transferred from gas and see face to water on winners 12 X our soaking up awaited I’m eating an elevated yes also face it’ll be our so cool diploma similarly when its valuation in particular for water knows what our sample of targets interact with similar variation the net as a attenuate so when they shall by absorption and scattering effect to go oxygen inertia when water therefore when water droplets enough away they create a rainbow and because of severe I want to make freak your expulsion way to eat the youths user stresses and inviting close of a poor and the oxygen last into it is interesting smoke when whatever guitar playing in bayonne animals interact with a smokier due to the dynamic order to the gasoline for designing whatever system no in the channel if you wanted or something does not love to to design

the system and to assess it based on user on your lip on the only basis of its own characteristic of the term character for assessing urban design which is design now you must tell me out but we are estranged this campaign as lecithin say it is very expensive it is very difficult to comment on the database is even difficult Remini’s so the objectives of our research is to improve our understanding of the corner and we come to find them objective is to evaluate the generality of computational tools for Brittany color change of strategy which is to determine the potential contribution of computational tools for assessing rottenness system and we chose in thousand two thousand eight new year called EDS the well-known of EDS iodine and simulator developed by least incorporation disability for screen mechanisms first it is sweet and it’s an open source it means that you can look at the sources what is done exactly you can know the university results you can even change a model white is a silver lake is widely used by the scientific community in other words there is an important feedback in the scientific literature and in particular in the channel configuration and we are lost there was already a water spray model included in the quote when I speak about the waters Cornell I mean what good would inform searching rockets in the computational domain wouldn’t for transporting droplets order for pitching and the way season and if you can see at the sky you can win that this car is the block for serving practical coyote buttes and we considered in 2008 subchannel fires and particular efforts so who use it directly at the time scale and we were quick you clear this long because we are the database available in fact the chest pain was conducted between two thousand four two thousand see you no more tuna as a scale ratio video or 11 so it’s a it is a little present there it is 40 scimitar wrong it is modulated matter to Denali thanks to a population with the same which is more teeth as a dog website developed which crackin what’s become fundamental minion waves as priority is always located insisted each other’s 70 meters from the up-train extremity and this take a screen magnification of Troy and home pages for this recipe we consider that as a visa we are feet flew in the child and eat alone it’s a disk when chip the chairs are quite cheap so these yourself is all this stairways you’re a lot to to install mason sauce in the child and to characterize the file via wats in most particular in at several distances from the fire location and in this campaign we uses most of and more than 200 song cells in the journey and the sort also many committees among some gas temperature at several locations like better of train collector amplifier for 8-12 24 nature’s at nostrand invader yes no city wide legged fix our guesstimate in situ its way to the fire by two approaches first my wedding soon of Fire Lord unsaved only by misleading the gas composition of matter 22 mature adults to fly occasion and waterless system was installed in the channel all the results were the same there were 55 / feasts winners or losers it means that each other produce produces fat wet and in a premium we study we got sucked each way can be characterized by a nitride roll my friends with min damaged iphone 40

manager and special comment one way to to 285 during sisters campaign are we come to teach 28 chest and the English is for search we are only used by just swings out what amazed modules and today I will only speak about searches from civil if you look at the 20 HS many many parties and if things are charging first multinational vision we tested the media manipulation wanting to be no confirmation big cities for getting the different population vision as it is one which is called the subcritical vision whereas there is a back is a bit of a clearing and is it on screen in the upstream side of the channel it means that there is smoke earlier which the cultural flows out of the children and there is a masa which is called super critical information visioning when I also smoke which is produced by Sofia specialize in only one side reception which Angelo Sosa file out we tested many many iron also to good creeps are at that pool and another crimson activated order to more Chinese and the particular distance salon it was a complete is a supercritical modulation regime with a little pool and six months three mothers proper located after the fire on swings are located through the am the approach we use for Moses researcher consisted in two stages in two steps annexed I one posted that and after competition well we first had ejected the curb and after also validation whether we use the code for better understanding the situation from country hiding five sons of anemia and for understanding the interaction between missed application and foundations of teleportation so today I’m USB cable protein is a validation and expensive as the expensive use of the word for chess 20 serum for simulating such best such an investor concern many for kind for aspects Percy tolerate said it’s dementia of some uncharacteristic for pit Wars sonali is relevant alaysha it is that the sites to evolve volumetric flow rate which is defined as the year don’t win 8 20 of the charm so what’s wrong system and the last see it wins wait the file opportunity we are defines a combustion reaction and also see it please wait just explain ok they see it wins wet supply of it now so sign in whether it’s as usual the estimating it rains weight from is leaving it Wiz rate estimated from initial weight loss of frozen final load in green yellow it wins later estimated 400 oxygen consumption and in the black include data we define and as you can see before miss activation we the input data fits both cures and after a Whitley speculation it only fits mirror the curve for the oxygen consumption sense in your watering system is located as a job as a final vacations so droplets can deposit into full fire and it is certainly Phoebus book you cannot see your a version of the it was right before my situation it’s a way no no liquid or liquid fire so very fast final girls and after all develop and after an activation it is such a treat of this vest it does not involve a very high volume low water quantity Susie betweens wait does not actually speculation so it allows to students apply online ordering a couple minutes after mr. finish and during this vacation so for when you’re dating to cut our success campaign we compare the top or evolution of gas temperature of it fixes and of every city a different location and today speak about to gas release into your

differences so if so I compared the gas temperature at eight meter las 12 12 meters down stray and 24 meters downstream and from height of the articles in July as a wet curve corresponds to the highest measure one point as you’re going to jerk the responds to the global mobile factories in modern fiction so before this activation there is a general good agreement in predictor and in measurement if you look resources the slope apples and magnitudes are almost were collected if you look more into the Jays you can see discrepancies first the process of a small section is the better a QA c is the coop plans to under predict sleep gastropod shells very close to the fire location secondly there is a bigger super see incoming know between predictions and visuals at mid height of the channel this sleeve hype corresponds to the interface height between the smoke which is flowing two worlds in don’t from extremity and the fresh air it’s a mixing zone between small cans of freshman and this is due to the reason to the width and length after Miss activation already say precisely this bus is particularly interesting same situation does not induce and in very sweet and as you can see there is a important gasoline that’s any speculation and that’s for substances the gas temperature Iran become a roni roni from over HTML on section which illustrates of this certification the same of this notification of the environment after the couple of minutes after 420 civils where it wins rate increases again you can observe a difference between temperature measured in the upper part of the chair and in the local cultures of journal it is greater risk codification of Z on my own if you look at a prediction by EDS products pretty brown eyes the first decrease after this situation SI unit gravity of tommyrot shots and jealousy most education of the young male 1 no zip it fixes and surrounded rostrum supply location of personal materials downstream and if you look at the prediction it fits in the Madras uptrend and alternative a location after a lot before I left on this activation and you can see especially after activation there is a very easily drop of it fixes especially answer salamat all stems from the faith so now at the stage we were quite confident in our simulation and we try to use it to better understanding the situation and we made an extensive use of professional so we study relations and France fancy girls your capital Kofi yes the seller as a toxic photography Susie of the door between steiner and we compare here 045 vertical profile of the culture of gas composition and so in mergers dr. lang it’s a response to Tausha up to location yet metals in the weather and 12 watchers in blue in the loony glance corresponds to gas composition on the left true oxygen concentration and on the left on the way to carbon monoxide concentration and other testing the reserve eruptions hours oops our observation based on silicon measure out of just conquered of just for mature we saw a difference between measurements in the upper part of the channel and easily walk part of the charm and if you look at the vertical profile of totalitarianism anything else quadrant in temperature as these two distances and it will wait a bit and despite this obvious question in a burger shop the result uniformity in your gas composition and location so there is a very strong direct when we at Wendy’s system is activated services from drinking between

the seller and eclectic array of when mist is activated so we saw very sincere local measurements that list as an inference as just dr. Archer and it fixes and it appears based on silicon much more starts eating juice with gasoline and a reduction in it freezes so we did an energy balance from measuring for country files on this contribution if you look at it the energy balance with ultra Chinese so the each is released by the fire this it goes to his own gas was in such an this it’s got to eat such other places is transported who’s the openings what needs is activated it is transported by water aerobics or is up so and these are solved by virtual focus this Zener which rebalance before saturation is represented on the left and after activation to focus on TV on the right each color of corresponds to the ground the big falls in it which goes to its gaseous reasons of challan bring were truly clusters faces or wash to eat transport expose your burning issues ullu to absorb iron water droplets choose some or this contribution you can get the black dots and see fire it possible to do when those so first if you compels and black dots with the web dots your traveler urge you can yourselves up there it is conserved my integral which is not to be because in accuracy equation of a conservation is not result is not so explicitly so in sometimes it’s one of the criticism against a spoon and reform saturation we can see that half of the player it goes to inter surface and artists on projects with the opening so we can observe I wish of turn on wars in this situation and after almost activation series distribution is completely modified since it’s absorbed by your pets your present what present half of the pirated so under the photos it was clear Sutan also faces so texts that workers absorb a very important it’s in this configuration and it will and miss induce Cesar hydrazine equals two suitcases so now if you go for frozen we can astonish sweet phenomena first situated evaluations even is Alaska last two facebooking and we can see the coefficient which fill me in on energy it absorption by virtual voters and we can see on this figure that it is named yourself by your guests kalinga and after by your wedded continuation but under sofa screaming working online personals it also by emotional minutes but this results it’s just going in control is just opposite to the last resort railways save that Miss activation induces a high addiction in interest of places so to better understand result we plotted the it was two suitcases back compaction and languid election in swizzles of the Jail visit of suicide in the fare evasion and in the gorge myself and so welcome actions on the left by foundation is on the wife webcam response to the fire journal blue crab to the Apes transom and the wedding ring craft was it don’t friend zone and as you can see Miss activation intercessor i would actually what you people to suppress it into a higher version it meant that water droplets that are spraying this fragile act as a relative ship between flames and cello wars how we can see the salon results is that it is that convective heat loss work of Artemis activation in the job front side it’s because of the gas cream is all that it is its own main phenomenon in this configuration and because the smoke which is bring two runs the dogs cream soda is done as it is the addiction so lost two successive musisi so we shows that MDS

was quite a shelter as a transcript but what about the phenomenal escape from LA to escape and we stood in many cases pharmacist wrongful of we we try to assess which was invalid for transporting drawbacks which white process in order for it’s a last transfer I was a relative morning and today I will only speak about you whether the SS models whether to unit because n equal time for assessing the wider jamuna I mr. something clicked in my unit virtually system is quite for displaying me know those seven others this play is a deserve peace between a black body which is a me to the relation and the spectrometer and we compute their generation of the rejection by comparing the transmission was mrs Cuellar when users flip off and black virgin and spectrometers are from native red quick at different distances from 20 centimeter to one of the 20 meters and is it we can see that this is the divergence can just see you pressure is very low only four baths preparing to the channel screen where the operating pressures on 90 bus but we are still working with a Miss system sensors the center line metals all month on the world micrometers if you look at the prediction of the attenuation hour vs. the distance from the injection advance so from 22 songs matter to 100 centimeters so black that corresponds with the experimental boot corrupt with you’ve ever gone to a culture which is an email scott which movie from Africa road approach for some things near the wagon from frank question and the recurve response to fgs we can see that ed esters to interpret evaluation attenuation and this is Cooper see with majorities own search persons mr. Hart it is alone watching person so based on this evaluation we looked more to the drives in the world and in shall we if you want to compute the relative properties of the support of drugs you have to integrate efficiencies over also spectrum and reversal size distribution of these pockets in EDS it was assumed that this scientist official was not really a quote to the one used as the injection point so it was considered a locally it was even low advice with a local mean diameter or media was beauty in each gas sale and purpose of an education based on trees it extracted from the information suppose that this software of targets can be considered as as a soccer ball topes a model disappear sample of your pets whose diameter is a cultural or social innovator and it exemplifies the oppression so it’s allowed to me we wanted to assess this modification and especially because it went to see congressional time and we each the exactly the same well the same test I we separated exactly the same chest and listen sep teresa and important improvement in pressure and for 40-plus alter we the disco ball swing is in short is equal to Louisa wall to lose any person Sosa rally as a channel scale we saw that EDS was also able to predict the south on high amount and the gas flow but with so few discrepancies especially before the circulation which says that it’s a contribution of mist is important and that the main mechanism is a gas control relative attenuation but it does not mean that we stood up to injustice to face cream and we contacted the pressure soldiers to to say that e we think that computer shop Excalibur of candy people very good compared to experimentation from better understand the situation under certain condition if you know certain parameters at our fundamental programs we tried to assess few models

of watch Oprah moment of fgs and the some modification have been proposed based on this work no we want to continue our work from the god of the sources we want to characterize a situation and especially in Western Asia operating for sure I am fighters for watching to walk on the visibility both with and without fortune family we’ve computers campaign in the test setup turn and install it in a CSV consisting of more connected to al Qaeda and at the same could be on Thursday mysteries and rolls of quarters on pay attention confirmation

LIVE! SENATE HEARING Committee on Finance 9/14/17

the current rebuilding plan genuine policy power totally unlike home a transmission link at 16 kV you can supply the power from this area to that area a publisher area to this area so what you’re saying sir is existing even if there is a micro grid maker being a five megawatts in 13% isn’t closed but let’s limit the main islands lean upon to hang I need requirement PO I he gets an idea negawatt I don’t understand the system once of 13 how how do you measure that means 13 13 plan up when you see me so consumers paper or anything behind the third person she needs a whole opinion laughing couple years he’s a Latin for eastward passion beyond that you cannot collect that from the consumers I still may be because I am NOT an expert like senator surges mania in the energy sector I just want to understand the management and operation of the nation’s ability of manage this new transmission lines that you will be it on capex nonlinear for transmission line and then um every year will be from your operating budget even parity and Sulu let’s make Roslin 2.1 million as I forgot in the community linen and capex policy governor Sabra Islam determine mr. young [Applause] we’ll be in and Atari and the banner Basile and so these are considered but high risk yes scary yes yes your honor just to go back to your question earlier it’s actually policy question and connected to Nia to be honest about it you micro-grid can go as high as more

than five megawatts telega I’ve seen some some systems nominal 10 megawatts guy in LA the micro grid delegate 20 megawatts guy in LA and although in your case especially for sap 11% the vasila mommy so I think the strategy there really is to upgrade the lines know and expand the lines no because present says yeah but there’s no limitation to micro grid especially now the capacities are getting bigger and bigger but they have to work hand in hand with me and because they I will ultimately say in a con toda el Cayenne and I’m in my micro grid sohn jung-in de Panama will happen policy they have to keep on connecting and connecting and who you trying and a Polish not in C T no but it’s really a policy question of yes yes to be a listener I just mean ability to play don’t aim winner investments because honey nice I’m supposed to be opinion but you IP finiman the purpose of developing up rebellion yes II have a budget of 15 billion a year on subsidies for diesel is basically 6.7 billion is there a 15 billion a year on visa you should be nearly six point seven million give up with then is a contra some of our diesel generators in PC do we have to spend for this so James’s Mohammed on a per kilowatt hour and post that be so Jim well as an estimate sorry if there is competitive price non diesel non-fossil fuel and then we give it to them like in Lubin the mayor asked me if he came out with a solar company you would you should not be ready but I know it’s buddy

bless the Lawrence in the sun’s ball unserved areas back at the ODC the Renewable Energy competitive of course yes I mean you selected a CTO so pursuing NASA minimum it on NPP yes but they still get subsidies they still get subsidies the post more ancient church customer socially acceptable machinist Arvind I mean we can five to six spaces you knew proper Bible and production questions are given ten pesos for example diesel they were not charged

to consumers 12 only charged them let’s say nothing nobody begin enough of course a new PP in short and conquer in Depot enormous PSAs 11 pesos I began tuna tuna in a caustic ionamin lemon pieces vemana hindi neelam young agent must be gaining some cydia – I think that in our immune battle Kahneman Solomon P so yes ESP NPP doubling pressure we’re actually made subsidy a penicillin marami John banker at diesel in the back of it healthier way but are you and Sephardi synagogue Oh any administrator indeed that part that’s part of the areas that should be supposed to be penicillin but you must remain anonymous I understand that gallant to do in summer but then again when you contract this gensets going back to the course with us obviously when you buy this small box of mementos are you see em if you don’t it’s so open yes so for example I need five megawatts in Basildon yeah seven yes now

[Applause] security may happen chin region Amin rebuy panels installed Ian Warren Avenue father Shalini [Laughter] we’re gonna be testing the fossa and on you know there’s an 800 million allotment for transmission lines yes Ellen huddy Tom for Palawan 20 you portunity it in Wapping Paulo Bandito giving giving side panel for vinegar on a minion so Allah for Bella Bandito well upon us or being they take a neato 6096 in substation but I’m a similar and interested number three they only the six nine kV transmission line yeah you know you know substation dominated ionamin supposed to be finished by 2019 from the substation nearing completion on us so will happen in 2018 will happen allocation for a deadline for Pilar online for transmission line we write the P rosette I have to contract construction construction for the raw site I would be from from there will happen will happen transmission lines well a palapa these are all just right-of-way acquisition station muna suppose your engineering side right-of-way sorbet parish Alicia sorbet procurement in can you submit to us Cerritos Apple are online yes you sure can say we have a pending investigation and I think in line with mum Laurence opinion about micro grid I think there’s no more merit looking at yes micro grid in this type of areas Kessinger 770 kilometers Seamus a micro read by next week so any more autonomy may be similar on micro grid by next week I’ve been I think I mean our my cook I’m not sure I ask Energy Committee and Finance Committee do I understand it clearly

mr. Benavides no an npc marijuana can mandate but it does not plan to expand its the third generation facilities on your private sector in an open transparent yes neutral process to obtain the least cost to consumers in Depot not in apparel and implement appear on visa or bunker generator contracts that was meronym and private sector must if the price is lower than 6 pesos saying in the corridor was if there’s a private sector that could give competitive or lower cost more reliable than diesel or bunker or cleaner you will see some diesel Jensen Noreen Perry go ahead they should be appointed by the board yeah just change the daemon and also in mati in GM Ella penal attending pointed me on my private San Covenant that was direction in GM don’t hate on me or directors locally they hire locally income mechanism because it’s supposed to be a independent board that chooses in supposed to the US so just to recap from my hand or I just need your cleansing programs for palawan just you understand and then we leave you said just to connect young lugar that’s your honor telling what makes me into a t-shape from my endo 3.5 billion to connect to connect kasi pathologist do all the money

but until my crew grid melon bucket profile Amanat a clue hollows up up at the balloon the Magellan acquisition panna cotta got a gal a Ellen the cabinet envelope orienting and towel environmental a cool be a wave not the CSP pre bottom scepter micro billion from our budget you wait years position take so long we I’m sure we’ll be sitting here next year hindi parminder go again Korean TMC belt on I’m sunburned and our son be sent to all of these areas Eddie a wave number K importante employee singer in wave Nicole pian but we’re agreed hi I’m a baboon October 27 impression as a technical guys is it 170 kilometres coming for you generator support o Penang El Nido hang them Roja spell correct no in terms of grant from Puerto Sonny’s Rojas and Vicente el nido a Thai Thai any burn a battle and attack Nobunaga up at the billion P soul at Babel in a new packs a 200 plus kilometers now masala mo what dial may IP issues for attacks the Kampala one I am certain we will be sitting here next year Hindi papa can you work on that with Nia to convince the coop to just wave the area and Leggett on tanning one year this should be CSP of private sector mag Livan Livan salam ala had to do that instead of 4 billion from government funds we expand along Inc UTP meronym mechanism a just to expand yuki TP cerium Swiss challenge yunxiang problema the other thing that we can do is ask Nicole to sign a PSA for a per per pound so indeed our producers didn’t I Union Union didn’t need her aunty and so the work the waves you don’t need the funds we save money and the people will have electricity sooner to be pregnant they have different eyes I don’t have any more questions [Laughter] so just to clarify your commitment to

the gensets and a b-cell banker in the roomie I putting away mama terminate in the long term contracts Nana hepato pod on panoramas Malini’s Nomura okay here thank you Subaru owners at Miraval five minutes er my lunch broke from Inspira you

International Business Model Transformation: How this Company Succeeded [in 7 Moves]| Luca Orlassino

Wouldn’t it be great if you could take a money-losing business out of your international portfolio and turn it into a profitable one? Wouldn’t it be even better if you could extend its success to a dozen countries, with a very limited investment, in a two-year time? How would you define your objectives? How would you choose the right team? What would your priorities and approach be? If you are an entrepreneur, a manager or a business consultant and you are facing a somewhat similar challenge, stay tuned and you may get some useful hints The story I’m about to tell you is based on a real case. As such, it is not intended to describe a one-size-fits-all method Still, my intention here is to make you reflect on contextual similarities and differences with other real cases that you may be aware of or that you may be personally facing right now. My hope is that you’ll get at least some kind of learning from it And, yes, the story has a happy ending After long years of one-business, one-country successes, this B2B services company has started diversifying and growing internationally, by means of acquisitions. It is now reflecting on its achievements. In particular, the second largest of its lines of business is dramatically underperforming. To make it even more complicated, results are inconsistent volatile and hard to predict across different country markets Reasons for success or failure are unclear As a result of incomplete post-merger integration, more than 10 countries operate a variety of heterogeneous and apparently inconsistent models. Country-by-country market conditions are hardly known at a central level. Rather than an organically designed building, the business looks more like an assemblage of business models, having little to share with each other Although results may vary from country to country, the overall balance for the business line is heavily negative It’s time to do something Let’s see how our case developed from goal-setting to governing a successful international business The underperforming business line, as we said, is under the spotlight The company’s top management believes that, besides reversing negative performances, setting up clear and international business model rules and guidelines would bring in at least three advantages: 1) Providing a shared language and a communication framework, to quickly read actual practices and performances 2) Orienting local managers toward improved results, by setting clear expectations upfront 3) Establishing a straightforward and convincing cross-country positioning of the business, especially when facing large international customers Having these three strategic objectives in mind, a team is now needed to carry on a transformation project The company CEO is the first character of our story. She hires a new global head for the business line Let’s call him Jack to steer it away from troubled waters Jack’s vision about how things should work is clear. His solid market knowledge and his sales experience make him qualified for the job and a credible interlocutor of the top management. On the other side, Jack’s strong interpersonal skills will help him in winning business people’s confidence throughout the change process, at an international level Jack gathers a small and well assorted task force to participate to his endeavor. He calls Amy to join him Amy is an experienced, hands-on operations manager, coming from the field The team wouldn’t be complete without Bill, the person who told me the story in the first place Bill is a project manager and a former management consultant. He’s got a strong methodology mindset, a structured approach to problem solving, and a deep and diversified international background The first decision to be taken is about where to start from. Again, the company

top management will provide a key hint Besides being the largest revenue generator, one of the countries involved is also responsible for the largest net cash outflow. Total revenues even fail to equal the biggest cost component, that is personnel. It quickly becomes evident that this country is the priority Despite the widespread sense of urgency, however, the team doesn’t forget that the ultimate goal is international by scope So, before digging deep into their immediate priority, they decide to invest two full weeks in a round of remote interviews, with a sample of the other countries. The intended objective of the preliminary interview round is to quickly produce a first high-level picture of the main commonalities and differences across the existing business models. The team spends some time to prepare an interview guide, which they will use consistently with all the analyzed country businesses. Rather than just sending out a questionnaire, they prefer to hear stories from the live voice of the local business managers they interview. To keep costs low, though, they choose to do so via videoconference The sample of countries which participate to the interview round is chosen with a view to capture the largest possible degree of variability. One geographically dispersed, and one concentrated country. One mature, and one emerging market. One large, and one small local team The interview round allows the team to reach at least a broad understanding of the many different models being operated under a common denomination. They also get a comparative understanding of the main factors that determine differences As it turns out, some of the differences appear to be induced by objective environmental and market variables, while some others are the result of discretionary management choices at a local level. Then, as soon as the team has completed the international interview round, the first real challenge is to fix the prioritized country. First analyze, and then redesign their business model Before moving into design, Jack, Amy, and Bill make a thorough analysis of the ‘as-is’ business model of the first country. They base their approach on two distinctive pillars. On one side, they conduct deep, in person interviews, with selected key people at all levels of the organisation. On the other side, they use a solid conceptual framework to collect and organize results in a meaningful manner The conceptual framework the team uses is Alexander Osterwalder’s and Yves Pigneur’s Business Model Canvas, or BMC. This very same framework is suitable for both business model analysis and for its redesign. It is not my purpose here to give a full explanation of the Business Model Canvas framework. I’ll just point out that it is a practical and intuitive way to divide the key aspects of a business model into nine main components. These nine building blocks allow to describe both the front stage, or client facing, peculiarities of a business and its backstage, that is its functioning mechanisms. In a nutshell, the what and the how of value creation and capture After a thorough analysis of the starting point, the team use the same framework to blueprint the desired target business model. At this point, while still starting from the nine building blocks of the business model canvas, in a further effort to make results even more intelligible by a broader audience, the team then concentrate their attention on five specific focus areas. And these include: 1) the value proposition, 2) the organization structure, 3) business processes 4) the data model, and 5) information systems To make an example, let’s take one key business model feature: the decision to structure the organization into three specialized divisions. Each of the three divisions is focused on one defined and specific area of expertise. As a consequence, clients that need different areas to be covered, will have to get in touch with multiple points of contact This choice directly affects both front stage and backstage components of the

business model. In particular, value proposition, customer segments, and the way customer relationships are managed belong to the first or front stage category. On the other side the way people are organized, how sales and operation processes and information systems are designed and configured, as well as the underlying data model, belong to the second or backstage category As soon as the new design is ready the team works on the transition to the new model Change management activities take place, and the business is quickly reorganized according to the new guidelines Meanwhile, a new local manager has been hired to lead the business in the first country. She will make it grow according to the newly established model. Jack and his team can now proceed to the next steps of their international business challenge, and the very next one will be the definition of a truly global model The first country implementation has worked as a pilot test and a design laboratory for the new business model With that experience in mind, Jack, Amy, and Bill take the model, piece by piece, and revise it critically The team tries to figure out, at the best of their capacity, what would work and what would require adaptation to work elsewhere Different market conditions, company sizes, regulatory constraints and maturity are some of the elements that they take into account. Rather than a monolithic construct, the international business model is made up of three layers, implying a certain degree of flexibility The first layer includes the foundational aspects, and the core mandatory features that are deemed essential to the very identity of the business across all the countries. The second layer defines alternative choices as a function of known contextual conditions that may vary marketwise, and the third layer includes generic guidelines for choices that can be left to the discretion of the local business managers We can explain this sort of stratified modelling approach in more concrete terms Let’s take into account the divisional model that we’ve mentioned before. The core layer defines the global list of possible specialized divisions, their main internal organization rules, their standard names, the related management reporting dimensions, and minimal information systems requirements. The intermediate layer defines how the number and size of actually implemented divisions can vary from country to country, as a function of demand, business size, and maturity. And the third layer defines recommended growth patterns for division startups, and career paths for key roles in the organization. These are mere indications and can be overridden by local country management rules and local HR policies Having a model doesn’t mean you have to apply it mechanically and regardless of specific market conditions. Moreover, model flexibility is a good starting point, but even a flexible model must undergo a market test, before you can say it works Jack wants to act as serious value generator. His aim is always directed at the customer. He knows that the model, after its initial definition, is still just a theoretical construct, built up in a reasonably short time. The final international business model will be an emerging consequence of its application to the real world The team is now eager to test it in new markets, and a big part of the work is still there to be carried out Besides the initial pilot country, the company operates in ten more. These need to be aligned to the international model Jack and his team prepare a roadmap Priority setting, now, takes into account two factors: the first is the contribution of each country to global revenues, and the second is project logistics. In fact, the team being small and geographical distances being large, travel needs to be carefully planned. The model scaling approach is quite straightforward and can be roughly summarized in three steps: 1) engagement, 2) training, and 3) follow-up

First, Jack engages the Country Manager and the local business manager. He illustrates the content of the project, the reference model, the expected benefits, and the timeframe. They also identify the key people that will act as focal points in the subsequent phases. The second step is about training people. Peter, a guy from the IT department, has now come on board and is part of the team. While the main focus of training sessions is on how to successfully apply the new business model, Peter’s job is to explain how to leverage the functionalities of the newly configured information system in day-by-day operations. After the first training wave, and after allowing some time to digest the new way of working, the team delivers one or more follow-up sessions. The number of the sessions depends on the country’s feedbacks, and on usage data extracted from the information systems. The sequence is repeated for each of the 10 countries This modular rollout approach allows working with a two to four countries in parallel at a time. By now, the team is traveling a lot around the world Meanwhile, Bill, the project manager, has changed his role from an active part of the game to an observer. He is particularly interested in verifying the effectiveness of the method, its strengths, and its weaknesses. He is also working to formalize it in a way that it can be possibly reused for similar jobs in other businesses of the group. This is not a one-way game. Each new country implementation contributes to enriching the understanding of the business. With new experiences and best practices, the model depicts a living organism. One that needs to be governed As Jack Amy and the new entry Peter deliver their training rounds across the countries, the need to move on, from implementation to governance, starts emerging. Meanwhile, the first financial results are coming from the countries that have seen the earliest implementations of the new business model Following up the adoption of the global model involves performing a number of activities, such as observing and measuring alignment, reinforcing messages accordingly, sharing good practices, keeping an eye on market evolution, updating the model. The utility of the data model, one of the five focus areas defined during the design phase, becomes evident here. It facilitates “reading” the newly rolled out businesses, and measuring their results. Identified gaps between theory and practice lead to one of two outcomes: either the message about the model is reinforced and the local business fills the gap, or, if good reasons are recognized behind the gap, the model itself is adjusted, to encompass the new variant. Sometimes, this kind of adjustment provides the occasion to share a new good practice among other countries that turn out to be in similar conditions. Another factor determining the evolution of the business model is the observation of significant changes in the external environment. Far from being a mere attempt to preserve a new level of stability, the whole governance process is one of continuous adjustment for the better The above-mentioned governance activities imply intense two-way communication between the center and the peripheral businesses. It turns out that a common language has been established. Thanks to the very formalization of the model, the newly introduced vernacular proves a formidable enabler in model adoption A dozen countries can now produce simple standardized reports, that everybody can make sense of. In most cases, they use shared information systems and operate common procedures without major misunderstandings in their application, and without feeling penalized by the typical rigidity of centralized governance models. Three years after its implementation, the model is now generating a net positive cashflow, at an international level. Revenues have now reached one hundred and fifty percent of their value at the start of the work Countries operating the model for longer are the same that show the best performances. At the company headquarters, the top management has started considering extending the application of the newly developed methodology to other businesses in the group Let’s now review the sequence of seven steps, followed by Jack and his team. This

is easier done if we group the seven steps into three parts or phases: setup, execution, and governance In the project setup phase, the dedicated team 1) set clear objectives upfront, 2) hired a team with a good cross-functional mix of skills and a strong degree of cross-cultural awareness, 3) quickly gained a first level of understanding of the bigger picture, doing this with a view to prevent major rework in the subsequent generalization phase of the model. In the execution phase, they 4) gave priority to one country and did a good job there, 5) generalized the results, in the form of a formalized flexible model, 6) rolled out the model, knowing that, at times, this involves challenging business as usual. In the governance phase, they 7) initiated model governance, by following-up change, sharing good practices, and incorporating field-generated hints into the model itself. Besides the overall purpose of reversing the negative performances of the business line at the beginning, let’s also recall the initial objectives that have been associated to the introduction of the business model These were: 1) providing a shared language and a communication framework, to quickly read actual practices and performances, 2) orienting the local managers towards improved results, by setting clear expectations upfront, and 3) establishing a straightforward and convincing cross-country positioning of the business, especially when facing large international customers. Evidence suggests that at least the first two objectives have been definitely reached On the other side, evaluating the success on the third point will require a longer period of observation Now, retrospectively deducing the general validity of a method from one success case can be a tempting hobby. A self-convincing exercise in the first place and secondly a self-selling one As for me, I tend to be wary of those who try to persuade me of their theories, based on alleged cause-and-effect relationships between their actions and their successes. Reality is that there are always a number of plausible determinants of the success or of the failure of a business transformation initiative. In fact, modern organizations are complex socio-technical systems. We cannot reduce them to a set of linearly interacting parts. We will never really know which has been more important, whether the rigorous application of a structure method, the inspirational effect of a committed and trustworthy team, or the charisma of a leader. Or else, the self-reinforcing loops arising among all of these factors together Nevertheless, if I had to face a similar case, there are at least a few points I would definitely borrow from the story I’ve told you, and save for future use These are: 1) context dependence, what works in one place whether the geographical or metaphorical, won’t necessarily work the same way elsewhere; 2) dynamic adaptability, what works today, even if conditions look the same, won’t necessarily work the same way tomorrow; 3) centrality of language, before being able to fully communicate something, you need to give names to the concept you want to use; 4) market first, no business idea is good in itself or, as Steve Blank brilliantly put it in a 2010 blog post, “no business plan survives first contact with a customer”, and the market will always be the ultimate referee of your game. Finally, allow me to briefly trespass the border of opinion and belief, just to say that the effort of building trust and acting as a giver can profoundly influence the outcome of a business initiative, such as the one I have just described. I’ve told you my story because I believe it contains a number of useful hints. Every business is a different case and it would be plain wrong assuming that what has worked so well in a certain context would automatically lead to success elsewhere. With the necessary adaptations, though, you might find at least some elements of the approach that has proven successful in that context effectively applicable to a number of similar ones I can only presume that, among those of you who had the patience to stay with me to this point, a fundamental doubt may have arisen in your minds. The latent question

that may be more or less consciously puzzling you is: do you really need one model? Have I read your minds? Well, I want to be straight and clear: my answer is no. Expanding a bit on this, I would add that it depends on the context and on your objectives. The business world is full of success cases in which the model shrinks down to a few accounting and reporting rules, and a code of ethics, maybe On the opposite side, you may find companies that rigidly impose one model, without any adaptation to market conditions, even if this implies renouncing to a share of the business, in favor of a simplified center-periphery information exchange and understanding. Within an ideal continuum between stiffness and near anarchy, in order to choose your “whats” you will have to answer your “whys” Has this story made you think of some personal experience of yours? Would you be able to list one similarity and one difference with any of your cases? Feel free to use the comment space below, to share your remarks! Also, what is the element of the approach, among the ones described, that you would most likely want to replicate in your case. Please, use the same comment space to let me know Subscribe to my channel, if you haven’t done it yet, so you make sure you won’t miss the next episode. If you want to read the full text of this post, you can find it on my website: lucaorlassino.com

IPR Strategies for the Environmental Technology Industry

[ Silence ] >> Barbara: Welcome and thank you for standing by At this time all participants are in listen only mode During the question and answer session please press star, 1 on your touchtone phone Today’s call is being recorded If you have any objections you may disconnect at this time I would not like to turn the meeting over to Mr. Matthew Sullivan from International Trade Administration You may begin >> Matthew: Thank you Barbara I just want to thank everyone for staying on the line We apologize for some of the delays We just had some technical difficulties but thank you for joining us this afternoon for the Department of Commerce China Webinar Series Our speaker today will be Alex Koff a Partner and a Chair of International Practice at Whiteford/Taylor/Preston in Baltimore Alex will speak today about the state of our protection for environmental technologies, the companies in China and strategies companies in that industry can take to protect their IPR With that I’ll turn it over to Alex and thank him for joining today Thank you >> Alex Koff: All right, thanks Matt This is Alex Koff As Matt said I’m a partner and Attorney at Whiteford/Taylor/Preston We have 8 offices in the United States and some alliances worldwide My email [email protected] and my U.S. mobile is (202) 262-1197 Thanks for staying online despite the technical difficulties What we will do is we have a number of slides here We’re going to break up the presentation into 4 discreet parts The first part is an overview of the U.S /China relationship and I’m going to flip this slide so you can see the outline Part 2 is a discussion of the rise of environmental and clean technologies Part 3 is a discussion of Chinese imports into the United States and options to protect your business So it’s a flow from China to the United States And the third is protecting your IP in China If you decided to make the jump into the Chinese market, what can you do to protect yourself and what are some ideas in advance, as well as when you’re there And then I’ll conclude with a couple of suggestions And in general, slides zero to 24 are dealing with the China overall U.S. China relationship, 24 to 40 are talking about the rise in environmental clean tech, 41 to 50 are dealing with trade remedies and 51 to 63 are dealing with some direct IP considerations so, in total 41 to 63 is that part 3 Part 4 is 65-80 and conclusions are on page 81, contact information is the last slide on page 82 I’m going to also watch the time, if you have questions please, by all means, send them through and we can try and stop then and get them The slide that you’re now looking at, slide 3 the U.S /China Relationship, many of these issues are not new Sinophobia a fear of China is a big issue The theme, I’m going to move quickly through a lot of these slides as background for you The economy really is number 1 in jobs focus, and assessment and recommendations are what we’re going to look at but the idea is really the Presidential Elections coming up in November 2012 is really going to shape our entire U.S /China relationship I think And to understand where we’re going to go forward in U.S. policy, I think we need to know where we’ve been and what we’re going to do now is look at some particular slides that I wanted to point out to you the key focus For instance, I don’t have a pointer so I don’t think you can see what I’m looking at, but on the Committee On Finance page on this slide, Max Baucus is the ranking member on the Committee On Finance and here he was talking about trading rights, distribution rights, services, these things are quite important to him and later he was quite effective at instituting a change to trade adjustment systems if you can see at the bottom the note, 19USC2272 There’s an issue that’s of importance to a particular, powerful member in a finance committee like Max Baucus’ is, it’s something that will be reoccurring, likely will be going forward successful U.S. policy I would suspect China trade history deficit, if you look at this slide, I’m going back here to the history 2005, Senator Dorgan and Senator Clinton at the time was introducing measures to cap the trade deficit at 5% The third bullet on there, talks about the Companion Bill introduced by at that time, House Representative Benjamin Cardin from Maryland Benjamin Cardin is now a U.S. Senator and Senator Clinton is now Secretary of State They were both interested in the trade deficit issue and I think they are watching it closely I think U.S. policy is also something to watch on trade deficit issues The next slide deals with CNOOC, China’s National Offshore Oil Corporation

For those who remember there was a big off keyhole in the South Swell against China buying U.S. ports There is something called CFIUS review which is the Committee on Foreign Investment in the United States and even President Clinton was involved in looking at the CNOOC issues, energy, oil, clean tech, hot button issues when it comes to China at least historically And the next slide deals with U.S /China relations and it’s a slide that lists Foreign Affairs Magazine and some of the titles in the runoff to election when President Obama was running in 2007/2008, he wrote an article called “Renewing American Leadership” in July and August of 2007 It had nothing to do with economics it was mostly the war in Iraq and in the Middle East but the titles in 2008 and the actual year of the election, I took some of those titles, they are Changing China, China, and is America in Decline and then Handling the Economic Challenge From China So the issues all dealt with China and where China is going to India and Africa and nestled within there, there’s a question is America in decline? He was at the forefront, I think back then and even more so in this election today Obama and Trade Policy, he gets elected One of the first visits he makes is to Canada and he had a conversation there in Canada with Prime Minister Steven Harper and this was in February of 2009 He talked about NAFTA and NAFTA being different than China and saying that the NAFTA relationship has to be important But again, China was put off to a different category Next slide here is Protectionist Sentiments and so again, 4 years ago and it’s looking at steel issues and the steelmakers And if you look at the slide from the Wall Street Journal, it talks about there’s a rafter of steel trade remedy cases which are coming in and there’s a lot of China bashing going on at that time Protectionism Buy America issues, China Currency Practice if you remember in February of 2005 Senator Schumer, again a very strong, important Senator, he was very key and very on the democratic side for the Senatorial Campaign Committee for the Democrats He adjures the bill, a bipartisan bill with Senator Graham allowing for U.S /China currency re-evaluation in position of 27.5% tariffs This was in February 2005 because the Bush Administration had not gone forward to try and change Chinese policies as it dealt with those issues And all of this will be quite relevant when [inaudible] of the IP focus and I know that those companies that are here will wonder why am I getting this education on China? I think it’s going to be important background to understand if you’re there, how your business may be affected Some other major issues which we’ve touched on Consumer Product Safety, if you remember Mattel had a big problem with toys and a massive recall And then I wanted to talk a little bit about the State of the Union Address and Jobs and Exports The President in January 2010 when he was announcing the NEI, the National Export Initiative which is now about 2 years into it, said, “Jobs must be our number one focus in 2010” And he said we need to double our exports over the next five years And then he also mentioned clean energy a number of times in that same speech He said, “We can put Americans to work today building infrastructure of tomorrow, there’s no reason Europe or China should have the fastest trains and new factories for clean energy products.” He said China is not waiting to revamp its economy They’re making serious investments in clean energy because they want those jobs and he was looking clearly to try and put American’s to work building clean energy facilities And he said it’s time to finally slash the tax breaks for companies that ship our jobs overseas and give those breaks instead to those that create jobs right here in America This was from January 27, 2010, two years ago And if you look and understand the way Washington works and the way the power-players are working, again, this is Max Baucus that Chairman of the Finance Committee, he held a hearing and he said our most– this was on March 3rd said, “Our most urgent economic goal must be to create jobs.” Echoing what the President said “We must do everything possible to open markets and promote our exports.” And he proposed a 5 step plan First was export promotion, trade agreements, key markets, and he wanted and expects more calls on access in China, and so forth The trade balance is something that I think people are looking at as well, and if you look at the trade balance on this particular slide for those who aren’t on audio only, there’s a trade balance here that shows China versus the overall U.S. trade balance And I’ve just inserted a line which shows that generally the U.S. trade balance with China not being seasonally adjusted is flat

But the overall U.S. Trade balance has decreased since July of 2008 compared to today And a lawyer or someone looking at this from a policy perspective may say, well the U.S. trade balance with China has become a larger proportion of the overall trade balance and they say that the U.S. needs to fix that when it comes to China Jobs and Trade is squarely on everyone’s mind, and it’s hard to exaggerate how bad the job market is I gave a couple quotes, once in the Wall Street Journal that says 1 out of every 5 men, 25-54 isn’t working and this is across all the major papers Washington Post said well who is going to double their imports It’s just not clear and the New York Times said at that time many of the jobs lost during the recession are just not coming back Well what are the jobs right now? In May of 2011, a year after this, we still had 9% unemployment and this was looking at the U.S. Jobs Rate coming out from the Bureau of Labor Statistics And one year later, you guessed it, that’s still squarely and centered unemployment rate unchanged one year after that at 9.2% No U.S. President has won re-election with over 9% unemployment So that’s a big magic number that people are looking at Brought forward to just this past week, March 9th, the unemployment rate remained unchanged from February, but it had dipped and now it’s at a historically high 8.3% but with more than 12.8 million people out of work But it’s lower and it’s dipped below that 9% so things are improving but it’s still squarely in front of where people are supposed to be The slide, for those on audio, has a picture of a large horse, and a small horse and the slide says Global Economy, then and now And at the top is says, “Different Beds, Same Dream” That’s a play on a Chinese phrase that says, “Different Dream, Same Bed” Between a husband and wife who may be together but have different ideas on the world Here, what the slide is meant to show is that we’re all in it together and China and the United States really need each other to move forward, and then we’re going to talk a little bit about the rise in environmental technology now that you have a background to the U.S. overall China relationship In The Rise to U.S Relationship in Clean Tech, when Senator Clinton, who then became Secretary of State Clinton went to China for the first time, she said in February of ’09, President Obama hopes to make climate change the centerpiece of a broader, more vigorous engagement with China and she said the opportunities in China are unmatched anywhere in the world And she invited China to join the United States as she toured energy efficient power plants in Beijing and gas powered fire plants that sold GE turbines which she said were nearly twice as efficient as coal-fired plants Now take this, and what I wanted to do is show you how the U.S. Policy perspective works directly with how your business can flourish I’ll have some slides and I’ll run through them quickly given the time on Smart Grid Technology Smart Grid in the place called the Tres Amigas project for those who may not know about it is the electricity transmission project that a powerful substation connecting the eastern/western Texas grids in the United States, and there was a discussion about whether cold cables that are produced by American superconductor could help in that transmission line There was a bill introduced by House Majority Leader Steny Hoyer, who is also from Maryland, that we can authorize federal government funding to cover half the cost of high-voltage transmission projects Cold Cables move large amounts of electricity in a small space They do it with low-resistance wire through pipes that are chilled with liquid nitrogen that brings the temperature down to minus 321 degrees Fahrenheit So it provides more slippery medium for moving electricity than conventional copper or aluminum wire whose efficiency degrades as they heat up But the problem is it’s quite expensive to move it and as the liquefied gas cools, it requires pumps and refrigeration units every some miles to keep them efficient Well, if you look at who produces this American Semiconductor, they also have the international locations Korea, India, and Beijing as well as some other places And now if you look at it over time, overlaying what we said in the first section to what we have here today, you have a timeline from February 1, 2009 Secretary Clinton goes to China and says that climate change is going to be centerpiece of more vigorous engagement with China On the 12th of May, 2009 Steny Hoyer introduces the clean energy bill, on the 13th of October, 2009 there’s a number of arsenals that come out in the popular press talking about the Tres Amigas Project and asking is this the future of clean energy? In January of 2010, the President issues his State

of the Union Speech saying that he’s not going to seed the clean energy sector to these other nations and then in March of 2010, you have Senator Max Baucus talking about the Senate Hearings Regarding the Presidents Trade Agenda What I’m trying to say here is there’s a number of ways to read the tea leaves and you can understand what’s going to happen in the future by looking at what happened in the past So March 3rd of 2010, at the same time that Senator Max Baucus talked about his trade agenda, mirroring what the President said there was a bill introduced by Senator Schumer, that strong, powerful democrat talking about the limiting of ARRA Funds, the American Recovery and Reinvestment Act of 2009 funds because he said that reportedly of the 80% of the 2 billion are stimulus and renewable energy grants, they were all awarded for companies they presented that stimulus rewards for companies And he singled out the U.S. Consortium with Chinese Consortium to build a 648megawatt wind farm in Texas for half a billion dollars This is participation of China’s Shenyang Power Group with highlighted on Capitol Hill where you have the folks on Capitol Hill saying, “If we have energy projects, why aren’t we really giving you stimulus funds to American companies as opposed to Chinese companies Well, let’s go forward Today, in March 6th of the particular year CNOOC gives back and Sinopec is back Remember CNOOC had created a lot of luster in 2005, and their deal was turned away, well just this past week in the Wall Street Journal it was reported that since 2010, Chinese companies have invested more than 17 billion with a B, into oil and gas deals in the U.S. and Canada And recent investments have been positioned intentionally as “non-threatening” to get back into America And this is one of the CEO of Chesapeake Energy How? The Chinese have sought minority states, played a passive role and in a nod to U.S. regulators, kept Chinese personnel at arm’s length from the advancement of technology which moves into the IP States Fu Chengyu has been leading this push and he was the chairman of CNOOC and learned his lessons and stuff in 5 and now he’s the chairman of CNOOC at another large state-owned energy related company You also have other Chinese companies like BYD which is a name that now people sort of call Build Your Dreams; they want to be the largest Chinese automaker by 2015, and the world’s largest automaker by 2025 They were able to jumpstart their production and the company never existed, I think, before 19 maybe 96, they got their start making cellphone batteries for Motorola Cellphones in Cheng Yen, China And from there they got their jumpstart because they made batteries for cars and a new way to go forward with electric cars So this is a new way that China’s moving forward and flash forward to our own State of the Union Address with President Obama saying, I will not cede the wind or solar or battery industry to China or Germany because we refuse to make the same commitment here And he’s tried to back that up, it seems like things like ARPA-E ARPA-E is the Advanced Research Project Agency dealing with Energy, the House within the Department of Energy and a fund development of transformational and disruptive energy technologies and systems They want to find the new thing and focus us on a high-risk concept with high rewards and by law ARPA-E is required to stand at least 5% of its appropriated funds and technology transfer and outreach activities It’s looking like clean energy is becoming a national security issue and as things are heightening with the U.S /China overall relationship and clean tech being that next focus, and really where America has its drive It’s looking like energy is maybe that next real strong focus There was publication of the Federal Register that came out today, March 13th, that talked a Federal Interest in Environmental Technologies that the U.S. is pushing trade in environmental technologies and the notice set forth the schedule and proposed agenda for meeting of the Environmental Technologies Trade Advisory Committee called the ETTAC and that’s going to be scheduled on Thursday, April 27th, 2012, 9 a. m. Eastern Time in Washington They’re going to talk about globalizing environmental relations, issue related innovation, trade liberalization negotiations and it’s something to be aware of if you’re trying to take advantage of clean tech and United States technologies But the United States isn’t the only one to deal with this I’ve put together a list of certain technologies that at least our law firm has been involved with just

to give you a sampling U.S. Technologies that have gone from here to the world, Sludge to Oil Technologies in Brazil, Floating Wetland Technologies in Canada, Technologies coming from the World to the United States, we’ve been involved with Ocean Wave to Air Technology that came over from Ireland and Scotland and have been employed here, in Maryland and along the eastern sea border, and there has also been Chinese Technologies that have gone out to the world like Vacuum Tube Solar Technology which is actually a joint effort between the Germans and the Chinese, and the Germans have since divested of it and this deals primarily with hot water heaters and so forth and it appeared on top of the Olympic Village at the Beijing Olympics So this isn’t just only a U.S push as you probably know As well, the next slide talks about the International Energy Agency said that countries are spending $17 billion on renewable energy and energy efficiency research in the next 10 years Fifty-six billion was spent on nuclear energy research and $22 billion on fossil fuel research during the same period So there is more that needs to be spent on Clean Tech, but if you looked at Global Conferences, just a small sampling, Singapore has International Energy Week, and International Water Week, Copenhagen has the International Clean Energy Fair and there are a number of different technologies out there The U.S. isn’t alone in trying to promote the Clean Tech wagon So now we’re going to look at Chinese Imports coming into the United States and this where after 24 minutes, we’re going to move quite quickly to issues where you, your company can move forward and protect yourselves and what are the issues that you have First, if there’s Chinese Technologies coming to the United States, Chinese Clean Tech that’s coming to the United States and flooding the market at below cost sales or Chinese companies are using your own IP against you to enter the U.S Market, what do you do about? So, first is the trade remedy issue which is Chinese products are coming into the United States What do we do? Well this slide here is an overview that shows for those on the audio, the date from the timeline, from filing a petition in an antidumping or countervailing duty case, to issuance of an order and release The total time is somewhere between 10 to 14 months depending on whether extensions are granted or not There has been a recent case that was filed about photovoltaic cells in the United States and in 2010 the imports of solar cells from China were valued at $1.5 U.S. billion and there was a preliminary finding in December that the case should move forward and not be considered by the Department of Commerce We’re going to move from that and move straight into something called Section 421 If China is playing by all the right rules, there’s something called Section 421, filing the petition, until you get relief, 5 1/2 months Half the time in an antidumping case so it gives effective relief very quickly, but there are some tradeoffs compared to the antidumping The 421 Cases are unique It’s based off of the Global Safeguard Action but it was a type of deal that was negotiated as part of China’s WTO Accession Package part of joining the World Trade Organization This clause will expire in December of 2016, so it’s something that will go away and this slide talks about all of the cases that have been filed to date dealing with section 421 There 6 cases that were filed under President Bush, 1 case under President Obama President Bush did not give any relief in the 6 cases that were filed under his tenure I was actually Council and represented some of the Chinese respondents in 3 out of those 6 There is only 1 case filed under President Obama and the President did grant relief and it dealt with the tire industry The petition was filed exactly 3 months after President Obama’s inauguration and relief was granted in about September 2009, 5 months later In the 421 Tires Case, the next slide talks about the scope of what it includes and the relief that was imposed was 35% duties added at the border then it dropped to 30% in the second year, and this year, the last year, it’s 25% and it will end in September of this year just before the U.S. election Politically the USTR Ron Kirk, the Officer of the U.S Trade Representative, he said when China came into the WTO, the U.S. negotiated the ability to pose remedies in situations just like this and the administration did was necessary to keep American jobs, and the USW President said, “Well for far too long,

workers across the country have been victimized by bad trade policies.” And Obama made clear he’s going to enforce American trade laws and help people like us, and that was followed up in April saying business is improving, production has increased, but employment has not always been maintained The workers have been recalled to work, these positive results would not have been possible without 421 Relief If you look at what’s coming up next, in the State of the Union Address President Obama specifically cited the tires case and said over 1,000 Americans are working today because we stopped a surge in Chinese tires My question is understanding the huge political issues between the United States and China, understanding that there’s going to be a ramp-up in tensions as the elections draw nearer and also understanding that there’s 5 1/2 month window until this hits the President’s desk and is actually a decision that goes all the way up to the President to decide whether or not to grant relief will an environmental case, because that’s the next foreground I think in the battle ground between the U.S. and China policy Will one, be coming in roughly May 1 because if so, if you calculate 5 to 5 1/2 months later, it’s going to be hitting in mid-October, the President’s desk, just prior to the election So if you’re someone out there who is looking to deal with Clean Tech and you’re having some trouble either with IP or other issues, the 421 maybe an effective remedy for you, understanding the political environment, and I wouldn’t be surprised if one of these cases came in The word on the street, at least from those who understand the Washington scene and what’s going on there is that President Obama imposed relief here, but there was an informal message that went out and said, you know, I’ve given you this relief but I don’t want to see any other cases I mean, there’s no other way, it seems to me, to explain why more 421 cases have not been filed And I would be surprised if one did not come in the next couple of months Now we’re moving straight to the IP issues Overtime, from 1975 80% of a company was made up of hard goods and 20% of a company was made up with intellectual property or intangible assets In 2005, that percentage has been turned directly upside down so that 20% of a company is made up of tangible assets, brick and mortar, and about 80% of intangible assets That leads us to what’s called Section 337 and if you’re a company who has your IP stolen, and it’s used against you to import goods into the United States, 337 could be a great remedy for you It’s extremely fast from the filing of the complaint until the remedy, you’ve got about 12 to 15 months Compare that to Patent Litigation or other things in Rocket Dockets like the Eastern District of Virginia, or in Texas and Section 337 is much, much faster Section 337 cases are argued before what’s called The U.S International Trade Commission which is a Quasi-judicial Federal Agency with fraud investigative responsibilities in matters of trade The ITC remedy is different than in a district court, in a U.S. District Court before a Judge, you get money judgments You actually have to show how much– first you have to show that your patent was infringed and then you have to show to what extent your damages were, and in the ITC they don’t have that Instead what they do is they can exclude your competitors products from the United States and they setup a barrier the ITC has jurisdiction throughout the United States, nationwide, what we call in-rem jurisdiction which deals over the product that’s being imported and they send a note directly to customs service and then customs service will exclude the product from coming in at all This has become the new Rocket Docket for cases Apple, HP, Glaxo SmithKline Beecham, even in biological suits Zippo lighters for trademark disputes and even individuals, there’s a guy named Geoffrey Lee McCabe who made fulcrum tremolos on stringed musical instruments, he brought his own case as well to the ITC And the one thing to note is the ITC cases are argued before knowledgeable administrative law judges, people who deal exclusively with IP issues and I think the figure is that 1 every 6 years, 1 U.S. Judge, 1 U.S. Federal Judge will hear an IP case every 6 years, but the ITC Administrative Law Judges hear those cases routinely every day and I think that they’re hearing perhaps upwards of 25% of all the cases that do appeal intellectual property that defy the United States each year In the legal issues, I’m looking at the time and I want

to keep moving but there are three cases here that affect 337 cases and actually will push and funnel more cases to this particular jurisdiction The first is eBay where it was before the U.S. Supreme Court in 2006, and the decision in that case made it more difficult to obtain an injunction in District Court and there you look at Justice Kennedy’s concurrent In Kyocera, another key case in 2008 it overturned a longstanding ITC practice of downstream product relief and it’s giving more attention to limited and general exclusion orders which are type of orders issued in the case What that generally means is there has been a slight change in practice at the ITC and people are actually being forced to name more defendants and more respondents at the same time, or more selectively pick who they’re bringing cases against, and the last case, which is quite recent at the end of last year was the Tianrui case which dealt with trade secret violators and the ITC said that, the Federal Circuit said that the ITC may impose relief when a trade secret violation occurs completely outside the United States but the goods are imported into the U.S and they applied what they called uniform federal common law For those who are listening on audio only, I’m showing a set of slides here which gives you a sense of the trends for ITC cases I compiled data dealing with cases which are filed at the ITC through 2008 and 2009 and this shows that there’s been a peak of cases in ’84, it went down into a trough in the mid ’90s and it’s certainly on the rise now at the late 2000 and early 2010-2012 The next slide shows maybe the top 6 or 7 top 6 countries who have filed cases and then what I do is isolate to show you the number of cases that were actually filed against China, and as you can see there were, for those who are on the video, in the late 80’s and early 90’s there were very little cases Starting in about ’95, ’96 the cases started to trickle in and by 2008 it had ramped up like a jet engine It just immediately took off and China is now dominated in 57 cases that are being filed at the ITC And what I did, I took a look at the cases in China and I tried to make a comparison to where China is going to be with IP cases to where Korea was in the same point in their own development years ago I looked at cases in Korea from ’77 to 2007 and found that there were 60 cases related to Korea The two most important things to take away from this is that there was a sharp rise in cases in the last five years, and there’s a sharp rise in Foreign Complaints and I’ll show you what I mean The first is that in the case filings on the slide in period 2002 to 2007 were double the number that in the previous two five-year periods combined So cases started to really increase in the last five-year period for the Koreans, and you can also see the one thing that was quite interesting was in that same last five-year period from 2002 to 2005, a large number of cases for the first time were actually filed by Foreign Complainants and the conclusion seems to be that the Korean’s themselves are starting to bring the cases, they started to see that they don’t need to be the defendants all of the time, they can also be the complainants themselves And, I took knowledge and that basis because Korea has moved from developing to developed country has joined the OECD and asked when will China make the same move to High Tech production and will they be making the same, sophisticated choices that the Korean’s had been making? The cases have been filed against Chinese companies had been mainly in the low tech areas like Ground Fault Circuit Interrupters, Nitrile Rubber Gloves, Sweeteners, DVD Players and Recorders, there have been some exceptions with Semiconductor companies like the TSMC versus the SMIC, those are the very large semiconductor manufacturers; one is from Taiwan and one that’s from China And there is a move to higher tech products I wrote an article that was published in Managing IP Magazine; Will China Follow Korea to the ITC? And I looked at about 6 different factors I said the first is innovation, second is a long-term strategy, third are what lessons have they learned from litigation, fourth is how do the Chinese companies value intellectual property, fifth is where, within the company is IP being treated? Are the folks in charge of intellectual property and R and D having direct reports up to the leaders of the corporations and the CEO level and the C-Band? What are they relegated to outside non-reporting positions

and third is how is technology being treated? And in Korea they have taken these models to heart which is what you see often times in U.S. companies and they’re employing in their own companies China hasn’t happened yet, but I expect that it probably will, the question is how long it will take? The last point for you is that even if you’re here in the United States and this is an article from March 10, just three days ago, in the Wall Street Journal front page, you need to understand that the Chinese companies will try and come after trade secrets Many other companies are dealing with trade secrets, I don’t mean to just think about China, but that’s what the press is reporting right now And this particular case which is filed by the FBI marks the first time that U.S. Officials have filed criminal espionage charges against a State owned foreign company And those charges allege that the Chinese Government and company officials asked U.S. citizens to compile DuPont proprietary information used for the manufacturing of something called titanium dioxide and it was a long-term wide-spread effort to collect older information It’s hardly cutting edge stuff, but if you’re here in the United States, you still must be vigilant to try and protect trade secrets and try and protect your intellectual property even if you haven’t set foot in China, but you may be interested in the future So if you do decide to go to China, I haven’t seen any questions If there are any, of course feel free to email them and I will try to look at the Q & A slide, but it looks like we’re on time The last thing is how do you protect your IP if you’re going to China into our part four? The first is you’ve got to understand that the United States, way back when, was a hotbed; the world’s premier, legal sanctuary for industrial pirates Pat Choate who is Ross Perot’s running mate for President in, I think, the late ’80s, early ’90s This is after not Admiral Stockdale, but was somebody else, he actually wrote a book called “Hot Property: The Stealing of Ideas in an Age of Globalization” and in that book he did something which I think is relevant to those who are thinking about entering China He said, in this he’s talking about the United States at the beginning of our nation-building, he said, [inaudible] Congress wanted to rapidly industrialize the United States and to do so by whatever means necessary A practice we now call nation-building America thus became a national policy and legislative act, the world’s premier legal sanctuary for industrial pirates.” And he was talking, and I won’t take the time to read the highlighted portions that I have from his book “Hot Property” but I commend page 30 to 31 to you What he’s saying here, and he gives an example of how the United States and there was someone called Francis Cabot Lowell who is an author and entrepreneur, he went and he stole the plan for weaving cloth and as Pat Choate says, he became a hero for bringing England the most valuable industrial secret to the United States and at that point in time he was welcomed in Well, many view Chinese companies in the same way to try and deal, try and take any of the IP, and certainly in the Clean Tech field, the United States innovates These slides for those who are not on audio are the typical slides you might see in psychology studies For those on the video, what is it that you see here? Often times, people will look at this slide and they’ll say well, I see vase and I see a face very clearly The next slide is the one that looks like your IP Many people will immediately see the vase and they see the face as well, the vase and the face, but they do not necessarily see the two Spanish looking people, one on either side of the vase and on the right hand of the picture seeing someone come out of a doorway and on the left handed face, the earrings actually a bottle, tequila bottle next to them The issue here is do you know what you have and often times you may not recognize the intellectual property that you, your own company has, your own crown jewels and it may make sense to try and go to someone who is an attorney or go to somebody who is well versed in understanding the intellectual property that you don’t inadvertently give away something that you didn’t recognize was gold So the first thing is, understand what it is that you have The second is, today, the stealing of ideas is trying to not become a new premier legal sanctuary for industrial pirates Again, this is in February, in the New York Times, Ken Lieberthal, I studied with him at the University of Michigan 20 years ago, he has since gone to the White House, he’s now at Brookings, his routine, and this is straight from the New York Times article, “Seems straight from a spy film.” When he travels to China and he is a key China scholar, the article said that he leaves his cellphone and laptop at home

and instead brings loaner devices which he erases before he leaves the United States and he wipes clean the minute he returns In China he disables Bluetooth and Wi-Fi, never lets his phone out of his sight and in meetings, not only turns off his phone but removes the battery for fear his microphone could be turned on remotely He can exit the internet only through an encrypted password protected channel, and copies and pastes his password from a USB Thumb drive He never types in a password directly because he said the Chinese are very good at installing key logging software on your laptop and he’s not alone Folks have said only the top counter intelligence official at the Officer of the Director of National Intelligence has said if a company has a significant IP that the Chinese and Russians are interested in, you go over there with a mobile device, your device will get penetrated So consider, whether or not if you’re leaving the United States to go to China, you should have a company policy cleaned devices that are sent over and certainly if you’re in areas that deal with export control and ITAR related controls, you’re aware of non-export related issues and you can’t bring a laptop with you, but that’s something to be aware of Question: China and IP is it time to believe? I wrote an article for the Washington Legal Foundation called The Legal Pulse Tian Lipu who was the Commissioner of China’s IP office declared in December 2010, the Wall Street Journal an opinion article that China is serious about protecting IP rights, and the Commerce Secretary at the time, Gary Lock who is now the U.S. Ambassador to China, the next month publically lamented China’s lax IP protection and enforcement and the question is it really time to believe that China’s doing better? Well Commissioner of Lipu’s declaration was nothing new In 2006 he told China’s state-run website that more IP protection helps Chinese companies and promotes innovation But within a few short years of that there was lax enforcement and that escalated to a trade dispute that both U.S and China claimed victory in January 2009 This slide gives the front page from the report of the panel It was not appealed to the [inaudible] and it was left as it is Issues are heating up with the coming Presidential Election Remember where we were in the future and I think you’re going to see a lot certainly in the realm of Clean Tech, as well as in the sources for Clean Tech Again this is cutting edge as of yesterday and today March 12th the U.S., Japan, and the European Union filed a request for consultations with China at the WTO concerning rare earth metals and other issues that are going to probably receive attention in the coming months are auto parts, cars and other issues that deal with Clean Tech and also on March 13th the President signed a retroactively to revise the trade laws to allow for filing of anti-subsidy cases There are some improvements that do need to be noted For instance in March of 2011 the Wall Street Journal reported that there were software makers Microsoft, Adobe, and Autodesk, they settled some copyright infringement suits against a midsize Chinese steel structure engineering company but while it’s good on its face, it’s also important to put that into perspective, the settlement was just shy of U.S. $200,000 and that was split three ways So, although supposedly this does not include undisclosed financial damages that weren’t reported in the paper the use of pirated software is still widely reported and is still pretty rampant That’s something to take into account when you’re going over So the question is, is your company going over? If you haven’t made the decision yet, should you establish operations in China? Is that the question for you? Some of the things you should think about are well, many companies decide to establish commercial presence in China despite the idea of theft for various company-specific reasons It could be the size of the relevant market or the cost advantages associated with opening a factory in that location Some [inaudible] prove cheaper employment, less environmental restrictions, cost of raw materials and so forth For them the issue is not whether to manufacture, but rather how to do so in a manner that adequately protects IPR For others, this balancing weighs against going into China There’s a company in Gaithersburg, Maryland that operates in perhaps 50 countries around the world Bobby Patton of Patton Electronics’ has chosen not to go into China In fact he has removed his product from China because he receives as an example, a device for a warranty replacement that was not his parts It was in his outside container from Patten Electronics for the inside guts weren’t his and things were getting ripped off too rapidly for him But for some industries, for instance mining, econometric modeling shows that there’s no positive association between IPR protection and enforcement and that, if you think about it, that seems to make sense because whether to open a copper mine should not be bases on whether a country has strong IPR rules or enforcement but more on global demand for copper and location

of those deposits, and you go where those deposits are So I think your industry will affect and your particular company will affect whether you want to source from there There are other places that can also be just as affective I’m not promoting Singapore or Hong Kong or other regions, but there are places that I wrote a report that was co-sponsored, jointly funded by the U.S. Patent Trademark Conference and the International-Intellectual Properties too which is a think-tank in Washington and this paper was presented on the first day of the annual public form of the World Trade Organization in Geneva in September of 2010 and it looked at a number of things It asked, how do you stimulate innovation among other things? It looked at tax incentives, grants, research centering’s, clustering, IP resources, a strong Intellectual Property Office, ready financing and effective rule of law And we had termed it the Singapore Model because how Singapore tried to structure everything around it to attract IP Centers and Hong Kong and Singapore have been quite effective in doing that So if China gives you some concerns, there are other areas that could be a regional hub just as easily The other thing which is now a new trend is be sure to register and protect your IP in China because even if you don’t go to China now, a plan to in the near future, you need to register and protect your rights because others in China will squat on those rights China is a first-to-file system, you’re also seeing this in some other countries like in Korea for instance and it could be costly for you later if you then need to negotiate with the person who’s taken your company logo and you have to negotiate for that back Registering in the United States also helps with what we call statutory remedies which means you don’t have to prove that there’s been a certain amount of damage to you in each case Instead, by law, if you’ve registered you’re entitled to a certain amount or fixed sum for every day or for every violation and that could help to a great extent in certainly leverage for negotiations and certainly collecting as well, and proving your damages If you do go and you are already there, you should consider segmenting your production process What this means is that if you’re manufacturing in China, you can produce only elements of a product in the weak IPR environment and if– the example is for instance there was Firestone Tire, the General Council and I sat on a panel in Chicago in October 2006 and he was talking about how Firestone maintained their secret sauce for their tire production back in the United States and they had separate factories in different locations in China Each which produced a discreet segment of tire and those super factories had no idea that the other one existed So they would produce part A of the tire for instance Shenzhen and then they would produce part B of the tire in say Shanghai and both of those would be sent down to be assembled in the United States and added with a protective sealant for instance that was a trade secret that the company did not want to disclose By segmenting your production you can still take advantage of cost of production and other things that are important to you and at the same time maintain your intellectual property which could be a big deal You should also keep apprised of developing trends which Chinese companies are using against United States companies and other companies in general, and often times it’s used in retaliation if you’re a consumer products company or a company that sells to big-box [inaudible] killers Wal-Mart or Sam’s Club or something like that You should understand one of these trends For instance, in China many are downloading patent applications filed in the United States and elsewhere, literally copying them and then filing for Chinese patent based on the blatantly copied prior art and what they’re doing is this is in retaliation for U.S. suits that are brought under Section 337 that we talked about or for some other things, for instance The U.S. firms will be able to invalidate the patent often times and they’re going to get the rights back, but they have to file for them and they have to engage council abroad and understand what the process is And the problem here is if you’re trying to seek production to get into a Wal-Mart on a Black Friday in November, you have to understand that the orders are going to come in and you need to start your own production and you’re in shipping lines as far back as perhaps March or February in order to fulfill those orders in time If there’s a cloud on your intellectual property right and your competitor has filed and actually obtained a patent and you hadn’t been watching for that abroad and they essentially can forestall you until you can unwind what has been done, they will essentially be able to prevent you

from actually producing that product in that buying season even though you will be able to successfully complete your rights in future And this is something that I understand has been happening to folks overseas and [inaudible] with people So you need to keep apprised of what’s going on and also it ties into I think knowledgeable council who can help to forestall that and be proactive in policies for you So key considerations when engaging China; Number 1, make the decision Give serious thought about whether you want to make the leap eastward and go there I think cultural dynamics are going to matter If you are going to go spend time learning the unique characteristics of doing business in China, such as understanding that, whom you know, is going to be critical The 3rd issue is, proceed with caution Build relationships of potential partners, conduct your due diligence, identify those who know China and can help you achieve your goals And the 4th is; protect yourself Of course protect your IP, but understand that managing relationships based strictly on legal documents could be a mistake While getting it in writing helps, personal relationships are going to matter So think about what can go wrong, as well, and consider your exit strategy before you go to China Arbitration may not always be possible for all disputes because arbitration means that the parties must agree to arbitrate the dispute and when you’re enforcing IP rights against people who are counterfeiting or trademarking violators, where there is no contract, often you have to the local courts and arbitration won’t be an option for you and that could be expensive Or you could be before a local court that is less receptive to a foreign body China is getting better on that, but it’s not yet where it entirely needs to be according to many of the business contacts I speak with So how do you deal with them? Sometimes you need to be creative First, dealing with counterfeiters, the Technical Supervision Bureau in China has the authority and ability to examine documents, conduct investigations, and they can also conduct raid and do so quickly For instance the Shanghai TSB, The Technical Supervision Bureau make a wage decision in several days as opposed to 3 to 6 months to navigate a system to a formal decision And raids can shut down the competitors and be quite helpful for you You could also do that through the U.S. [inaudible] in Shanghai, the U.S. AMJAM and some of your other sources of knowledge base supply You can also deal with trademark violations through the State Administration of Industry of Commerce or the AIC that has many of the same powers in the trademark concept that TSB has against counterfeits in order of seizing products but also be creative and there was an article that I saw in the American Bar Association publication that said look, why don’t you call the fire department? You can look at child labor, safety, fire codes, and although these departments may not have the power to seize the counterfeit products, there may be an advantage to you in having them conduct a surprise inspection which disrupts the production line at the factory and could also result in penalties for non-IP issues So if counterfeits are also found during the surprise visit, they trademark or copyright an owner may be in a better position to lodge a complaint with relevant administrative body that does have control So that’s another suggestion when you’re actually in China So the final conclusions, is 4 Number 1, read the tea leaves; larger political dynamics do affect the U.S /China relationship and it will impact your business You should take advantage of those If you spot where people have been you can identify, I think, where the next cases are going to be, where the next funding programs will be from the United States and where the best pressure points will be for your business Second is, protect your IP and R&D Think about travel abroad policies, departmentalize your production, be proactive in patent enforcement, register your trademarks and copyrights with Global Watch services in whether your own council or others I’m more than happy to help Execute your NDAs, your licensing requests and so forth The third is, understand that you’re in China Negotiation is key, politics are important particularly if you can call on a powerful trade association, embassy official whether it’s U.S. or Chinese, local government contact to act as a pressure point for what is right That’s going to be important And the fourth is legal proceedings including arbitration to help your negotiation, sometimes, but they may not be a total fix You need to be prepared to compromise and you also need proactive to actively seek to avoid disputes That’s the presentation Our contact information is here Again, my mobile number is (202) 262-1197 My email is [email protected] and we’re also on the web and www.wtplaw.com I’m happy to take any questions that you may have and thank you for letting me present I apologize that we got started a little bit late due

to some technical difficulties >> Matthew: Well thank you very much Alex for your presentation, and with that, Barbara if we could probably open up for about 10 minutes of questions or so if anybody would like to ask a question to Alex >> Barbara: Thank you Once again to ask a question press star, then 1 [ Silence ] >> Barbara: Glenn Roberts, you may ask your question >> Glenn: Hey Alex, great presentation and I’m with a commercial service out here in California and of course, you know, China is just across the waters from us and, you know we have quite a few potential buyers coming through and one of the strategies I’ve seen to effectively kind of control you touched on a little bit today I’d like to find out from you how effective is the strategy in regard to trademark and branding? Is it like the first line of defense to protect your IP as a strong brand and a trademark? >> Alex: Did you have more of the question or was that [inaudible] for that one? >> Glenn: Not much more to the question, a little bit of the background You know somebody IP Technology is well over the head of the people who are enforcing such technologies and I think it’s kind of easier for our companies to develop a strong branding strategy and if you can elaborate on that and let me know if that’s a good strategy and a good first line of defense for protecting IP >> Alex: It’s a great question I also serve as a Vice Chair of the Maryland DEC, District Expo Council, the Maryland DEC, so I appreciate all the great work Commercial Service is doing and under Secretary Sanchez, was in Maryland yesterday at a White House briefing and gave a strong presentation, and we had a number of discussion on these and other issues Ambassador, Chuck Ford, also came to Maryland I think about three weeks ago entered our DEC meeting and it was great to see the support of the Commercial Service I think you guys do phenomenal work for companies out there If companies are not taking advantage; who are listening to this, the Gold Key Services that the Commercial Service offers or the bedding processes, you should be To answer your question Glenn, I was in Korea and this will relate to China, in February and I was with Commercial Service and I was talking with people in that office and they said please, please, please insure that companies will go forward and protect their trademarks, and protect their branding, protect their rights because Korea, like China, the first-to-file system and what’s interesting in Korea is the government officials, it was reported to us and I was doing this through a study as well that was sponsored by the U.S. Patent and Trademark Office to look at U.S. [inaudible] assistance and how that helps U.S Exporters sell their goods overseas What was important to us is that it’s interesting, the strong tactical assistance programs that were employed in Korea in the ’60s and ’70s and throughout, as well as through the UNDP have educated the Koreans and they’re full onboard, the Koreans are now proselytizing to the Chinese and what the Koreans reported back to those meetings which were at the Embassy, is that although in many first file systems, people can go in and register rights and squat on those rights In China, they’re actually doing uniformity and as a practice So in Korea, before someone has entered the market there may be less of a risk to actually setup your branding strategy and to register your trademark and to register you intellectual property rights in those foreign markets until you actually go in In China the reports are coming in is that the Chinese are actively going outside their borders, identifying some unique companies that have no intention of going to China for another 5 to 10 years and actually registering those rights today because they know that the folks will be coming to China in the future So I think it’s extremely important for a trademark and branding as a first-line defense to immediately register those rights in China as abroad because it’s such a huge market, and although it may be an expensive process in the class of service, if you identify what could be your most profitable in the future, that would really be quite helpful Another example is Under Armor which is based here out of Baltimore When they had limited budget, instead of pursuing patent protection or some other intellectual property rights protection versus to try and protect their wicking technology and other things for their clothing, they said grow all of the revenue and early development

into the trademarking and global brand building and recognition, and I think that’s paid off quite well for them It depends, I think on the market It depends on the company, but for many companies I think it would be square front-and-center, consumer goods really being that front-and-center one Does that answer your question? >> Glenn: It certainly does and it’s a great confirmation and then what is the average cost? I mean maybe a high ballpark side to register your trademark? >> Alex: It’s not a huge cost in general There are two components to it and I have the figures, our firm has standard flat rates Other firms do as well, I’m sure I don’t have them in front of me but, there are two things, the response comes in two components The first is you’re going to want to do U.S. Trademark Registration, as well as what we call Third Party Watch Services I can’t talk about other firms, but I can tell you from our experience, we monitor so many trademarks that the Watch Services, the Third Party Services give us a discount, like a 6% discount which gets passed on and I think it’s in the realm of less than $1,000 or maybe $350-$500 for multiple Watch Services So your company would then say, could you please look at our trademark as in this class of service for this internet extension and domain name and they will go out and monitor and produce reports that get sent to your lawyer or directly to you on a weekly basis to say, this particular trademark has been filed in this other country, in this other jurisdiction, this web address has been filed which is confusingly similar with yours So that you’re out there, actively monitoring and trying to protect you mark which I think is important That’s the first component which relatively nominal I think in the grand scheme The second, which is more dicey and I won’t give you a great answer for that, I apologize is how expensive will it be? That depends on where you’re going to decide your Key Markets are and that’s a very company specific issue If you say that I need to protect myself in Korea, in China, in the following Asian countries, or if I want to go to some of the Brit countries, wherever it may be, each one will have their own separate national fee because there’s no general international wide trademark that you apply for You actually have to go into each country and specifically reserve your rights and you have to do it by class of service, meaning I’m going to do something like a T-shirt as opposed to a software technology and you can register in each of those classes and there you need to decide carefully and cost effectively where am I going to get the biggest bang for my buck in terms of protection? Again, it’s not very expensive in the grand scheme if you’re going to try and do a development rollout and you’re going to enter another foreign market and you have a number of different things at your disposal that the [inaudible] can assist you with, as well as [inaudible] bank financing Many people who enter those foreign markets are scared by it but if you actually do some export control, if you do some export work and you focus on how you’re going to deal with exports the benefits driven to your bottom line will far exceed, or should far exceed the cost of registering your trademark >> Glenn: Correct and thank you so much >> Matthew: Okay, thank you very much I think that’s about all the time we have for questions today but thank you again to Alex and thank you for everyone for listening, and just to let you know of course feel free to follow-up with Alex if you think of any questions And a copy of this webinar will be posted on our website www.stopdates.gov if you want to listen to it or review any other parts of the webinar later Thank you again and have a good day

AWS April 2016 Webinar Series – S3 Best Practices – A Decade of Field Experience

welcome to today’s webinar s3 best practices a decade of field experience our presenter today is omair geelani senior product manager today with us we also have drew myers Senior Product Marketing Manager Carl summer software engineer both wom senior manager developers and student Chan product manager as our webinar moderators who will be engaging with you and answer your questions throughout the presentation omair welcome the floor is now yours thank you and everyone welcome to the webinar for Amazon s3 so we have a lot of great content today a quick look at what we’re going to be talking about I’m going to spend quite a bit of time talking about some of the new capabilities we’ve added to the s3 platform specifically helping our customers manage their data on s3 and get more value in raipur more value from their data on s3 will talk about a new storage class to be announced last here specifically designed for in frequently accessed data we will also along the way talk through some of the best practices that can help you improve performance of your data on s3 and also leverage some of the great security capabilities we have to make your data more secure on s3 will also talk about one of the new service we announced just recently which is three transfer acceleration we can help you move data faster over the Internet to s3 so with that let’s get started but I did want to call out we have a lot of great folks helping us out with the webinar today and I would encourage you to enter your questions and engage with us and we’ll make sure that we get to this question at the end at the end of the webinar and talk about so with that let’s get started so before we dive into s3 so Amazon s3 has is actually one of the first storage services that we did launch for AWS but very early on we heard from customers that not all data is equal data comes in many different forms in many different formats and F search if you have files that you want to store in the cloud you need a different set of capabilities as opposed to saving video video or objects and so while we saw a great growth and you continue to see great growth in Amazon s3 in clay sure we wanted to make sure that we think about storage offering formative years in a breadth-first manner and so as such we’re going to continue to add more capabilities to our platform and really we think about at a degree as we think about storage as freedom of choice right this is about providing our customer to the right set of capabilities so that the data is able to leverage that the leverage the right set of capabilities for that form format of data so if you have data that is in file format we had amazon EFS that can help you have a dynamically scaling file server offering we also have elastic block storage that you can use with your ec2 instances for persistent storage and then of course we have Amazon s3 and glacier which is ideal for object storage in the cloud scaling dynamically with high durability and high availability but storage offerings themselves are not enough right a lot of our customers specifically the customers who are kind of on their journey to transitioning and transforming their data I’m moving to the cloud ingestion of that data and ability to transfer the data from on premises or from there from their end users to the cloud is very important and we focus a lot on that we offer our customers the ability to use direct connect to establish that high bandwidth pipeline between your on-premises data center and the cloud we last year announced amazon snowball which is a device that you can have shipped directly to your on-premises data center connect that to your to your network and then ship that bad 228 in VA standing and use that as a way to ingest data into s3 we work very closely there I sleep partners to make sure that all the storage appliances that you know in the works seamlessly with Amazon s3 as well as our other storage services we announced amazon kinesis firehose a way for you to ingest streams directly into s3 so you can start building your data lake around history and in leverage all the great services and the ecosystem that we have to offer on top of AWS we’ll talk more about the new servers that we just analogies s3 transfer acceleration which allows you to move data across long distances in an accelerated manner and in just that directly into s3 in airport we have Storage Gateway again allows you to establish that connectivity from your on-premises data center to to the club so kind of in all you see a lot of choice up and in the focus really is you can make sure you have the right set of capabilities for your specific use case to have an overall and comprehensive

storage offering from either the rest so with that let’s shift our focus to s3 and before I talked about some of the new capabilities that we’ve been adding I wanted to touch base on a core customer promise at the heart of it s3 is a dynamically scaling service that provides our customers with extremely high durability our promises 11 lines of durability and high availability or promise of Android design four nines of availability to make sure that when you need your data it is it is always available it is always there for you so you never have to worry about ever losing your data and you put it in s3 it is available for you so as soon as you need your data you are able to get it using a very simple rest interface and of course and this is a big one too for us that the service is extremely scalable we have customers who talked to who have you know all the way from gigabytes to hundreds of petabytes all the way through exabytes on us and a big part of the value that we deliver is a completely dynamic scale you never have to worry about hitting limits or never have to worry about working provisioning more storage and that’s a big part of the value too is that any of the capabilities we add to the platform we make sure that it can handle s3 scale it can handle scale all the way from gigabytes to exabytes and you know customers never have to worry about that so let’s talk about some of the new capabilities we’ve been adding to the platform this is not a comprehensive list in fact there’s a lot of capabilities we’ve been adding and we’ve been very busy with this and so I would urge you to go look at some of the documentation that we have available there’s a lot of great talks available and for each one of these each one of these services that we launched we have focused documentation that you can go look at it should help you get started with this is over I’m not going to go in detail for a lot of these but to kind of just quickly give you a summary view of some of these we launched the event notification which is the ability for you to trigger notifications to multiple destinations when specific action happens in your data in your s3 bucket for example if you put a new object in s3 you have the ability to configure its 32 trigger notification to do multiple destinations including SMS SQS and lambda so for instance you can write a specific lambda function to take action when a new object shows up in in s3 so if I want to update a secondary index if I want to transform data if I want to transport a video file and so forth you can now use this event based computing paradigm instead of polling and trying to find out whether you have new information in s3 simply having s3 trigger the right end function will assess us learn or SMS and then use that to perform and proceed in your workflow a lot of our customers are doing very interesting things with with event notifications and you continue to make this platform better we added the capability to trigger notifications on to beat actions in addition to put but also the ability to filter on specific prefixes so you now have the ability to trigger a put notification on one of your prefixes and then maybe a different notification on delete min and optic gets deleted from a different prefix which really gives you a lot more control in terms of wind specific triggers happen and so that your workflows become become enriched and you can do a lot more interesting things we also so a lot of our customers kind of came to us and told us that for compliance reasons a lot of times they have the requirement to have a copy of their data replicated hundreds of miles apart and so far for use cases like those we announced the ability to set up cross regions replication which is very easy to configure policy on your bucket there that then essentially ensures that any new objects that are added to your bucket are then replicated to a different destination bucket in a different region so specific specifically for those customers who have compliance reasons need a second copy of their data in a different region they’re able to easily set up a policy and then it’s three then ensures that all of that data that’s coming into their bucket gets replicated to a different destination bucket in a different region there are other advantages of this as well one of the use cases we’ve heard from customers is additional security and protection against rule users a lot of times our customers since they use is three as a data leak put a lot of very important and sensitive data into h3 and such as logging information in and allow access to that to a large part of their organizations to poke so there’s or their users can can leverage other services in need of us and start analyzing and getting more value from the data but at the same time you want to make sure your data is protected even against the even for the people that you’ve granted access to and so creating a secondary copy replicating in a different region which is completely locked down allows you to get that additional peace of mind and that additional layer of protection we also announced VPC endpoints for Amazon s3

specifically for customers who are using the virtual private cloud capability which allows you to have an isolated network on here WS with your own IP range n IP subnet for those customers they had to set up firewalls and NAT instances and actually managers Matt instances if the application running in the VPC needed to talk with s3 and not only that in order to get access to s3 even if it’s in the same region we aren’t going through the firewall or firewall or the net instances the traffic actually had to go over the internet to talk to s3 so we heard that feedback and we added the capability recalling VPC endpoints for amazon s3 it really ensures that when an application within a VPC needs to talk to the ns3 book in the same region traffic never goes over the internet and customers do not have to worry about setting up internet gateways or instances so that makes it very easy to set up while still keeping a lot of control so you have the ability to set a PPC policy on your VPC endpoint which ensures and limits specifically which VPC widget which s3 buckets this VPC is allowed to talk to but you can also go set a policy on your s3 bucket which lists out all the VPC endpoints that you’re granting permission to that can then talk to this s chief okay so you have a lot of control and you can make sure that the right VPC is talking to the right s3 bucket and of course the traffic never goes over the Internet is the attitude factors where there’s a lot of our customers grow on history in fact s3 now has trillions of objects which is extremely high scale we hear from a lot of our customers that they would like to get more visibility intuitive data it has a customer I would like to understand what data I have on is three because the volume is continuously growing and actually be able to understand my usage office trees well so to that end we added the ability we added a good three new cloud watch metrics of cloud watches are is they’re monitoring service on uws now by default all s3 buckets in the cloud watch metrics for total number of objects within a bucket as well as total number of bytes for standard or new storage class gender in frequent access as well as the reduce redundancy storage so now you can track over time how your storage is changing how the number of objects is training with is going up or down and take meaningful action based on that we also integrated with our audit trail service which is it obvious cloud trip so so now here you can enable a free service with cloud trail where you will see all of your bucket level API is API such as creating a bucket deleting a bucket changing ackles or policies or lifecycle policy on your bucket all of those ap is are now audited and interact as part of your overall cloud trail log so you have a way to go back and know when where who changed my bucket when did that change happen and then what was the end result of that which really going to gives you that an additional auditability for compliance reasons that you can then use and know exactly when your bucket was changing and how that was actually changed time for some specific use cases some of our customers told us that they would like to increase the default bucket limit on their AWS account so by default you can have up to 100 buckets and there are specific use cases where you might have different end users and we want to isolate your specific users by bucket in those cases we work with you now and you allow you to kind of create more than 100 buckets the process is actually very simple you can simply open a support case if you think that you need more than more than 100 buckets and we will work with you understand you use case and increase your bucket limit accordingly now s3 is a very large distributed system and as such for our update and delete operations s3 is eventually consistent and what that means is if I update an object if I delete an object that action and I am I get a success that for that action eventually as three will get to a state steady state where that object will actually be removed or appropriately updated however if I update an object and I should get request immediately I may get the previous version because again it’s a distributed system so you need to build your applications to be able to handle these situations this is also true for putting the new object however based on customer feedback you change of platform and now all endpoints in all regions in s3 or read after write consistent which means if you put a new object in history you can immediately read that object and be ensured that you will be able to see that object and see the right version of that object of course it doesn’t stop there we also announced a new storage class for s3 specifically designed for in frequently accessed data and we’ll talk more about that as you proceed in the in the presentation we also continued in a great honor lifecycle platform which is really the focus of the lifecycle

platforms to help our customers you know focus on building their applications rather than managing their data on s3 especially at s3 scale there’s a lot of repetitive tasks that you might have to do to be able to effectively manage your data but lifecycle platform though we take those those repetitive and those cumbersome tasks away from our customers such as you know I wanted to your data in a year after it gets a year old from s3 to glacier because I want to simply archive it or I want to delete log files after after a year because my policy does not require me to to keep those so those actions that that you would have to perform then on hundreds of billions of objects potential you ate all of your data image3 you can simply set up lifecycle policy and has three will then in an efficient manner perform those repetitive and those those housekeeping actions so that you can focus on building great applications on top of history now we can continue innovate on on the lifecycle platform and there’s two new policies that we’ve announced this year that can again help you manage and key to manage your data effectively on s3 one is what we’re calling expired object weak market specifically for version 4 version buckets you have the ability to set a lifecycle policy to to expire an existing object and move it from a current version to a non current version in that case and then in addition to that you can set an expiration policy to I should delete the non current versions as well however there is a situation where you might end up with an object that has a sweet marker which is really an indicator that there is no current version of this object in all the the previous versions have been deleted based on the policies that you have said however you can end up in a situation where you have delete markers with no current versions no non current versions under it and you can simply list those and delete those yourself or you can leverage this new policy to simply say for if you there is an object that does not have any non current versions under it and then I end up deleting the current version is current version as well as you will take care of cleaning up the delete marker for you similarly we have the ability to upload multiple parts of a large object which they really have you get better performance is unless you talk about how you can leverage multi-part to improve your performance as well later on the presentation but when you do use multi parts there are situations where your application might die or you might not end up completing multi-part uploaded and you have some parts of your of your object uploaded but not others in that case of course you do incur storage costs for the incomplete multiple multi parts that are uploaded you can now set a simple life cycle policy to go ahead and delete incomplete multi-part parts that are dead that if you don’t complete in X amount of days which again is an easy way for for housekeeping and managing large volumes of data where the customer now does not have to go list all incomplete multi parts make a determination and then go ahead and delete some of those and then of course I mentioned earlier a new service the view that we announce which helps customers move data with with high performance in a more efficient way across large tits distances so let’s talk in more detail around the new storage class standard in frequent access so we heard from a lot for customers that data actually evolves over time if I put a new object in s3 the characteristics of that object and the usage patterns that that object sees does evolve over time right based on what I’m actually doing with my application you know 11 canonical example we hear from lot of our customers is if for instance I have a video file and I upload that to s3 and in my application and this is this is new content more often than not that content is going to be very heavily used it might be a document that I upload that I you that a lot of people are editing because it’s a new document that we’re working on but over time this document in the access frequency and how interactive I am with this object changes and reduces all the way through the end where is Rachel you are kind of course we have s3 standard designed for active data where your data is available at the best possible performance available immediately for you to use and we also have a majority which is on the other end of the spectrum which helps you store archive data at extremely low cost but there is a retrieval time to get that data back but what we heard from customers is data does not change from being active data to archive data overnight right usually it’s a process where the access frequency of a data changes and reduces what I’m all the way to the point is not be used at all and is ready to be archived there is an opportunity for for customers to save on storage costs if they weren’t if it doesn’t make sense for them to put their data in history standard because while they do need their data immediately when it’s required it is in frequently accessed and that is really the goal for the

storage class that we have you that we’ve designed specifically around in frequently accessed data and so now you have this additional storage class which is lower costs industry standard that allows you to too which is which is particular data that is that when you need that data it is available at the best possible performance but it isn’t frequently accessed therefore you know I might put data in history standard and then move it to standard in frequent access when the excess frequency goes down and then when that data is ready to be archived I can simply move that integration and we talked about some of these use cases already read but some interesting use cases the view from customers for this new storage class infrequent axis is file sync and share specifically around consumer file storage use cases with a new document when I am working on it when I’m editing it or interacting with it or collaborating on the document it is very frequently accessed I want to make sure that my application can can get the best possible in my users can get the best possible performance and it is obviously always available but over time my files do get less pretty quickly access I still want to make sure that my end users get and are able to load this file as soon as possible with the best possible performance but it is but it is an access every day or it is actress access error at a much lower frequency same thing is true for back up north I specifically around disaster recovery where I might be creating a backup of my of my data and may not get to that for very long large amounts of time but when I do need to do that recovery operation I need to make time you to make sure that that data is available immediately because this ties into my business continuity and then also for longer tane dinner I data such as logging information where I I keep a lot of logging information because when I do run into an issue this is very useful for me to go debug or go troubleshoot or even just analyze my usage patterns in my application to make it better but it’s these large volumes of data that I keep over very long periods of time I don’t necessarily analyze all logs every day and so that is another use case where it makes a lot of sense to use the standard and frequent access storage pod class and save and save on storage costs so let’s dive a bit deeper into the infrequent access storage class we made we wanted to make sure that customers never have to worry about ever losing their data on it so you regardless of whether you’re saving that data on standard on the new standard in frequent access storage class or glacier so with standard in pretty good access you get the same good durability you get with s3 standard which is 11 lines of durability Saturday entry code access is designed for three nines of availability which really means that ninety-nine point nine percent of the time when you do a get operation on your data it’s going to be available for the point one percent when it isn’t you can simply do another get request immediately after and we actually be able to get this get the data and it’s going to be available for you to use one key part though is we wanted to make sure that when customers access and request and interact with a data that’s in standard infrequent axis we get the best possible performance right essentially the same performance you get for s3 standard and everything is a big value added value proposition for for infrequent access because you want to make sure that your users do not pay in terms of lower performance when you when you’re trying to save on storage costs in frequent access you essentially get the same throughput the same performance as you do with with standard the other big part of this is regardless of what storage class you choose in s3 you essentially have the ability to put your data in the same bucket right so your application never really has to change when you’re transitioning data from standard is done in frequent access and we leverage your life cycle platform to be able to do this is the same race interface the identity of the object does not change needed at the location so you don’t need to move your data from one bucket to another that’s focus on infrequent axis which we have which of course will be work for you and you would have to change your application to be able to handle that the identity of the object is exactly the same the bucket is exactly the same you simply just tear the data from standard during frequent access without the application never knowing and so you need no changes to your application you can use the same same simple rest interface the other thing I wanted to call out is you can in a lot of use cases that it does make sense for you to put data that happens to be in frequently accessed from day one directly into into this new low-cost storage class and you can do that using the same put-put object REST API simply by specifying the storage class security is always top of mind for for s3 NE WS and as such all the different security capabilities that we’ve added 2 2’s few over the year such as the ability to encrypt using suicide encryption bringing your own keys or using KMS managed encryption keys you can use all of these just like you do with standard with its new storage class and of course all the other capabilities that that

make data management on s3 easy such as life cycle the ability to version your objects and amid notifications all work with infrequent access just like they do with standard so with that let’s switch gears a little bit of talk about the other new announcement that we’ve recently done a new service that allows you to optimize their transfer to s3 over the Internet especially when your mock now especially when you’re moving data across the internet and across large distances it’s very hard for you to optimize performance over the over the Internet specifically because you you know a lot of times you don’t have control right you don’t know where did where your traffic is going to be routed and your throughput is only as good as the weakest link and so there are solutions out there that do help you optimize this is throughput and specifically those propriety solutions but they are very complicated to set up and manage a lot of times they do require you to install clients install software on your client side or even at the server end so that the end to end the optimization happens but then you need to install the software configure it and maintain the software both at both ends and of course these are expensive solutions raise a lot of times there’s upfront fees or fees for per gigabyte transfer all the way up to to the order of dollars and so kind of hearing the feedback from customers we we decided to work on and use this new storage this new offering which is Amazon s3 transfer acceleration which is really a service focused on optimizing your performance and throughput from for for those use cases where you’re uploading data and larger files across large distances and by large distance distances we really mean uploading the cross continents now when you do upload get up from your client to an s3 bucket in a different continent or over over a large distance you don’t really know the exact route that your data is going to the end that your traffic is going to take and like I said before it’s only as good as the as the weakest link we transfer acceleration however we leverage the global infrastructure we have for AWS and allow you to make a short hop to the nearest interviews education from where we used a lot of optimizations on routing as well as on the networking layer and actually leveraging the the infrastructure in the backbone that amazon has to then send that data directly to your s3 bucket in a different region and different continent had a significantly higher performance in fact in a lot of our internal testing and feedback that we provided from beta customers under certain conditions you can see up to four hundred percent improvement in performance over large distances we also wanted to make sure that is extremely easy to be able to use as free transfer acceleration so you don’t need to need to make any changes to your code or your application you simply need to point it to a different end point and then and then if there’s an is to make sure that we automatically identify the edge location closest to you and actually leverage that when we when we upload the data to to your s3 bucket first and then because of that you don’t need to worry about any firewall exceptions you don’t need to worry about making any changes or adding any software to your clients so here’s a quick look at the flow on how how this works end-to-end so if your client is if you decide that you want to use a three transfer acceleration on your client you simply need to just change the endpoint wave so instead of saying my bucket named s3.amazonaws.com musically say bucket name dot s 3 / accelerate dot amazon AWS comment that’s pretty much the only sure you have to make in addition to a checkbox on your bucket to say that this bucket I want to enable chatter acceleration and the way this works is that so in order to do that put API call we resolve using rad52 the IP address for an auto detect and resolve the IP address for the late for the closest it obvious each location to you and then leverage that education and the Amazon backbone backbone to then send data or HTTPS to all the way to to the end region you want to put your data in and then we have an easy to proxy fleet running under the covers which will then make sure that we put your data to your SUV okay now all of this is completely transparent to you as the user right the only thing you need to do is check a box on your on your s3 bucket to enable transfer acceleration and simply point to a different head money the rest of it we make sure the v potter detect and get to the closest it was edge location to get the best possible performance and other optimizations that we’ve done to make sure that we can

leverage the database backbone in an optimal ways such as we leverage the maximize the the connections that are therein avoid reconnects so that time to set up is is saved and so forth but really this has been designed for sending larger objects to our long distances way and that is a key point and you know in some here’s an example of some of the internal testing that we’ve done where if you put data over the public internet the further away you are from your end point in this case you are putting data into in Singapore the further you are from your end destination the higher delays you see by the higher impact you see on your performance with a suit transfer acceleration however you’re able to leverage all the work that we’re doing and handle the Amazon backbone and get significant performance improvements and of course they get better as you move as you move further away from your destination bucket so as I mentioned before is extremely easy which really was one of the goals for the service and so really those two steps right one is on your bucket you need to go in able s to transfer acceleration which is a simple checkbox and then for your application when you do the put request you can simply specify a separate endpoint and that’s it we also have a tool that’s available to for you to determine how much performance improvement you will get so we urge you to go look at that and leverage that to determine if standard history class or acceleration is a is a good fit for it for your use case alright so let’s focus a bit on some of the other capabilities we’ve been adding to the platform and let’s start with with lifecycle policy so we talked a little bit about this already with the goal for life cycle policy is really being making data management easier for customers on its three and helping customers derive more value from there from the data on honest three and there are two core actions that you can take with lifecycle policies one is what’s called transition which is really changing storage class of your of your data from from one surface to another based on a specific rule which is which is really the object age so for instance I can set a lifecycle rule that says I need to tear data from standard to glacier when the data becomes a year old and that we will evaluate on every single one of your objects regardless of the number of objects in your bucket and s3 will automatically take your action instead of you as the customer having to list your data and then try to determine which one of those you want to move to move to the different storage class the other actually you can take is exploration and so there you can delete objects after a specific specified amount of time again same rule but the action you take is different you can expire or delete objects based on with what rules you said you have the ability to combine multiple actions so you in a given pocket you can decide to you can actually set these at the prefix level so you can decide to tier objects in a year in one prefix where is wait a few months like 90 days to tear objects to a different storage class prefix so you have a lot of control in terms of how you can figure bicycle policies to manage your data in your and your s3 bucket for the new storage class that we’ve that we’ve announced as I mentioned earlier it fully supports life cycle policy so you do have the ability to transition data from standard too infrequent access or data that since that’s in standard infrequent axis and archiving it directly to to creation based on the object age you can also set exploration policies on data that is in standard frequent access storage class and you can also use it with versioning right what do you need the capability that we where we allow you to ensure that you if you do replace or delete a version of your your object you don’t accidentally remove that and we actually keep that object as a non current version so it’s identified you can get to the latest version at any point in time but since we’re not removing or deleting view or overriding the objects you have previous versions of your object available for you so you can always recover those and that again works just like it does with s3 standard with standard frequent access and the other thing of course is where it makes sense for your use case you do have the ability to put data to retain to stand on your frequent access and start saving on storage costs from day one so here’s an example of a lifecycle policy that you can set and in this example in the highlighted area i’m specifying a rule in action for me to transition data that is in standard once it becomes 30 days all i want to tear that into said infrequent acts because for my specific use case i know that my data becomes infrequently access about 30 days and of course you can pick a number that’s appropriate for your

specific use case I can then set a different action to transition data that is a year old and applied that integration so both of these rules then work together to provide you the two to provide you the ability to first here all data that’s in standard to standard in frequent access once it becomes a month old and then a year a year after that you can there then transition the data into to glacier and both of these rules work together so let’s talk a bit more about versioning for s3 objects as I mentioned earlier this is the ability the versioning gives you the ability to ensure that you never accidentally remover overwrite your data and you always have the ability to recover from an accidental operation and then really it achieves that by ensuring that we never really remove a version of the object so if I had virgin enables on my bucket and I have an object there which is the current version of my object if I issue a delete against set object history will not actually remove that object will rather just make it a non current version and then place a treat marker on top which identifies that there is no current version for this object if I were to overwrite this object then the latest version would replace the current version but we will not remove the the previous version it’s simply simply becomes a non current version and both of these in Turnham ensures that you can always you always have the ability to recover from an accidental operation and this of course is one of the best practices we recommend we recommend our customers in a lot of times it’s extremely valuable for those for those accidental overrides it means it’s very easy for you to retrieve n roll retrieve your deleted objects and roll back and there’s basically three states in your version when you version in a bucket you by default the buckets are not version so you do have enabled that and that change in state to versioning enable after that any operation you do on your buckets it is it agreed it will not actually remove the non current version simply just mark it as not current you do have the ability to suspend versioning if you want in which case we will retain all non current versions but any operations you do afterwards once the bucket is in this state we will go ahead and actually remove those versions remove those objects a lot of our customers actually use versioning along with life cycle to actually get recycle bin functionality right which which really kind of helps you easily recover for those accidentally deletes so you so with life cycle pause using versioning together you can make sure that you have a sliding window of say 30 or 60 days where your data is completely recoverable and you can always roll back at any point in time and it’s it depends on your business lead and how big you want to make that window but you can use versioning along with lifecycle policies to make sure you don’t have infinite number of versions lying around but at the same time you have a buffer period for your recycle bin now since we talked a bit about about versioning there is a specific use case where when I have originally enabled for my bucket and I delete the latest version of my of my object s 3 indicates that there is no current version of this object by placing a delete marker and then making the the current version and marking it as marking it as non current now in addition to that you can actually say you can either directly delete a non current version by using the API or you can use a lifecycle policy to say a week or 30 days or whatever however many days you think that makes sense after an object becomes non current go ahead and clean that up for me go ahead and remove that for me so in this situation you end up with a delete marker which has no non current versions or no previous versions under it there is something that you can buy specifically a specific indicator you can flag you can list all get to be more present if you can clean those up yourself but again at s3 scale that is a cumbersome task that customers would have to probably write code for to be able to do with the expired object delete marker policy however you can simply tell s3 to delete any mark delete any delete markers that do not have any non current versions under it again taking that taking the owners away from customers so that they can focus on writing their applications as opposed to doing housekeeping on on the d-line s3 here is an example of a lifecycle policy where we’re indicating there we were indicating the vini that we want to remove non current versions you can simply specify a flag that says non current version exploration and you can in this case I’m saying after 30 days 30 days after an object becomes non current we want to remove that object in this case is 3 will make sure that when we do remove the when we expire an object placed delete mercury and remove the non current version we actually automatically clean up the delete marker

but if I was not leveraging lifecycle policies to delete my my latest version and my non current versions in this case for instance I’m not I don’t have a life cycle rules expire current objects I do that using my using the API directly as sweet as the last leg of platform does not know to remove the delete markers so i can simply add a bag that says i wanted to lead expired object delete markers and set that to true with this if there is a delete mark group with no non current versions hundred s3 will automatically go ahead and and clean that up for you speaking of the ability to recover from accident of the leads one of the best practices I want to call out is vs3 does have the ability to enable multi-factor authentication specifically put to the API which again gives you that additional layer of protection to protect against accidental the leaks when you enable multi-factor authentication you will be required to not only enter your security credentials but also a code from an approval of dedicated device for the specific delete API which if you have this enabled will make sure that that you avoid accidental deletes and permanent of needs for for urgent objects alright so let’s switch gears a bit again and focus a bit on some of the best practices that we’ve seen specifically around performance improvement so one of the things I wanted to talk about was the ability to upload multiple parts of an object so this is something that is three allow you to do where if you have large objects you can paralyze the upload of those objects using the put API but essentially breaking the object down into multiple parts this allows you to remove bottlenecks on on your phone from your from your community network and really can leverage the power of multi-threading and parallelism so that you are uploading multiple parts together another added advantage of this of course is that you can increase the resiliency of your network errors specifically for those networks there are error-prone so first is if you’re uploading using a mobile network it’s possible that when you’re trying to upload a large object you see issues with your with your connectivity and you would have to restart entire upload again but if you’re uploading multiple parts then you just need to restart the flippers for a smaller part right which makes it much more efficient and as resiliency to your network so so with multi-part upload you can you can add parallelism to your upload that really makes it faster and more flexible for you to do uploads and allows you to upload a set of parts and then of course once that upload is complete s3 will automatically present all of those arts as a single object for you um so it’s irrelevant to this so there are cases when if you’re leveraging multi-part upload it’s possible that your application might die or run into an issue where the application diet and does not complete all parts successfully upload all parts of your object in those situations because you do have some parts of your object uploaded and you are paying for storage for those and in these cases you have the ability to specify a flag in your list API to explicitly call out and list the multiple parts that are not complete and then you can go ahead and and remove those who think that makes sense but again that is a repetitive task that our customers have to do and based on that feedback the added listing policy to our life cycle platform which is the ability to automatically expire incomplete multi-part multi-part uploads based on the each of the upload so you can simply specify a policy that says if my multi-part upload has not completed in say seven days or ten days like a platform can expired the the upload and delete all the incomplete multi-part to take you back to a non good state so as an example it’s very simple to set that lifecycle policy you can simply specify a flag for an action for a board incomplete multi-part upload and specify the age so seven days after initiation of my multi-part upload if the multi-part upload has not completed go ahead and about that and delete all the incomplete multi parts for me another optimization you can do is perform range based guess right so in addition to doing multi-part upload where you upload multiple parts of your object you actually have the ability to get specific ranges in your object which is which is especially useful when getting large objects because you can use the power you can use a powerful parallelism here as well and when download parts of your object at the same time which again compensates for unreliable networks because when you do have an issue with a

specific part specifically in that you’re getting you only need to restart the gate for that specific range in so you can add resiliency and benefit from the prepared ilysm the same thing is true for list you can paralyze your list API and you can do that by specifying the social for list API you can get up to a thousand objects at a time but you do get an indicator that tells you where to a marker where to start from for the next set of thousand eight pockets so if for instance you wanted to list a bucket with five different prefixes you could at the very least kick off five different instances of list focus on specific prefixes or if you should understand how your your bucket is set up and your prefix is set up you can actually paralyzed this request for a gift for a specific prefix as well another best practice in a very common usage pattern we see from our customers to optimize this is actually using secondary indexes you can use services like dynamodb to actually set up a secondary index when you put an object ministry you keep an indicator for that in in your in your secondary index and then of course you have the ability to add additional metadata as well not only does it make lists much faster but you can also get additional benefits such as sorting based sorting sorting and searching based on metadata or other properties of the object there are a few best practices we talked about as far as SSL and using a Sicilian optimally is concerned first and foremost if you use the SDK we make sure that your SDK always has all these best practices and more it is always up to date in leveraging all of these best practices to give you the best possible performance so if you where you can we encourage you to leverage the SDK to to talk to history but there are specific actions you can take such as if you are using ec2 as your client to talk to s3 you can leverage hardware acceleration on specific instances and that gives you better performance when it comes to encoding for for ssl you should also try and avoid timeouts where you can you can do connection pooling so that you don’t need to do the word to set up the connection every time you can even send keeper lives to to avoid handshakes which are expensive and can add up especially on mobile networks or networks with with limited bandwidth another best practice is around how you name your bucket and that really is that we do see usage patterns in our custom where customers try and leverage prefixes like folders if you try and use prefixes to organize your information well it makes it easier for you to discover your information especially at higher TPS and specifically Hunger GPS and higher you start seeing performance since occasions because a lot of your data is is under the covers on similar or say or the same partition now s3 automatically tries to distribute your data pinion is minion as as many partitions as possible and it automatically if you start seeing high traffic on a specific partition we automatically move it around as well but that does take time so so in the meantime you will see start seeing implications on your on your throughput in and start seeing performance issues and so here i have an example of a prefix which is very common especially in when people are putting logging information on history that you organize it by by year month and day and what that does is all of these objects end up being on similar or the same partition so all of the objects that have 2013 on here as the first part of the prefix will end up going to the same partition so on this on this prefix if I start getting extremely high TPS I might start seeing performance implications for some amount of time so the recommendation there is to add as much randomness to your key naming as possible so here is another example where I’m not using a specific year range but rather I’ve randomized all of the queues for my objects and so this ensures that inherently you end up on different partitions in s3 and you never end up having a partition especially as you as you scale TPS and your traffic very very fast you will never see that that implication because you will not have too many of those going to the same set of partitions there are a couple of techniques that customers use to be able to randomize the keys one is to hash the entire name my object so if you do that then pour the object name does become does become randomized and therefore I don’t have the same object name showing up and that data ending ending up ending up on the on the same partition another approach is to prepend Yorkie named with a short hash and that again of course randomizes

the initial part of my key name automatically sending me to a different set of partitions yet another approach is trying to reverse Yorkie named specifically in the cases when you’re using dates and so if you’re doing that Union just reversing that me may help you introduce randomness but really it’s about randomizing your keys and with that let’s let’s open up for questions and let’s see let’s see what we have there thanks O’Mara for a great presentation Andrew Carl though and Susan for engaging with our audience throughout presentation Jin so on your screen you will see three live poll question please provide us with your feedback by clicking the radio button next to your reply for each pole we have about 10 minutes laps so please feel free to continue submitting questions we will address some of these as they are submitted but apologize in advance that we might not be able to get to all of them so I will hand it off back to a me Andrew to answer some of the Q&A we proceed thanks Jenny we’re glad to go to be here everybody so we’ve got a number coming and while you’ve got us here experts on the phone we’re happy to stick around and take care of them so we’ll do the best we can ayo Mara I’ve seen a couple of conversations about life cycle policies one of them particularly can you apply a life cycle policy to trance to move stuff into SI a at a bucket level and not just a file level mmm that make sense yeah and the answer so with nicely the policies you have a lot of control in terms of where you apply what actions as I mentioned earlier you can specify these at a prefix level in fact you can specify different actions to different prefixes but you can always decide not to specify a prefix right and so the bucket the policy can then apply to every prefix within your bucket that essentially makes it a packet level policy as far as the action is concerned you do have the ability to tier data from standard to standard infrequent axis and that simply is one of the one of the actions that you place in under transition under under transition okay cool and then halo I think there’s some questions here about s3 transfer acceleration particularly around performance so if you could just kind of talk us through why in some cases it may be you know one to fifteen percent better in other cases it may show dramatically better and also how can I test it when I’m thinking about whether this will affect my own particular situation okay so s3 transfer accelerate gives benefit to over regular s3 uploads if the distance between where you are uploading from to your destination s3 bucket is far away I think an example given is if you were on the East Coast let’s say Marilyn and you’re uploading to the eye on the Virginia region you’re not likely to see a large benefit if you go further way for example you’re trying to upload from Los Angeles to the I to the Virginia region you’ll see a a moderate benefit and then if you’re even going further away for example if you’re going from Japan and you’re trying to upload to your Virginia bucket you’re going to see this you should see it extreme benefit so your benefit is going to be based off of how far away you are from the bucket that’s really hoping that ass and the other interns let’s be comparison there is a speed comparison tool that is on the s3 page that you can look at when you click on your bucket that should give you a good indication the benefits you will see so you can run comparisons uploading to different buckets excellent so I can just run a test from wherever my client is and see what my results might be that is correct okay omair can you talk about best practices on bucket naming or or using folders and their names I think you touched on it but it seems to be a really important issue absolutely so the key remember there is you want to make your bucket your your prefix names as random as soon as possible right so a little bit about buckets and prefixes so once you name your bucket s3 essentially is a flat structure right so when I put an object in industry I can add more more prefixes to it which kind of show up as folders in the UI and I can say you know for bucket be one I want to put all of my JPEGs under a folder or prefix called images so be my bucket name / images /

all my you know all of my objects the issue with that then is if you think about it the prefix starts with starts right after the bucket name so the bucket name followed by all all of my billion images that I put in this three are going to start with IM h s right images and so the issue with that then is as you start pumping high TPS to these set of images because because the prefix initially is this is similar is the same as three tries to put them in the same setup in the same set of partitions very possible now the impact that of course is if I start hitting 300 TPS and start getting you know lots of these images for my for my specific workload it’s going to be the same set of hardware the same set of partitions under the covers that are going to start pulling this information and sending it back to you which witch then does mean because of the same partition there are physical limits and you start seeing performance implications there now on the flip side if I had said you know what I’m going to hash or add a four-character hash in front of the folder name images and so my bucket name that looks like you know my bucket name /a hash for witches which is obviously random the underscore images and so forth in that case I the data under the covers is going to be 1 on Brandon partitions because it’s going to be the first few characters that we look at before when you’re putting the data in specific projects so when you do a putt request and you look at okay what is this we’re just doesn’t end up we first look at the first random four digits and it goes it ends up in a it organically ends up in a different partition so now when i drive the 300 or 400 TPS to my bucket excessively to be talking to a different set of hardware which which inherently make sure that you do not see any any performance implications there so so really it’s about when you sit down and think about naming your buckets and the prefix is within your bucket specifically it’s always a good idea especially for those high TPS and high volume volume prefixes and buckets to randomize the initial part of your of your key name where possible how do you so I understand the randomization value but how do it then you handle customer challenges around navigation when the when the logical structure gets sort of at a random level your files naming yeah okay that’s where one of the common patterns we see for our customers is using a secondary index if I were using a dynamo DB index where I said you know what I want I wanted everything to be under the folder images for slash the name of the name of the file every time I put an object in this three I can simply add a add a row to my to my my table in dynamo DB to say for an object that that I want to put under bucket the images you know image 1 the actual name in is 3 is bucket name random hash and so forth right so that there for you keep that mapping around and so every time you want to do a look of you can simply go to to DynamoDB to know ok for for my image 1 which i know is under images what is the actual key name for it alright so once you maintain that mapping between key name and the logical structure you want to have then you can simply look at the fed up back and go back and forth earlier thank you all right and I think the time is just about up and unfortunately we’re not going to be able to get to everyone so Jenny do we have time for one more do we need to pull the plug right away we are asked out of time so I’d like to say thank you everyone thank you 20 mirror drew Carl’s well as Susan for answering the audience’s questions you have any comments on our previous or upcoming webinars or suggestions on topics please feel free to email us at AWS webcast at amazon.com your feedback will help us improve our webinar programming thank you to the audience for taking your time to join us today enjoy the rest of your day

Multilayer Perceptron In 3 Hours | Back Propagation In Neural Networks | Great Learning

the perceptron is the basic unit powering water study known as deep learning so multi-layer perceptron is a perceptron that teams up with additional perceptrons stag din several layers to solve complex problems understanding the importance of this we have come up with this tutorial on ealier perceptrons now before we go ahead with the session I like to inform you guys that we have launched a completely free platform called as Greek Learning Academy they have access to free courses such as a iCloud and digital marketing you can check out the details in the description below now let’s have a glance at the agenda we’ll start off with an introduction to perceptron and neural networks then we learn about activation functions after that we’ll understand the concept of forward propagation and back propagation and finally we’ll have a demo on neural networks we’ll start with some problems first of all why you need neural networks the problems with we will say machine learning so let’s try to list out let us say some three to four or five problems that you people feel machine learning will not be able to any medium so it’s always good to know the problem and then go on to why are we doing that so any idea what machine learning is lacking or when there’s a large data has one thing okay so you are saying data size is an issue yeah all right fine we’ll see how to put this as a problem anybody else and then we have non numeric data okay so I will say non-numeric data perfect where we don’t have enough features all right featuring issues anything else Corrections do people see that do we have I think we have only two algorithms in available in machine learning who are actually capable enough to correct themselves correct what are those algorithms we have SVM and boosting do people remember these two here they have a capacity to learn yeah so this is one issue our ml is you know as one of them which we have to train and then we can use it it does not have Auto hearing or auto correction system kind right and fifth one I will see is somebody already featured non-numeric data so I will say what if you want to process a sequential data so like I was looking at one of the codes sent by Ashish so Ashish of you’ve used LST M’s there so are you sequel to Dallas team of them I use him like I work on that a lot of a what a lot of videos and so you have an idea Felicity right yeah yeah so this is one of the purpose so guys what is LS what are these things if you see our machine and stuff I cannot say this is a sequential data sequence in the sense we I showed you people Arriva for example timeseriesforecasting that was a sequential stuff sequence means what happened some 10 years back there is chances that it the same month or same time the same thing which might happen the seasonality part of it what if I want to take care of something like this on to machine learning sometimes it is not possible to do it and there are hundreds of very more reasons so I will say for our course the best reason could be non numeric data what do you mean by non numeric data say pictures say text say voice we cannot process it was up to some point we can process using machine learning after that we have a limitation here all right so now what we are going to do is let’s try to build some algorithm into machine learning see also one more thing is even though in our course we say machine learning is over but actually machine learning as the whole box under which you have two parts or three parts here which we call it supervised learning unsupervised learning reinforcement learning after that we have something called artificial intelligence and then I will say some mixed bag of yeah so still it’s a part of a melody but from now what we are going to do is say I have a box I will not say what it is we have some empty box here and I am giving some input to that box yeah so let us say my input is

defined like this now when I give an input to this I am expecting a certain output so before I get an output I have some expectations let us say my expectations are exp I am getting an output here I’ll say expected out whether we get confused now what we are going to do is we are going to analyze this output is it as per our expectation or not so how do we analyze let us say the analyzed using a function I use a function basically to operate on our output versus expected output all right now if this function is of certain value let us say the two values can be good another one value could be bad let us say if the function isn’t bad value what are we going to do we are going to go back and make some corrections to this box now what made this output bad can I say this particular box made it bad because of this box we are getting some unwanted junk over here agreed everybody yes so what is our duty back go back and try to fix it you keep doing this in we come to the phase which we said good so we will come to this what are these things but till we find that this and this is matching he believes this is what is an example of a neural network a very simplistic approach all right any fusion to this no no okay now there are two two paths that you people observe this one path was called forward path forward path walls where I gave my inputs and this box amplified it or D amplified if you don’t know if the box made some changes on it and give us an output this is called forward propagation now since we found out that this function is of not a good value we went back and we corrected our box this is called back propagation as simple as this all right so these are the most important parts within your neural network to understand that we will give some input will get some output and will keep training it till we are satisfied that yes we are getting expected output out of it okay one more example I usually use this very silly examples let us say this is the kid for example okay and you are teaching some tables to the kid say you have got five team so I’m given input that kid is learning yeah so you made you did something such a way the kid knows the people now you asked her the gate five five Zach for example the kid says five five thirty – so what do we do immediately we go and cut we keep doing it till the kid comes and says Phi Phi is at 25 this is also another example of neural network forward prop and a backward prop to correct it and keep doing it till you are happy with that okay now let’s talk about that box that we were talking about now what exactly is this box so this is give me a moment I will open up my PPT one question here yesterday when we go that B go for the backward direction in that case it is assumed that we know the input as well as the output and ZF for putting into the box and very find out whether the box is giving me the output but in the case of prediction we really do not know basically the future prediction we really do not know the output so how does this really work okay so for any machine learning whatever box I showed you know whatever we did supervised unsupervised AI reinforce whatever we saw the very first phase of any machine learned technique is machine learning PC so for that we need a set of independent data and a set of target data no matter what you are doing if you are doing text processing if you are doing voice processing image processing no matter what the input matrix remains C alright so what do you have you have got some input for training that is we call it an independent data then you have got some tags if you remember I said Y actual right this is what will be our reference point and whatever we get output we will have we will call it as Y pred see we are going to compare always vibrate with y actuals all the ways no matter what algorithm you’re doing and then if you are asking your future predictions one our machine is learned once the difference between both of them is

reduced to our satisfied level and the machine is learned you have to disconnect this you have to disconnect our input and you give only your futuristic output which will be only your independent data so that you will get Y all right because that’s more of a learning phase what you exactly yep so this is a common stuff for all the machine learning no matter what it is the first thing we have to make it learn now there are some algorithms like LS TM and all that I was just talking about yes you know those things we can predict some something without learning much of it but will come down and so when we come to NLP we will look at let me not confuse you there is right ok for now this is what is the architecture we are going to look like now there are some questions let me see ok all right ok so now let’s go to this PB how to interpret your network so let’s keep let’s take the inspiration this is this particular topic is being picked up from biology ok and neither of us are expert on to that so we will make our own version of new let us say your brain neurons the neurons that are there and please remember there is no neuron in your brain which is not connected to other one some other other way it can reach from here to let us say to some other you know yeah so that is the first part of it which is called fully connected neural networks this is the basic of all the neural networks and that means each neuron is connected to all other neurons in the network alright now let us say you divide this into four parts this is part number 1 2 3 & 4 now when this brain was newly formed that means when this was a baby at that point of time all these 4 web of almost similar these zones were not differentiate now what are these zones you may ask me let us say this is my happy zone for example this is my sad zone ok we will not compare it with real biology because if we do it we are going to fail on to this so I’m just showing you certain neurons belong to happy certains on sardines want to say relax part of it or sleepy part of it and certain zone to the anxious part of it for example okay let us say when the baby is newly born they have not seen much out of the world so what they do is they mix up all the emotions sometimes they keep crying sometimes they are happy they don’t get it now when when lot of events happen day to day life and the person who grows up to some level there are certain learnings that the person does and the person tries to store those learnings in each of these knots depending on the zones so let us say after let us say at the age of 30 if a kid passes with a very good score definitely the neurons over here will be active alright so please remember a neuron can be activated and a neuron can be deactivated that means in the happy phase all of these will be deactivated all right they don’t continuously keep running the electrical signal is sent from one to another one only in certain triggers situations all right so this is what is the theory behind why we use neural network so from here now we will try to put this whatever I said onto the machine part of it so let’s see so what basically you will learn at the end of this particular session as intro to neural networks working of a neural network forward propagation backward propagation and types of neural networks so we will see do this on a very detail because now whatever for five months you have from your course we will try to explore almost every type of neural network available in the market right now next one what exactly so I explain you what is a neural network why do we use neural networks so please remember it’s a very simple neural network can adapt to changing input so the network generates the best possible result without needing to redesign the output criteria what do I mean now let us say for example we take linear regression yeah in linear regression if we have baselined of our equation people remember this equation right I think we started our interaction from this session itself yeah now is it possible for me to change the inputs if I change the inputs will this work we are not sure if the inputs are very close to what we trained it on yes it worked if they are not it will not work this model will crash what if I tell you that there is a possibility of making a machine learning model no matter what you are going to give me I will relearn myself always and I will deliver it output what if I have what a

self learning kind of stuff yeah so I will say this is more robust to changes this is more adaptable and this is forth quicker in the terms of learning yeah so this is where you are going to need your neural network so the box that I just draw on the whiteboard that was nothing but your usual guys do stop me if you find it difficult to understand yeah now next one what is the application of it so basically I hail from deep learning background this is what I do on my daily basis in my job now you can use deep learning for anything whatever you have done so far classification regression you can do the same thing on your network so the first thing I would say is decision making that is your prediction and decision second thing is if you want to go a little more complex you can talk about image processing so what do you what do I mean by image processor so let me quickly show you then then do you think you know the DL is much better than ml yes definitely okay then why why the people use ml then yeah I’ll tell you why why do we do that just give me a mint okay data science contains deep learning yeah so now let me take you guys through a very simple application of neural network it is not simple to code but yes down the line we are making it more and more simpler in a minute I have too much actually over here the main image recognition and image let us say I have a screenshot from a particular movie alright so this is a screen shot for me now I want to automate certain things over here so I will give you an example if you people are on Netflix or Amazon if you click on the star cast the particular scene which is going on they will give you the they will match it up who is there in the particular scene and they will give you the description of the star cast every guys observed that if you stop the screen immediately and check for the stars especially on that’s one prime you will get the names of only those people live who are there in the scene have you seen that no nope just is in prime you get to some correct how do they do it very simple what do we do this when you click onto that star cast button the Amazon Prime uses its own image recognition system and tries to create a boundary can it mean use a different color tries to locate a face in the picture you would have also seen these kind of stuff nowadays flying around on YouTube if you are talking about a I know they will see people who are walking and they would have made one box around the people who is walking and they would have said this is a person who’s walking it’s become a trend now to post all these things on YouTube and LinkedIn just to show that I work on a I part of it yeah but it’s not rocket science we are going to learn down the line now this is what I want to do so how to do this very simple first of all you need a big corpus corpus means a back-end data saying that we don’t we don’t know whether we’ll get image of this lady or not we don’t know whether is there in corpus but yes we will get lot of images of men in women and if it is a men the tag attached to it so if I say the independent variable is men the target variable attached to that will be men that is the image is men the target variable attack – it is men so this is how I tell my neural network that I am giving you an image as an input you have to adjust your whatever is inside such a way that you have to give me so whenever I give you this image you are supposed to tell me it’s a men you know so this is how we do it so let me now show you how do we do it so this is an untagged image let me try to copy the name of image and now we will try to push some kind of brackets on – yeah let’s see so all I need to do is this is my version of image processing okay so just hold on for a minute a kernel is dead just give me a minute guys why am i showing you this is so that you get to know why are we learning deep learning otherwise if I

would have start with neurons and all it will be the same as our video yes and here I will paste our image yeah so saying yes it is showing me there are two faces that he has detected did you see that and later on if I you go and see here this is my output image it is very simple it is showing me and we have detected two faces here the next version of this could be find out who the celebrity is so what we have to do we have to give an input corpus of all the cell and we have to retrain a network saying that if you get this particular image whoever this actor is all right so this is how Amazon and IMDB and all these people do it if you want to take one more example let us say we take a complex example look at this image it’s pretty huge right so what I will do is Liu push it across through a neural network so just let me copy the name of the image will input it so this is a pre trained neural network I am not real if you observe there is no neural network running here why because my neural network is currently lying into this particular format you don’t have to keep running your neural network every time you can store it pickle it and whenever you want it you can put it on in this case I have an XML file which does my job now if you take some time because there are a lot of faces there and it will show me how many faces are detected if you observe these saying we have 29 face is detected let’s go and see the output this is how the output looks like yeah and you will observe some of the faces are not detected in your in you spot those faces 1 2 3 4 what is common between all these faces can I say they are tilted yes yeah so there is no box around it that means the corpus that I have used to train this particular neural network needs me to have a straight face or full face for this particular person the full face is missing all right so this is how you this is the end application of a new little I will say clear we will go deep into it down the line in computer vision session anyway so this is how we do image processing using neural networks ok speech processing so I can talk something in from the corpus the network can pick up and it can tell me what I spoke about ok not sure you have seen this earlier but yeah so this is my speech-to-text converter so just observe I will talk something and after a once I talk it will populate here and it will tell me what I spoke about alright so hello everyone hello everyone vir hi everyone yeah so you observe it said hi everyone if I say some of your names say Raju are sheesh mohit Raju Ashish Mohit yeah did you see that so how does it take it it what are we doing is basically we are running some kind of Google audio in the backend so it is not only understanding what I’m saying but he’s also trying to translate it and move so basically this particular board what it does is it’s a payment bot so if I say I paid 300 or finally rupees it identifies what I’m saying and then tries to pull out the money oh yeah so this is one more application where you can use your neural network for voice mod takes part of it so text is not in our certification context but I will show you guys down the line how you can use this for text also yeah again gaming

I will say it’s very big industry when it’s out of my scope also it’s very complicated and on the futuristic next five to ten years you can expect something like this to grow up so I’ll give you one example there is a startup who is working on an eyewear for visually impaired people so that particular eyewear will recognize what is going around the world so well let us say the person is sitting in a park see if he wears those glasses the glasses will analyze is what is going around the park and those analysis will be sent via voice to blind post saying that there is a dog who’s white colored he’s running there are three kids were playing a football looks like it is going to rain if the person is walking it will question the person I am able to see there is a car coming please hold on you know so whatever challenges he has because of no vision they can overcome this so probably next four to five years you can expect this to be very common so computer vision and AI is working towards this so this is where you are going to use this I’m not sure how many of you are going to actually work on this industry but yes we’re very good no place to work in to our next fight Romania now coming on to the real part of it now what exactly is a neural network we saw the applications very good now this is how a simple neural network looks like this is an example of fully connected neural network F CNN because if you observe each neuron is connected to all the neurons in the next layer it each neuron is connected to all the neurons in the next layer so you’re not leaving anything we are fully connecting it now what actually happens let me show you so what we do is we give an input now each neuron or sorry each line that you see over here has got a weight some random number you can say which gets multiplied to my input and goes to the next neuron the whole game over here so if I ask you now let us say we talk about some of this machine learnt algorithm let us say we talk about SVM can somebody tell what exactly we learn in SVM we give some input right we train the model till it learns and then we say okay of the model is trained so what is those things that we are learning in his field if somebody can highlight that in the pattern in the input perfect but turn theories whatever which can define the whole bottom data so let me show you a small clipping and after that we will go into an excel file where we will simulate on your okay so just observe this clipping for now and then we will see what happens okay so you will observe there is one number seven which if you observe this number seven we have divided this seven into some kind of pixels or a numbers those pixels and numbers are going in the first layer as an input okay and if you observe some of these are on some of these are off so if I rewind it yeah so this is my some layers in the box and this is my output so if I give a number two you will observe this particular neuron will come as an output saying that I have identified that the image belongs to number two now the image will belong to number one so this one will be highlighted so this is a very simple example of classification do people agree this is what we do in machine learning also right yeah we classic the only change here you guys will feel something newest how do I fragment this number seven into the inputs how I give it to them what happens here and how the output is it that is what is the today’s challenge I will explain you guys how it happens another way of understanding a neural network classification could be as simple as this if you observe there is one image of a dog which gets penetrated into some of the environment and finally I say it’s a dog this is how we do object detection you give any object if the neural network is trained on that it will be able to identify no matter what breed what angle we have taken the photo whatever it is you will be able to detect that it is a animal which is a dog okay so now there are three layers which are most important here one is called the input layer one is called the hidden layer and one is called the output now let me shift you guys to that people file and we will simulate something so we’ll start from scratch yeah how does a neural network work as I said we have three layers input Haden and output so please remember as we are talking about neural networks the complete operation whatever it is it should be in the form of numbers no matter if we are dealing an

image or we are doing voice or we are dealing text whatever it is image number goes in number comes out yeah so let’s see how it does so the first thing is your input layer so represents the dimensions of the input vector now let us say I have a data set which looks like this it has around one two and three columns independent data and this is my target data yeah so if I have three columns that means I am supposed to have three inputs to this that means each neuron in input represents one dimension all right good and each neuron in the output represents my target column so let us say in target I if I have zero and one so this neuron is supposed to detect zero this neuron is supposed to detect one that means if I give a combination let us say the combination of 0 0 0 that means when all three of them are 0 it is going to highlight that this is your classified outside okay so this is input and output this is how we decide input an output now comes the hidden part of it now what exactly is a hidden layer hidden layer is randomly I will not say randomly but it is our choice that how many neurons you want here how many lists some some of them have 10 layers some of them have 2 layers for lists whatever these are nothing but the same neurons but the only thing they are connected in some other other different way all right so if you’ve got three less input/output in hidden any shoes on to this simple don’t worry about the calculations we’ll see but so far so good all three layers are ok ok how we yeah now we’ll talk about something called activation function now what is an activation function if you observe this PPT carefully you will observe that not always all the neurons are on did you see that some of the neurons are off correct and in output other neurons are off only one neurons on so you might be wondering okay how this happens this is the job of activation function activation function acts like a switch now let me define one activation function right up let us say we will design a function which says if my number is greater than one then I will activate that I will say pass the number to the next neuron if the number is less than one I will say switch it off now in my input if I give a number of one point to what’s going to happen this is your function you will choose this one and you will say that that current neuron is active pass the number to the next neuron allow the number to pass through these are called activation functions this is a simple one I’ve shown you okay let us dive deep into this activation function theory so there are so many functions available like this we’ll go one by one before that let me show you what happens in the neural neuron so imagine this is one neuron this one and this neuron I have represented here for example any of the neurons all of the neurons have same kind of look what am i doing and giving some input to anyone this is my input X is zero there is a weight and if you ask me what is this weight I will sell randomly chosen number okay during your forward propagation when your your network is built we randomly assign some number say our andn and we put some number yeah that random number gets multiplied with your input to become W 0 X 0 that W 0 X 0 from the other neurons we get W 1 x 1 W 2 X 2 so these are inputs from other neurons all right all three of them are coming and converging here so if you want to see an example just have a look here so if I say we are talking about say this particular neuron how many inputs are coming one two and three so this is my X 0 X 1 X 2 alright so you can imagine that way now inside this cell inside this neuron what do we have we have an activation function we call this as an activation function here F and also there is one random number called bias which we have to add it with our addition of all 3 of them so if you observe here what is this w 0 X 0 plus W 1 X 1 plus W 2 X 2 plus some random

number called biasing factor all right now if this particular operation is greater than certain value the L it is allowed to pass if it is if it does not satisfy my function it is not allowed to pass ok now let me see let me show you what are these functions so if you observe here what looks like very complicated activation functions but actually they are very easy we’ll start with the easy function this one have a look at this unless and until your output is greater than 0 it will be or else it is 0 this is called step function second one is sigmoidal function have we done this earlier in any of the algorithms can you reconnect this have you done in sigmoid in anywhere else regression regression logistic regression allow this correct log R and one more kernels oh yeah SVM kernels you remember one of the kernels tangential sigmoidal curve yes what does it mean if your input is between say this phase and this phase it will continuously increase if your input is less than this the value will be zero and if your input is more than this it will be a static value of one if you expand this a little bit more okay and one more thing this is not real loop I don’t know why I wear it already over here this is wrong this is sigmoid and if you expand this little more on the depth part of it but you have to cross through zero that becomes a tangential function alright third part is called rail okay so what is rail you basically rectified linear unit so if it is less than zero it is permanently zero if it is more than zero it is a linearly increasing curve this is one of the most popular activation function we have used we are using actually so depending on what type of inputs you are having you can use one of these so down the line when I train when I will show you guys how to write case studies and how to decide your networks at the time you will try to choose all this stuff as per our data yeah and there is one more important activation function called softmax whenever you have a categorical output now we just saw that there are two type of possibilities over here the output could be categorical and output could be regressive correct same thing we have done in machine learning also if your output is categorical the default function activation function you will use only on output layer is called softmax it is a categorical activation function if in case you have no categories you have regressed outputs regression as an output in that case you can use any of these are we clear all of you true yeah so how what are these functions and all we will see when I the stuff but for now I’ll be clear we have sigmoid and relu and threshold these are some of the most important ones and why do we need it just to manipulate our number and let it pass for me that’s it all right now moving on Duke so this is how overall on a very high level neural network works now let’s try to simulate one so I have simulated a neural network in excel file now let me show you what it is now please we cannot do this on a higher level but on a very simple neural network you can’t train it now observe my network so as an input what do I have I have got some alphabets from A to Z now please remember computer does not understand what is a what is it what we do is we do somewhat cheating at the back end we tell the computer that if this is the number it should be displayed as this figure if I get this as a number it should be displayed at this figure everybody agrees to that yeah binary logic and all if you have studied in the computer science and this is what it is same thing applies in neural networks also so say for our sake what I have done is I have taken A to Z and I have divided numbers between 0 and 1 in 26 equal distances all right what is my job now is my job is to train a network such a way that when I give when I say that WH Y okay in English spelling it should also give me the same numbers as a know but in the form of numbers and then we will map it up in the form of digits oh sorry in the form of letters all right now let me show you what exactly are these calculations and what we have done so before that how many inputs do I need

so if you remember I say total number of independent columns is equal to our inputs in this in this case can I say I just have one independent column which is a agreed so I will need only one input how many outputs do I need it is ok first of all what is it is it a rigorous or on a classifier it is a third classic mmm look at the numbers okay one way you are right because we are passing ABC but in this case we do not have softmax in Excel files so what I have done is I’ve made a regresar and then later on we’ll we will try to classify but perfect runs so it’s a classifier with one output so this is one input one output all right and then you may ask me okay what are these other things these are my Fidan layers you will see down the line what are these and how I designed this all right now let’s try to pull one number say I pulled out W what is W equal to 0.847 two that is what my actual value is this is my actual Y all right now what I am doing is I have to choose some random weights new people remember we talked about weights randomly chosen data now let me randomly choose some bits so I will make it say 0.8 4 5 4 245 say 80 67 some random number whatever we like alright so I have chosen some number as my weights now what are my weights way do my weights like my weights will lie on each of this line so this particular weight 0.7 that you are seeing is a weight given to this line so when my input 0.847 2 goes to this particular neuron it gets multiplied with this weight are we clear on to that everybody what are these weights this is my biasing factor and these are match you’ll beach okay nice do let me know have you okay with this any any issues there poetic or when the concept is fine but just from just one thing I don’t understand how this random number comes because when we say that the output is dependent on the input as well as the weight which is coming up so the weight the number which we are choosing for the way it is something which will influence output also exactly perfect so what is our job then to back propagate and adjust the ways such a way that my actuals and my prediction should match okay so basically that’s part of learning in this case so cooperate the brain with it okay exactly the whole game here is to adjust this weight such a way that the output that we get here matches my input all right do you see now output input matching not at all right look at the loss function that I put our MSE that is around 5,000 plus or something agree so whatever made so sorry what is what is the problem that we are taking here the actual and the MSE yeah okay so the first thing what I did was I multiplied my this with my current weight and activation function I have used sigmoid onto that now you have you may ask me from where I got sigmoid it very simple we wrote a logic here or a macro over here to simulate sigmoid now what is sigmoid on a very simple scale sigmoid is defined by 1 upon 1 plus E power of minus X what is my X X is one of my inputs given to the sigmoid function that’s it so what I’ve done is I have my own sigmoid function where what I am doing if you observe the formula bar be 5 into this one okay correct yep mine my this one what do you say bias factor and I sigmoid it up that means I whatever manipulation I am doing down the line I am putting a sigmoid onto that and when I put a sigmoid onto that the value that I get is 0.01 to which I have perfectly spoiled it up because my original value is 84 and the value which I do after multiplication is little weird okay then so this is my n1h one same way I take this up multiply it with next weight put a sigmoid onto that and get this okay third one is n 1 H 2 so this one so if you remember if you see here now now we are getting inputs from two of them we are getting an input out of n 2 H 1 and also W 3 same thing whatever goes out gets multiplied with the new weight 3

will put a sigmoid onto that as an activation and will throw it out alright so our data is propagating from here to here so when my this particular data propagates from here to the output it becomes minus 42 and if you ask me why it is all because of junk weights that we have put agreed everybody agrees to that the weights are causing this issue right so what what we should do now if weight is responsible for making this issue we are supposed to go back and change the waves now how do we do this in neural networks we call it back propagation I will explain you with formula now but for now in excel file what do we have is me and site data we have something called solver I don’t know if you people have used it my intention here is to reduce my rmse now everybody knows what is our MSE right root mean square error it’s a loss function what do you mean by loss function difference between my output at predicted output and actual output that is what we call is loss function everybody is okay with this our MSE all right now what is my intention here is to focus on my are MSE here and my intention is to reduce my animus II agreed yes or no yes yeah yes we’ll go to the formulas now after this we’ll I’ll show you in derivative way how to do that next what is my intention here what I need to change such a way that my RMS it should reduce can I say all of these junk weights I have given those weights over here now let me click on solve it will take some time but you observe the way it’s changed and the weights when the weights are changing now can you look at my rmse from 5,000 odd it has come down to almost zero you agree three point 2 into 10 power of minus 6 is almost zero yep and look at my output now and look at my input now can I say they are almost in sync yes or no so can I say I have designed a network here such a way that tomorrow if I give this particular value it is going to identify that it is W or at least if I give this value it is going to identify it is are you able to visualize that enough let test it let’s try to test it so for example I have given a value s here s is what what is the value of S is 0.6931 okay so when I give the value automatically all the calculations are done and the output that I get is zero point six nine zero four zero seven five how I got this using these weights and the error that I currently get is zero point zero zero three negligible amount so this is your forward propagation and what we did here using solver was your backward propagation one forward plus one backward represents one epoch in a neural network so neural network continuously goes front and back front and back front and back till I find my loss function to be minimal till my weights are adjusted to the very best values this is the high level overview orphan unit so on the solver whatever it did right so it’s this is very very easy and good to understand the solo was just trying to adjust the weights and keep keep iterating is it probably it will done whatever rate to ten epochs to get the right value current okay how does it I just adjust is there any way yes there is a way to adjust that so let me show you using formula but before that I’ll be clear with two processes forward and backward now if I say fall forward prop and backward prob you will be able to visualize what I’m talking about forward prop is your activation functions backward prop is adjusting your weights alright yes we’ll come to that we will come to that one by okay don’t worry about how do we put it and how to choose this and all will go deep this is not possible in Excel files they have a limitation of only using sigmoid so if I use relu so if you look at my rail ooh if you look at the rail you any value which is a negative will be switched off

agreed yes on us we will will this rail you allow us to pass any negative value I will see this is vital information and one more clarification can you just go back to that picture yeah say like from input to this NH one this w one is getting passed right is the activation function work here or from image in one h1 to win will it’s only the activation function will work division here activation function resides here and one h1 the multiplication happens here and then it goes to activation function is okay so from input the activation function will not work one lemon it goes away 1 x2 the activation function we get yes yes perfect perfect good now let’s come to some heavy maths so now you might you people might feel getting little heavier so what do we do so let me explain you forward propagation in a more complex way now so now when I talk about weight slayers activation function now you will be very clear ok so now will not repeat what is activation function now let us say I have got a weight or input X X gets multiplied with w1 weight of that line and I get alpha one which goes inside an activation function I get alpha which gets multiplied with my next weight and gets alpha 3 alpha 3 or a 3 goes into new activation function it becomes a 4 a 4 gets multiplied with some weight and I get my output this is your forward propagation okay now the same thing you can see over here this is my input this is my weight 1 and if you remember I said there there there could be a biasing factor it is again a random number we use it so this into this Plus this becomes this then I pass this through a sigmoid function so when I pass point eight one through sigmoid I get 0.693 that is your alpha two this one same process happens till you reach your output so if you now observe our input actually wall 0.84 and because of this whatever we did in middle we got some loss of 0.8 three numbers what random numbers random numbers if I change it if I this it will change did you observe its changed anything actually float is better integers there is a function which will allow us to put only integers but then there are some problems with that I will show you why all right okay so this is what is your forward propagation now any questions on to this how this value became this value any questions perfectly here so you imagine if there is a bigger network like this say you have a bigger network like this now if I ask you guys you know what goes in and what comes out are you clear with that forward prop yeah yep good now let’s talk about what happens in the backward prop and how those weights are learnt now so before moving on to backward prop I will ask you there is a very simple question what will type of loss functions you people know so far what are loss functions we know let us say at least three of them will talk about what do you mean by loss function why pred when compared with why actual that’s your loss function how do you compare it subtraction with distance F are there agreed MSE and si are MSE yes these are some of the loss functions that we know about right now let’s come to the complex part of it now how actually we learn the weights now it was very easy for me to use solver and show you very easy because you have a ready-made function it worked but what happens in the neural network how

network and the stretches so see this is your output okay so this is what you are supposed to get and so this is what you got and you are to get why as an author this is your actual this is your predicted so if I have a function which will take care of overseas why I will define this L as my loss function okay and what we need to adjust you put the things which are in scope so my current for my back propagation only the highlighted part in red is in scope that means we need to alter one of them or all of them or combination of them such a way that my loss function is decreased all right now which mathematical function helps us to do that my intention is by how much my loss function will change if I alter these what kind of function we can use here from our full days you can recall some functions we have done this you people know it but me we don’t know the real application of differentiation perfect of quickly identified we will say it by how much my loss function will change if I alter same wave number five everybody who is okay with this differentiations yes it’s a step function so how by how much I should change so that this will reduce very simple now if I want to develop a chain rule here okay now how do we do that so now you may ask me fine we got a differentiation but how do you say which way to change and how do you derive the equation so for derivation of equation we will put here so it’s not needed but it’s just for your understanding of backdrops I will say by how much my loss will change with respect to my actual output that I’ve got sorry a predicted output that effort that will be my first differentiation multiplied by how much my output will change with respect to my rate number five agreed multiplied by let’s write the chain rule how much my wait five should change with respect to my alpha for how much my alpha 4 will change with respect to my alpha 3 how much my alpha 3 should change with respect to my weight number 3 how much my weight number 3 should change with respect to my alpha 2 agreed as you’re getting it how much my alpha 2 will change with respect to my alpha 1 and how much my alpha 1 will change with respect to my weight 1 in P go anywhere further no right because wait and X is not dependent on each other in any way I did now just think about it if you want to just focus on how much you should change this wait what we should do cancel this and this what I will left with and I say we are left with dough off l / do off w5 agreed yes now keep canceling what if I cancel this this descenders gone descenders gone now what are we left with can I say doe off l / though off wait number three yes correct again keep changing it keep canceling canceling what are we left with doe of L with respect to do off wait number one simple so now can I say I have got a method to find out how to control my loss with respect to all three of my weights same way you imagine a larger network where there are so many weights the same rule happens and the weights up optimized in the backpropagation all right simple huh so you know we are changing the layer 1 wait you can fix whatever you want so if you say you want to change all of your weights you can change all of them if you say I want to freeze this weight and change these two you can do that as per the chain rule form right okay well you’re understanding the first week you can assume that yes all of the weights are getting changed in one first backpropagation with automatic calculation do be absolute specifies I’ll go it automatically one time and what it says automatically in Python we have something called optimizers they take care of one line will take care of

complete optimization the way we did it in solve are no in itself same way we have it in white yep once one line that’s it okay so today I will show you a neural network which can be designed in three levels that’s very funny your network classifier very simple you don’t need to go very complex models you can even do it a very 3 to 4 line that’s okay now any issues with back prop good alright so last one just from my side I know we have done this softmax function in somebody tell me what was soft max again one more thing this all the calculations everything in them into fights and fight and you don’t have to do anything so whatever mats I showed you know it is why you’re on this in Python within seconds you will be able to train your network right ok final one no more theory now we’ll go to practice what is soft max you people remember what was soft max okay soft max is an activation function please remember don’t forget this here soft Max is an activation function to be used only when you have an output which is a classifier if you have an output which is regressor you can use any of these alright and you don’t even have to worry about the formulas you just have to remember the name that’s it one name in Python is enough to do all of your calculation good so now can I say at least 30 40 percent you guys are clear on what our usual Network star like like softmax is mainly used for multiple categories right for Helsinki for Cindy classifier can be used sigmoid will that we find next yes you can do it for a single classifier even you can use a step function here okay Oh zero yes yes zero or zero or one that’s it even you don’t want to use the classifier no problem with build a filter of our own so we take the output and we’ll say if it is this answer is this or it answer it is anything but yes you for two classifications maybe you have two class defines you can use any of these not at all click on one question it is from the the video classes so in that video he told me that he told us that because the function as a constant value so we would not be able to calculate the flow and hence the grading method would not work yep my question is why do I’ll be using rail you and it has become more popular because half of the real you is also a flat like that do not be able to pursue the gradient ascent or better no see if you look at this rail ooh is continuously increasing value is not reaching or radio is not parallel to x axis did you see that negative value it’s flat correct correct perfectly right if you ask me yes yeah if a value zone we’ve not be able to create the gradient decide and I’m not be able to correct position no in that case I will say me let me correct your let me your approach here a little bit please remember we are talking about a fully connected Network so if I go back do you think this is fully connected yes each neuron is it connected to the next layer so even if one of the one of the function is clipped out there is another function which will help to pass it through you getting my point let us say this is a negative value yeah this is a positive value and let us say this was a real ooh in that case you are perfectly right if only negative value is going inside it will be clipped off but also along with that we are adding one more wire which will allow us to go forward correct since they are having two inputs there are chances that one of them is positive and it’ll allow us to go forward that’s it yes yeah once we move in the same concept would apply into the step function as well isn’t it he so he’s in the step function as well one value is positive one man is negative

when if one man is negative so we it would allow us to we go through the next next iteration but the way it was taught in the class it was from the point of view of the gradient descent so trying to minimize lost and you know in order to minimize the loss we need to know the slope of a or the gradient ascent of the location where we are in the plane and that is positive or negative that would depend of they will need to go in the real function and half of the times you will not be able to calculate if they we are mean this see in the proponent of a gradient descent I will not give all the model correct perfectly so now yeah yes possibly something no you are perfectly right Andre Lu and you perfectly your analog is perfectly right and created but I will try to reshape your gradient decent understanding once more so then you will be able to fit really one yeah but here’s you are perfectly right if in case the enter a negative zone definitely you are not going to get anything out of okay so let’s now try to get a very very detailed or I’ll say a sneak peak on do gradient descent so what as for you guys is gradient descent what exactly you mean by that in somebody quickly let me know what is your understanding on gradient descent basically calculating the slope of a V a V I in a particular plane and trying to minimize say trying to move into the right direction I love you go into negative right additional to the positive you know to minimize the loss or any function that we’re trying to okay perfect now let me reframe that let us say let’s not use technical too much technical term like slopes and orders when we do that people get confused so let us say these are my weights alright and this is my loss function there is one particular place where the loss function for this particular weight will be zero agreed every weight has one point where the loss is zero at least let us try to draw a diagram let us say we have a gradient descent diagram like this for example okay and also we’ll drop a line over there saying that this is the line of optimal weight now there are chances that you might be here when you chose a random weight you might be here so when you are here your loss is pretty high really or else there are chances you chose a very high weight we are not sure about it at the time for your loss is pretty high in both the cases as well as you were saying is you have to either reduce the rate or increase the way direction of the gradient listen in this case my direction is towards right in this case my direction will be towards left all right depending on the slope and on whatever C was say let us say there is one number or there is one factor called loading rate I tell my back propagator yeah whatever I am going to use in back propagation I tell him that every time you make a mistake every time you get a loss you reduce or you move by some number X so every time I have a backpropagation my rate value goals by X values down or in this case by X values in this side these this keeps happening till I reach a spot where the losses in control or I have reached an actual position of zero so this is on what I kept now the lines that I kept here these are called sweet zone sometimes it is not possible to reach here now I’ll tell you why let us say you have chosen a value of 0.1 and let us say when you kept doing back propagation one time you have reached here now if you add point one to this you are going to end up here and if you end up here what happens in next back propagation you subtract when you can subtract the learning rate you are going to end up here so this is what we call the oscillation zone there are certain time because of learning rate we will not be able to reach in that case we can say that we will put a boundary saying that yes we are ok with this set of gates this is what is my take on gradient descent are you are in sync now I think you people also have the same idea right the only challenge here is to understand how to choose this and how many I box on an average will take me to reach over here how many back propagations will happen to each of them

all right yeah so Roger one thing is you are perfectly right Andre loop so you’re gonna clip the negative part of it but also remember we are not using only relu in the network one layer could be using rayleigh another layer could be using a sigmoid also one layer could be using soft mixels since it is a combination of highly connected lengths it is sometimes okay to ignore certain neurons even if they are creating a mistake it is okay we can ignore them and they will not create much nuisance in backpropagation because of this this gradient is will fall I hope you got you got my point already yeah I just wanted to you know that a they’ve something other than available – yeah I just wanted to clarify that there’s nothing much I cannot do that I might have missed yes so it is very good I am very happy these you guys was my other batches we’re not very focused on activation functions but it’s very good to know this because if you have a good idea what is activation functions and when you use your neural network designing becomes very easy because you have to and make this yeah so you have to choose activation function there yep all right perfect good so now let’s try to design one of them so we are not done yet there are some more concepts which I will show you one by one so I think week 2 also have covered almost using this so I think your next week’s video the week two videos will be very simple for you guys now before moving on I would like you to introduce to some of the libraries within networks there are many actually and now there is a big boom every company is putting their own library yeah so the most important library for learning is tensorflow the latest version of tensorflow now it is just released it is 2.0 I am currently working on 1.0 I am yet to migrate to 2.0 it is like you learned Python 2 and suddenly Python 3 was launched ok so it is very critical for us to transfer this because some of our codes are very hid and since this is new I am also waiting for it to be stabilized and I am also waiting for all the solutions available online after that I will migrate the codes will change so this is tensorflow what is the use of tensorflow tensorflow is an end-to-end neural net library which is owned by Google for now it is free not sure later but it is owned by Google and what it does it gives me the statistical power so whatever back forward prob back prop activation function weights whatever you saw could be managed by tensor flow the calculations could be managed by tensors good and also if you observe there a certain hardware that we need we need some virtual neurons who can hold the values locations right even tensor flow can do that we have something called PF lon which can do that but to avoid it what do we have we have a better version we have got Charis Charis is a high level API which sits on the top of tensorflow and simulates a neural network for us good are we clear Enza flow versus chaos you can build a whole network here also without chaos why do I use chaos it’s easy to define impact alright we have one more called PI torch hello people heard about pythons any time yes yes I saw some very powerful library very powerful only issue it PI pouch is it’s a low-level programming language this is a high-level programming language like somebody just asked me that while doing back propagation do we need to do this manually I said no why because I have a simple command which will help me to do it whereas in PI toss you have to hand build everything not now now PI torch is evolving yeah so pi thoughts is again a very popular library for academicians and researchers so maximum majority of the science when you look at a AI based paper especially from a university or a research company they will be using white or short one why because they can handmade things they don’t have to depend on this I am a big fan of tensor flow so whenever this upgrades happen now I will be bound by this I have to upgrade it whereas in this

and handmade distance but it’s a low level one means you have everything beautiful from scratch okay so this is Phi dot apart from PI touch there are so many libraries I mean like if I if I show you guys a library called say Swift for example so whenever you want to deploy your AI on to iOS or Android this is one of the most popular libraries sonnet sonnet belongs to one company called mind learn if I’m not wrong and these company makes or publishes lot of AI based papers you can use their codes to learn the only issue is they do it using their own library called sonnet they do not use stencils no colossal so we have triple nothing like that there’s so many of them away alright so the most popular ones if you see now as for your starting purpose these two are very good please learn them thoroughly and after you master that you can go on to PI taught you so you have a complete explosion of all the libraries good so these are the libraries especially you should know the difference between those two so I’ll give you one example recently I was taking an interview was one of the candidates and when I asked him he said me I am from a back and this is the harsh reality of our field everybody says yeah I know artificial intelligence so I very simply asked him okay what are libraries you have you are aware about so he told me tensorflow and caris yeah so I asked him that can I use tensorflow independent without using cos he said no this is where people will pick you up saying that you have no idea about here so out of ten interviews that I take on AI nine of them I reject it because of these reasons so people would have seen the codes they would have done it but if we don’t know the basic difference way to use what then it is very difficult to build a networking ta because it’s not a straightforward method to be very front so guys please be careful with these two and try to learn documentation part of it so they have both of them host online some recommendations please keep it handy yep and keep in habit of refreshing your knowledge on both of them every three to four because they keep updating it they do lot of updates one question here as you said in this lab we’ll keep on updating so I was working on one library that they’re compatible little Python 3.5 so I’m able to run the template code on gu4 lab but I’m not able to run on my Python but the library so what is the solution out of that how do we take care of this kind of problem yes so what you can do is let us say this is the code that you are you have put your pythons a 3 for 3.5 for example and it has got everything underlined what we can do is we can pickle this you convert this whole thing as a function if possible and we will pick a letter that means we’ll store this function and we will use it as a how do you say API get my point so we will define some inputs and outputs of the function so say you are given like how do you say input file is linked so we have to name the file here and that file has to be worked on yeah so there is a function called pickling that is down the line that is the worst thing you can do but if you can see if it is understand under its Python 3 I don’t think so there should be a big problem did you check it up is it because of python version or anything else is initially are you sure it’s finally come for you here so I was watching a few videos on this in this particular library so most of the those videos were on 3.5 Python of it before hey that Python so they were clearly saying that it should be working on it by some 3.5 or 3.3 like that and that’s obvious calling it so I have Python 2.7 when I am Kryon is running it so it is showing that it is not a in addition to showing a lot of conflicts so I would easily I forgot the library name I shared the name with me I have another machine with me which I use it only for R&D let me down grade myself to Python 3.4 and then try and not even 3.5 I will go beyond that and let me try loading that let’s

yeah yeah oh yes there are sometimes possibilities like this sent you an example the child bought that I made no so I got one request from running saying that we want to use only TF infer flowed to build a chat pod we don’t want to use Cal so what I did is to use tensorflow end to end without Kerris we have to use something called TF lon okay TF LAN is not compatible with latest version of tensorflow because they have changed this library here yeah so what I had to do is I had to downgrade myself let us say from tensorflow terms of float to to say tensorflow one point something and it is exactly like what actually said as it is not working below certain levels so i had to actually come down to one level below that and then this thing fired up so what i did was i pickled it up because in a sense i know this is going to be a problem so I converted this to it into a runnable function onto any of this worse it was not that easy because the the files the stuff has it is a function commands like input TF learn but if I’m using the current version of TF in my Python definitely it’s gonna give me an error even this solution might not work that get yeah so yes since we are using something free this will happen and it has happened many times in my career where this is not working and you have deployed something so what we have to do is you to find a fix for that all right yeah one more question related to this so how is it be possible to downgrade the Python version or time salvation without losing the previous data that we have worked on yeah yeah you see if I downgrade I just need to know if I downgrade is it going to overwrite my if I have a control of where Python is going in copying the file if you know what I’m saying your C Drive it is going to create one folder if it is going to overwrite my current folder then yes you will lose it if it is not overwriting if you have a control over that to create another folder you don’t worry about it it’s it’s going to be as it is it’s all about a folder that’s so if I save my file into another folder and then try to download it then it would not have anything yeah I should not be wrong genius should not be a problem I mean it would I find a simple code usually kind of corner or fee if they do I have to uninstall Python and then install it again ID you need to uninstall Python and then you need to install the specific version or if just look it out just check out on Stack Overflow or geeks for geeks any of these websites is there is there a place to run two of them because I am using Mac right so in Mac I have an option okay the original Python that I’m running is two but I have an option to upgrade it to seven so I can run both of them okay in Mac it is possible so what you can do is you can have a Linux VM or if you can have virtual machine I don’t know whether you have done it earlier but if you have a VM you can do it easy you don’t have to worry about any of this but if you are doing in Windows life I will say store your Python as you say and then uninstall it and then reinstall the new version this is what I know okay but try to check on Stack Overflow if there is a command to have both of these versions and to choose which version to run I’m not sure about it how much storage you have in your back Krishna this is simple to see 128 this is five year old Mac so it is 128 30 okay you and I among the same thing so we just thought what it need to upgrade or not okay no no it is working it is working good okay but yes down the line if you if you talk about working on computer vision and all then either you have to work on collab or you have to get a lot which is GPU in it otherwise you put is going to die and in hours to just train your network but collab is now because of : lot of people are very happy we don’t have to spend money on GPU now but I don’t but if I go to aw so you need to pay high right aw yes now Amazon Azure that is Microsoft and Google they hate to head

on this in anybody wants to conquer this space basically as the reason Google is giving it for free right now so that they will get you addicted first of all for the cheese’s we are giving edu which is 32 GB imagine and there is no Python dependency there so you can go and install any version anything there you don’t have to worry about what you install your machine just no sink with it like but the only issue is tensorflow and will have both of them belong to Google so I’m not sure how long it is going to be sleep to be very frank and now as you is also picking up I don’t know whether you guys have heard about it as you is making an end-to-end analytics platform within their own cloud so it’s going to it’s going to be a big rival for our hey WS okay yep so if you guys can work III had a conversation somebody the Google itself so the idea of Google collab similar to Gmail so they want to move this free yes as long as poppies and calendar that they have a good year thank you good that’s good for us actually you know to spend on a GPU machine right yes because one of my learner’s from the other batch bought our machine specifically with GPU you costed him a lot actually but then this collab came in but that I feel there are some issues with collab also that it’s running in middle if it stops or your net fluctuates it restarts rather than restarting from where it was stopped there are some issues but yeah there this is a Vida version right so they will upgrade it ok alright so let’s come back here let me show you one simple example of how to build a network and then we’ll look close down there is effect prediction yeah ok see I have a very simple problem where I have got two columns one is some say number of days a person is working or a test tester is working for example and I have another one where I can say these many number of test cases the person is testing in this many number of days yeah and out of that the outcome the target column is defect prediction so this many of defects he is getting all of it yeah now what is my business problem here is what if I can make a predictor here saying that in future if I give these two values I should get an output like this okay let’s try to build a model onto that so what I’ve done first of all as I have used all of our machine learning techniques as usual just to show you guys the power of machine learning versus here so first we are using decision tree and since I have done pruning all not getting any overfitting issue and everything’s good as of such random forest is coming a little lower which is 88 boosting is 84 and the gradient is around 90 and bagging is 88 so if you compare all of this I feel bagging is sorry boosting is doing pretty good so we’ll choose one of them so we have made an ml prediction predictor with accuracy of say 90 percent for example so this is a very simple way of doing an ml so now somebody asked me write that this AI or a deep deep learning is pretty good then why are we learning M in these these are the things using which on a very simplistic level you can implement it you don’t have to you know go ahead and do back propagation for propagation and all that but some of the applications or I will say right now at least for next three years that is what is my prediction if you guys get on a real-time data science project also in your company the first two to three years you will be deploying only these kind of things to be very fun because if the company is new if the project is new and if the space is new the company does not take a big risk of implementing the high cost products when I say high cost it includes GPA it includes deep learning it includes training time lot of things are there which we have a really called as KP is for high risk products so what we do first we penetrate into clients trust by using these kind of simple micro services and once they are confident enough then we

take them to deep yeah so this is one example you can say that yes this is where ml has an edge over DF simple levels I don’t expect you know to do image processing on all using em as not possible yep now to counter this up what I have done is I have made a very simple neural network which starts from here so you can forget whatever we did about no problem we know that our part next what I am doing is from Charis and implementing a sequential a library what does it mean if you observe our excel file the simulation I’m sorry not this one the simulation that we did don’t you think this is sequential kind of stuff this depends on this this depends on this and this it’s like a sequence so the the first part to understand encoding is we have to tell Karass that our model is sequential one alright next is what layers you want under kill us so if I yeah what do you want under care us so to add a layer to keep adding those hidden layers no you have to use the command Holton’s next thing if you want to drop out and if you do want to do normalization even it’s possible under cannot so we’ll do it after that we don’t want SQL on and one one library has optimizer so this is used for back propagation or gradient descent or to adjust your weights anything you can call about all right now let’s go further so as usual we’ll import the data no change in that we will do we’ll define x and y no change in that the only thing you have to be very careful is even if the output is in the form of how do you say classes yet you have to do a categorical out of it category in the sense you have to convert this into category because if in case this looks like a category and it is not category neural network is going to fail so please be careful the output the Y of each and every data in your data set should be categorical no matter it is test or trying whatever it is so either you do it before split or you do it after speed after split the only issue is what I made a mistake here is they have to write two different lines what if I would have done earlier itself one line was enough to categorize all right after that coming on to the inputs now can somebody tell me if I want to design a neural network here yeah what should be my number of inputs so I have to have some inputs right so sorry Oh perfect two inputs and what will be my number of outputs one okay I did not show you there are actually three zero one and one – okay perfect so I’ll have three outputs now if I have three outputs what activation function I should use softmax please remember that no more relu anything so this is the input and output we design now let’s try to focus on the interior part so before that please remember till you are confident enough to visualize all this always print the shape of your test and drain so this will help you to design input this will help you to design output all right three outputs to induce the first thing you have to fire up is you have to tell Python that this is a sequential model and the name of the model you are defining is model itself and every time you make a change now you see model dot add what I need to add I need to add a dense layer what is the dense layer so first of all this is your first input layer now in input layer what we are doing is we are considering input layer itself as a neuron so you can say that yes we are having a neuron here and we are giving input to it so you can consider directly this as an input some of some sometimes we don’t take it off sometimes we start in new activation functions from it is this yeah in this case what I have done is I am saying I want an input layer of dimension 2 that means I need two neurons and I eat 64 rings now if you ask me why because my next layer is having 64 neurons if you remember each neuron is supposed to be connected to other one so like that if I have 64 inside I should have 64 lines of weights also agreed so this and this should be in sync next is I am saying the activation function I am going to use Israeli forma now coming to ok so please ignore drop out because this is the this is a topic from thirdly robot means it switches off

some of your new ones it’s a it’s a type of hyper parameterization of feature engineering in in Iran so please ignore for now these two next is I am saying is I will add another layer that is my hidden layer such a way that it will have 64 neurons and activation again is real and the third one is your output 1 output is I’m saying is I will add 3 layers sorry I will add 3 neurons as an output activation by default is sophomores what if I have user a Lu here I will get a regressed output if I use soft mix I get a classified output that’s it all right so this is your neural network on a very simplistic level 1 layer 2 layer 3 layer it is capable enough to deliver the same what ML gave us about next so we have done we have created a hardware so I can say we have created an empty Network now we are supposed to populate it now how do I populate it before populating it we have to design backdrop so I can say this is forward prop this is your back propagation back propagation is nothing but optimizes so there are so many optimizers available so if you want to see the list of optimizer you can just go to Google Kara’s optimizers ok this is the documentation which will help you to understand all the optimizers available within Kara’s I will show you each one in each case string ok rmsprop ad a delta ad a grad atom SGD all of them I’ll show you one by one so for now let’s talk about a very simple optimizer which is atom now if you ask me what is optimizer very simple remember doel by duo of weights the formula or the person who does this is this particular person called all right instead of atom you can have SGD rmsprop ad a grad lot of things we have depends on the application what we are choosing in this case it’s a very simple numeric application Adam works the best it is something like a rail loop a generic you can trust him for all the backdrops and so you have an option to define a non injury so that is the I keep it usually in two decimals 0.0 if you want to be written slow if you want to be very dynamic you can put large but if you have a large activation function please remember we just discussed it is going to oscillate it will not come to a rest point but yes your neural network training will be faster also in the larger learning rate so this is your atom now you are saying this is my model please attach this optimizer which is called atom the loss the way he is going to calculate loss function is nothing but it is a categorical data and we are going to use cross-entropy guys what is cross entropy we know entropy right what do you mean by cross entropy have you heard or I told you I think I have not discussed all right you know entropy all right so I’ll do one thing in next session I will give you some mathematic behind it and we’ll try to see why not do a Python examples there are so many lost functions available in neural nets one is called cross entropy if the data is categorical categorical cross entropy if your data is fast what do you mean by sparse there are lot of zeroes as an in the output in that case you can say that yes it has sparse underscore category underscore inter cross entropy you have multiple of them available for now whenever you see category use this one this is the best one who’s that what is what should be your checkpoint so I am saying accuracy will be my checkpoint do we know any other checkpoints we know precision we know recall we know support we know f1 score right so this is what you can use always I will say use accuracy accuracy is a very reliable factor so far next what I do is so my back prop is ready my forward prop is ready both of them are ready so you fire it up as usual from your machine learning fitting so just take the model fit it fit on to what your training data okay that size now this is something new that size means if your RAM is capable enough to fetch the input at a time I would like to fetch input for next 80 readings you know so if you have a good ramp even you can put 800 so if you are doing a computer vision image processing and all that we usually keep bad sizes very bit so that the fetching is faster in this case you can choose any random value like it’s something like our seed if you remember random seed but here it

is about fetched executions next comes your epochs now what is epoch please remember one epoch is equal to one forward prop plus one backward Rock that’s it now you may ask me fine how do we choose the Sun so what I would say is first of all you choose single-digit epochs a five six seven like that and once you do that look at the first five epochs so let us say these are my first fight epoch still here look at the loss function yeah can I say it is having a huge amount of change you observe there is a considerable amount of change a change a difference of two is huge enough so if you find in first five or 10 he parks that the changes are very huge then you say ok let’s use fifty hundred two hundred like that or else you can check the accuracy if you find the accuracy is continuously increasing by a good number then you say that yes if I have more deep ox more accuracy I’m going to get that means more number of he parks more back prompt more backprop better the weights better the weights nearer your accuracy is near at 100 riyals so in this case when I ran for 100 times I got an accuracy of 84 it is not as good as ml because in ml we got 90 but this is a very dynamic model what does it mean tomorrow when my inputs change let us say it takes no time for me to retrain this I know even our ml takes no time but in ml you could hyper parameters and all those things feature engineering and we take a lot of pain to again redo in this case it will auto take the changes and auto adjust the weights at the end it will try to show us results of many rules there are certain weird cases where you might find ml has an edge over this yes but more or less deep learning should be good all right now what do we do let me show you what is soft max so if you remember the output was soft max how many three of them so look at this this is the predicted output that we have got now what are these these are nothing but your probabilities soft max will become so all three of them will get a probability so let us say if I have given an input and ends up I am expecting an answer as three that means one two and three so soft max will give me a probability what is the probability that the answer is three let us say the probability is 0.9 this is point seven this is point three the softmax will pick up the neuron which has having highest probability so just to show you guys this is the probability that we have got and if you see the highest probability belongs to this particular question so the answer is zero so my classes are zero one and two all right you don’t have to do this just for your showing purpose or softmax that’s it and finally you do your testing part of it so you take your model and run it or you value it on your test data so when you do it I get an accuracy of around 88 so which is a little more than 84 I know so what we would have made a mistake here is why I shoot it up high is because of bad size or number of epochs or learning rate so you can go back and just adjust it somehow so I can say that yes this works good on my testing data I got an accuracy say of 88 was perfectly good so this is how you can train a very simple neural network so I will again request all of us pick up some of your machine learning or data and try to build a model like this and see how you are getting and then try playing around with more or less so now if you want to add one more layer what to do very simple copy this put it now when you put it you can keep the same number what a problem now you may ask how did you get 64 there is no logic for that so what I do is say I start with some number say I would have started with 100 if I have hundredweights my next layer should have 100 neurons to take it if I want my numbers to diminish I will say my next layer will have 50 neurons you can have one more hidden layer here saying that my next layer will have 25 neurons if you want one more you can say my next layer will have say 50 neurons you can even double it up not a problem and from 50 it will converge to thei yarosh alright so some of my other batches is also asking if I can give some mathematics behind it how to design this I will again say there is no proven math that how to choose this numbers optimally but yes I have got some links to some blogs where they have done some equations but that is specific only to

date it so you can just take some learnings and more or less this is the way you identify clear if you try this up on your old data and let me know how comfortable you guys are on forward and back alright so I am done so I think we have covered for one and half week now because forward back and all where part of our next week but good I feel if I show you something you should also understand so I don’t want you to miss out anything yeah so I think we are going good we will almost complete everything by next week or by third week starting so we will have some extra wanna and try to show you some deep learning more networks wonders yeah image classification takes classification will not good see down the line we do some social media analytics and all those things using deep learning got any questions for me already Christian unless and until we find out the correct number of hidden layers do you think the ml is going to advantage over a plan correct correct so what one of my learner did was he is very ambitious enough to find the answer for this what he did he did grid search on this he said he searched he put lot of layers so let us say in grid search he said okay I won number of it unless could be 10 it could be 20 it could be 40 like that he started and he ended up spending two or three days for the neural network just to run and complete the grid search that if we are confused about or if we have we have already obtained of mix matching this stuff up yes ml has an edge on to this but ml cannot work or I will say supervised learning cannot work on to computer vision and NLP so by default only backbone for image processing and expose okay okay yes so anyway people come back to this but yes if we don’t have time to and patience to basically do this you have ml has an edge on that I agree with you perfectly and I am clearly saying here in this case study I still feel boosting is having a big edge over the other one because it’s a very simple algorithm in four to five steps it is giving you everything whereas in ml we had to do a little bit of neck stencils perfect but or the long run this is far better I feel because it auto corrects auto launch itself auto corrects his own rates also yeah okay okay so again to summarize very quickly the whole game of neural network is to adjust their weights that’s it how do we get this weights we will do in v3 what a different type of weights for now please remember weights are random but also we have a provision where you can control your weights we will see down the line in your third week when you do hyper parameters all right and second thing is there is two type of propagations happening forward prop backward prop forward promise to spoil your output backward prop is to adjust your own or adjust your loss function good next is you have three layers input layer which should be in sync with number of columns or features available output layer should be in sync with total number of categories you have on your output column good and the hidden layer again is a big mint I will say start with a with a number like 100 120 went through and then keep reducing it Hafford divided by 3 you are divided by 4 orderly but yes it’s a cumbersome process in one short see I got this but not always we can get of 84 percent accuracy in a short to be very slim and still I stopped at 100 you know if I don’t want further I would have get got at least 90 plus accuracy yes so this is the overall summary for noodle it looks good so now can I say the the the concept is little that the border is gone written less on the understanding part now I think if you go back watch the videos noting the theory what professor is excluding computer computer everything is going to go on a different level we may need to use glasses to

correct that visible exactly yes next 5 to 10 years the the r4d so often yes where for machine Nonnie now they arise come for deep flux because you remember I showed you by Auto ml do you remember that code which will ought to take our inputs I don’t know some of you were there with me what why – so you can easily there’s a machine learning data scientist who knows only machine learning with that piece of code if you want me to show it again I will show you guys again that thing so only do my work and some of the codes that I do here for my newer batches I already use Auto a mine under that now I don’t need anymore I will use what I am meant to get my predictions and it is becoming very common now even Google has come out with something like an auto ml which is similar to as your already has it so now if you want edge yes deep learning is a next one because we do not have auto dealers of such it will take some time to come okay so take it in a slow manner we don’t have to rush that’s the reason we have this design this particular module so that you get adjusted to noodles and we will go or we will run basically from computer next one all right so please understand each and every factor here don’t miss any my new detail so any issues anything drop no post on your group I will try to answer that yeah and what I showed you today was very simple next session will scale it up we will try to design more complex models okay I’ll show you something want to text an image processing also right away in next session so that you get a good feel of not only categories and regressions but also other two but I’m not sure how many of you use this for iteration and categorization who’s already you have a mint for that so the edge comes in something onto image detection phase detection expression detection sentiment analysis but anyway if you do not use neural networks for this detection right you use convolution only right Chris no no no no I will show you so your first project for neural network know you equ KN for image processing for image detection okay oh yes and you are perfectly right and volution is the default one but we will make you use neural network to detect an image and opposite to that we will make you also do tienen to detect the image and then the reason we do it just to show the difference between ml nd you can use VM to detect the images or to I can say not detect I will say classify images so if I give you a number seven let us say you have an image screen shot right in that image has something like this I can use SVM to identify what it is and display this this image represents number setting you can do yes okay but convolution is a better way to do it so we will see in next model why it is better what makes convolution better than views and even we have better version and convolution we call it our CNN regional convolution neural network we have more better version this than that we call it yellow and ssp and it’s still evolving we are not happy with that also it still you know evolving now we are going to complete it week session or eight week module onto computer vision it’s a very long model will spend two months on to that alright okay effect so I think since let’s move on to I shall take you to your code and I have a very similar code to that I can help you to give that it also does the stock prediction only predict the next ten days the model is basically predicting values for the range of data I provided I want to predict the future values okay so you are using LST mg/l right I saw you use it so is there any issue here it you as such so if you go to the bottom of the food it is a chart that I have made it if you go to the bottom of the road yeah is it art so basically what is what is happening if I see the chart itself what is happening is this is even that happens in the market and the SKU is reacting to that even in a delay format so for example let’s say on a day

four or five the blue line if you see it Peaks out and after that the red line also Peaks out with it happening in Adelaide format so just doing repeating it is just beginning what has actually happened not predicting the future correct correct correct got your point called your point okay then I need to analyze the whole code step by step why it is happening so I’ll be missing any values here or the network is a shoe we need to check this it was a ok code because I just you borrow it for my display see then you try to put it together okay we’ll do one thing Ashish I have a solution for this meanwhile I work on this correcting this code I will give you one more code of mine so just between us sorry you can just modify it and then use it yes learning lsdm see yes so I have something like this which predict stock markets yeah it’s a little customized code so it is almost similar to what you’ve done except what do I have here is I have got a huge amount of corpus knife and look if I show you the corpus that I have might put exactly same as yours but you can use it one minute deep learning LSD and time series talks so this is what I have okay and this is the way each companies talks are there okay right so what I will do is I will shape its code with you it is a self explainatory code if you know LS TMS then I don’t think so it’ll be an issue for you what I am getting is a I’m getting yes next 18 or 20 days of prediction I am getting it so this is my prediction busy yeah so try it out if you are able to fix take some contents and fix it parallely I will work on line by line of this because I need to understand the whole flow first of all this sort of simple you know task again thing I need to plate this up because usually what listicles do is lsdm has two parts I don’t know if you know this a little or two parts you can use this plug-in pull and then people I know if you just used it for encoding part of it or a we have we put a decoding layer I think it is just an encoder basically we let me remind you and then let me analyze why that n days or whatever gap you are seeing is happening yeah one more thing I want to include in this code that me but that I think would be really helpful is to incorporate other features by gamma – yes other features would be able to help us be better so if we can include it because most of the LSE encoder that I saw on the website they only inflicted on one value either on the closing value opening price so if we can include other features that would be able to that would be that would help us to be at the okay perfect let me check it out okay so I’ll do one thing Ashish first let me solve this problem because it looks like to be a complex and not say very complex code but yes when LS came in I’m sorry when Ellis teams are involved it is little complex so first I need to understand the design of the network first of all why it’s 50 60 70 what is it making this way and probably once I get hold of it I don’t think so it should be big trouble editing this okay good Merilee I will share this with you my code of lsdm predictions check it out okay yeah all right perfect Ashish one more thing I just a reminder there is one guy called Ganesh who is just joined our bastard he is he may he has what you say this is a little difficult way to do time series was tomorrow something changes now you have to play in your network and I don’t know how much time it took for you I think the epochs are not heavy you know it actually took me it took me around you one ought to train this code correct now this imagine tomorrow is that condemned back in content increases no the training time is also increase so first option what I would request you is if you are deploying this somewhere or if you are doing for your own our knowledge perfectly good this this is very good but if you are doing it just then is to deploy it eventually if it eventually so just check the cost of it cost in the sense because whenever you

have you I have to retrain your model no you have to retire your initial knowledge first then you to pull it back and then you have to retrain it and then push it back so that non operational phase you have to calculate the cost for so if you are ok with it no issues first option should be ARIMA second option should be electives to be very very well he said i certain I trained it or not daily data is that because I it would not involve too much cost in exhibition because I didn’t think the manual this is an office it particularly weather so my intention was actually to do it on a live data on a pick up buy ticket on a second-by-second data so that would be I’m not sure because if it takes someone to run the code if not impossible to do a trade on that kind of data on live data so that’s what I was thinking so II is the is the deployment is different from see what we are doing so is first we are trying to create a code a once we are notified with the course how the deployment takes place all the way you want to deploy this when you would like to acquire you and deploy it on my phone so using them when when okay I will show you right away it’s a way right yeah yeah yeah let me show ya let me show you right away this is not our L STM’s but yes this is one one minute we use flasks basically a fish to do this and I will give you this code also so you can deploy it immediately and this is my tap so the overall architecture if I show you just give me a minute so the overall architecture is that my intention is here to make HTML web which I have done it okay so this is my HTML and this is my CSS some styling I’ve done to create an interface which looks like this okay okay and this particular thing is very simple you will just put a URL here this URL being given is an input to me in my code so I use flask to convert whatever I am talking in Python with respect to what HTML and CSS understands very simple so I am taking a link from a user why a flask I am in putting it into my model which is pre-drilled on some kind of data it’s a neural net model okay and after that I am saying I am replying back to the HTML using slash saying that it is a real news or it’s a fake that’s what we are doing all right now how do we do it let me show you the code the code is very simple let us say if I can’t run this before that yeah so this is the flask app so there should be one application which you have to make it after you are confident with your model what you have to do it you should pick tell it if you see this this is called pickling pickling is like converging your model and storing all the supporting files on to a location which can be called from anywhere so this is my pickled package which is nothing but my name Bayes model that I’ve designed above okay any model that you design could be pickled like this so this is the the two liner code which you can use to pickle and once you run it you get a pickle file like this now your job is to import this pickle file in your reel so this is my real application and also parallely I have deployed a webpage so this is the link to my web and is the CSS which I have written on that basically some styler so whoever is handling your application will be having access to both of these blue very near yours website is hosted there you have to go and drop this pickle file basically okay okay along with this the file is a cool so this is what we call it as flask now what does flash do is flash allows me to create an app inside that app it allows me to load my model if I have one this is my model a technically remember next is it allows me to connect my HTML page link here if you have an IP or if you have any other thing you just put it across here next is what you want to do with it so if this is the HTML page what do you want to do it so my HTML page what I want to do is let me run it if it is not ok it’s not active so let me run the code and actually it’ll take some time it’s not that fast but yeah you just have to run it once once it generates an IP IP is ready look at this this is a fake news predictor all I need to do is I just

need to go and for example let me say I will pull some junk out of it I have to just get the link go to my web publish the link here and say predict it might take some time the news is real so this is how I am Stelling I am telling my app that first of all this the model that you are supposed to use from Python this is the HTML page and inside HTML page what you are supposed to do you are supposed to get my text this is called web scrapping if you have heard about the Scourge read scrapping is like converting the data on the web onto an understandable which could be passed to a Python model something like a data set input data set or database this is happening whenever really you want to do with your inside your HTML you to define so in your case what you have to do in your case you to publish a graph or do some numbers that logic should come here so no actually in my case what I want to do is a it’s a given order is for example a model is predicting to that the price would go up so it to send that order to the broker town that is to buy the share and he also send us a simultaneous order giving you a stop loss or a target price order so that a some execution can take this and then II when the model says yeah it’s time to get out of the market then it exited loader and it fits the market perfect so in that case as usual you will need a help of a web developer basically to write this interface okay if we can do it in Python as in Python should output then no problem we will say whatever is the Python output will display it here in this junk but if engage lot of interactions are needed internally no you need a web developer to take you through those interactions because I’m not show a little boudoir I am not that good at HTML or I’m not sure even you are but the person is designed the weapon should be good enough to put interface this brings us to the end of this tutorial on multi-layer perceptron now before you guys sign off I like to inform you guys that we have launched a completely free platform called as create Learning Academy but you have access to free courses such as a iCloud and dead still marketing you can check out the details in the description below so guys thank you very much for attending the session and have a great learning ahead you