 Well, good afternoon all First of all welcome to Intel's AI launch My name is Bushan Desam. I am the senior business development manager at Lenovo's data center group My primary focus is in artificial intelligence business So, I mean there's a lot of excitement around artificial intelligence in recent years But at the same time there are several fears too especially when it comes to automating everything that we do today and also We're losing control on some other things that we're good at But on the but there are also a lot of social benefits that comes from artificial intelligence And I'm going to make a case for that the AI and computer vision for social good so If you take that picture, you know, what do you see there? A man and a passenger van. It's very simple. It's not quickly, you know So what is the context here? so a man talking on the phone is about to hit by a van while crossing the road and What would you do you run away to avoid the van and you just want to make sure that there's no traffic on the other side You know, that's the common thing that you do, you know instantaneously, you know without thinking much and What is e-power 91.2345 plus 25 power 2.2? This is a hard problem to solve for humans so We will ask the same questions to a computers that were Until until recent past, you know, like if you ask the same thing to computer if you say, what do you see? Doesn't have a clue. What is the context? No clue What would you advise? No clue But if you ask the same math problem is saying now we are talking this one is easy, you know 4.1941 times 10 power 39 so Who are intelligent here humans are computers the computers who could actually solve the hot math problems? Are the humans who could instantaneously analyze the situation? Reason it and take a quick decision. Who is intelligent? So let's just you know quickly look at what is intelligence first of all sorry the The slight thing, you know, so it's a very sensitive so the if you if you look into the Webster Webster dictionary of intelligence What it says is Apologies for this. It's not working. You know, so Yeah, okay Okay, so we are good. So if you if you look into a Webster dictionary of intelligence It says the ability to learn or understand or to deal with newer trying situations Reason also the skill use of reason. So that's what intelligence means So if you take this and apply between humans and computers So if you see humans are very good at sensing, you know, so you when we look at the picture We could immediately sense there is a man and you know, like is the car So and then we learn we learn from our experiences or we learn from others experiences You don't have to hit by a car to know that it's a dangerous situation You know that, you know, like so you somehow saw this you read somewhere somebody got killed, you know by when They were run by a car run over by a car So and then you reason it and then you acted no So you immediately run and then you adopt a situation that you don't want to run into another car But you want to run away from that so that you know you are safe So it's basically humans possess and you know, like exhibit intelligence, you know, that's what intelligence is Whereas until recent past what the computers are good at is basically doing complex math operations data storage and retrieval, you know, so it can Store terabytes to pedophiles of you know, like petabytes of data and it can retrieve it in milliseconds So that's what the computers are good at So there is a clear division of labor until now, you know So humans do some things computers do some things, you know, the things that computers could do we can't do You know, the things we could do computers can't do so that's actually a very nice division of labor between us and computers So what are computers good at, you know, so the computers are good at following rules created by humans so if you take in traditional way the way we interact with computers you actually Write a computer program that has some logic and what the logic does is it codifies the rules You know what rules you want the computer to follow and When you give those rules through a program the computer executes that program and it gives you, you know, whatever the desired output But the the real thing is what computers are good at is that very good at following rules created by humans at incredible speeds You know, like so that they can just execute So, you know millions and billions of operations, you know, like in the in the in a second So that's what computers were good at, you know, so that's just What what we've been used to But that's about to change. So this is a very popular visual recognition challenge in the computer vision community It's called image net. So image net is a large database of images. It has about 1.2 million images that belong to thousand different categories so the challenge is if you if the computer actually can Look at 100,000 samples and test what that images, you know, be able to tell what that images and at what accuracy So that is a challenge So if you look at the error rate the error rate in in doing these hundred thousand tests if you historically track that In 2010 the error rate from computers were about 28 percent. So that is the error rate, you know So they were getting, you know, 72, you know, like right, but the 28 wrong and then, you know, the error went down it keep going down until 2014 but something happened in 2015. So this is when Microsoft researchers could actually Bring the error rate Even below the human level slightly below the human level. So the human error rate is if I show you that thousand, you know 100,000 pictures you could get 95 percent right and 5 percent wrong But this is the first time the computers Did this at 4.9 percent. So the slightly exceeded the human capability and in 2016 it went further down so basically That in 2015 you see this whole thing turning around. So there is the press went wild Sorry, I'm the another person who's blaming press but They're just saying hey a is here is going to take over a job is going to change everything So that was the the main theme of the story But also keep in mind It's only doing one task. It's not doing everything isn't it? So all it is saying is hey, this is a picture of a cat. This is a picture of dog and that too In the in those thousand categories, so if you find something that's outside the category It says I don't know what it is Simply I don't know so it's very very narrow. You know keep in mind that you know, it's not solving every problem It's just taking one data set of images and kind of trying to find what this image belongs to So is this artificial intelligence really so let's let's look at some definitions of artificial intelligence now So we looked at intelligence, but no what is artificial intelligence in that so Artificial intelligence is basically the intelligence Exhibited by machines or software. So this is a Wikipedia definition very general definition and we already reviewed what intelligence is You know, so they built you to Samsung act reason, you know, like adopt and all of them are considered intelligence But another technical definition of artificial intelligence is the study and design of intelligent agents Where an intelligent agent is a system that perceives? You know, that's the key thing, you know, so it actually perceives the environment You know, it tries to look through it perceives the environment and take some action that maximize its chance of success, you know, so Perceiving and acting those are the two things So let's take this and try to apply to this image image net challenge. Is this really artificial intelligence? there's actually in fact there are two kinds of artificial intelligence one is narrow artificial intelligence The other is general artificial intelligence. There is a big gap between these two So this I think you know one had to realize that so the narrow AI is today is driven by industry, you know Like both large scale, you know, like large industries and also there's a lot of start-up working on it What it does is it tries to address one specific, you know, like area such as either it recognizes images Recognizes voice or it tries to, you know, like, you know, translate the language and run business analytics So that's that's the kind of you know, like typical use cases for the narrow AI So today it's already deployed by hyperscale companies like Facebook, Google's, you know, like Microsoft So when you upload an image, you know, Facebook image to Facebook, I mean An image to Facebook it suggests some tagging That's how it is doing, you know, it's actually doing image recognition, you know underneath And but now it's actually finding very good, you know Applications in medical diagnosis, education, finance and manufacturing. So I'm going to review some of those applications And it's poised to create some good societal and economic benefits. So that's the narrow AI scope is that's what actually happening today, you know, like In overall but there is this general AI it's kind of mainly driven by academia and scientists, you know So what they want to do is they want to they want computers or you know artificial intelligence to come to a level where It really exhibits the intelligent behavior that we exhibit, you know, taking multiple things, you know Multiple things together like in a sensing, acting, learning, reasoning and all of them and how we react to a situation We want the AI system to react to a situation But this is where, you know, if you track the history of AI, it went to booms and busts, you know, like twice Probably perhaps, you know, we are not there yet and they there's Still, you know, like little progress and it needs decades and decades of research in this area and and but it has Important and complex implications to society, but we are not there at least, you know, couple of decades But even if you get there, there are always Technology always poses several challenges is not so if you if you go back Very very early days is not so the steam engine actually revolutionized the the way, you know, like the muscle power is Replaced by, you know, like mechanical power So that has huge implications in productivity in the society But at the same time the the steam engine was actually powered by coal Coal was actually putting carbon dioxide into atmosphere and it's believed to, you know, cause this global warming But then if you compare to let's say 20 years ago from in a coal power plant versus today It's only putting 1% bad stuff out. So 99% is clean That is again driven by policy and the policy, you know, like in fact draw Environmental technologies to clean the bad stuff that's coming from the coal power plant So it's it's going to be it's going to be an evolving field There will be lots of technological progress, but at the same time there's going to be policy changes And they all have to come together really to deliver the right benefit either. It's the the general AI or the narrow AI so Let's come back to the image net challenges, not so this is the first time computers beat us In recognizing images, but what made it possible? So there are three key factors in there. One is the computing power itself Today's computers are very fast There's tons of data that's available today compared to, you know, like five years ago or ten years ago And there several algorithmic advances to, you know, so those are all three of them They came together to, you know, like make that kind of progress. So let's review each of them First of all, I mean we are coming from Lenovo, so we want to talk about computing power, and that's so we always say you need more And if you really what's showing on the on the right-hand side here is On my right-hand side is the top 500 supercomputers in the world so the the curve on the Middle one here is actually showing how individual Computers, you know, like the number one is actually growing at the capacity and the one in the bottom one is the 500th computer, what is the power of the 500th computer? And the top one is if you sum up all the computing power from top one to top 500 computer if you sum up all of them That's what the capacity is but keep in mind that keep in mind that this is a logarithmic scale It's not a linear scale, isn't it? So do not take this Advances lightly the reason I'm saying is If you take today's CPU if you take just one Intel CPU that one Intel CPU can almost deliver One teraflops per second so it can do one tera, you know, like flops of you know work in in one second and Interestingly if you track back 25 years ago If you take all top 500 supercomputers, I'm talking about supercomputers not the you know like Consumer computers in the world if you combine all of them together. That was only delivering one teraflop and If you take the same one teraflop 20 years ago the fastest supercomputer on earth was delivering what your one CPU today is delivering the astonishing progress in technology and And see where we are at today. We are today the fastest supercomputer in the world is actually in China So it's in a national supercomputing center in Wuxi, China that delivers about 100 teraflops of computing power per second and us Is actually building the next-gen supercomputers called summit and Aurora so they are going to be in that kind of you know hundred teraflops per second of computing power and The the one here is actually the Marconi supercomputer that Lenovo and Intel built together for the Italian Research institutions, it's about 11 teraflops per second this machine So this is in Bologna, Italy, and this is the world's Number 12, you know, so if you look in the 500 list, it's the world's top, you know 12th computer Here so if you can see just to sum up the computing power It's it's going up and up but at the same time the the prices are also going down You know if you for the same dollar you can actually afford more and more computing power So the next one the advance is the the computational side or algorithmic side of you know So there are these dupe neural networks So they the dupe near the the neural networks are actually Inspired by human brain so what what I mean is you can see that there are multiple of these You know like circles are you know like all these squares you think of those are yes You were neurons, you know so this is not exactly doesn't act like a neuron but you know it mimics a neuron and When you feed something to the neural network, you know, so you take one some image So the first layer is going to keep the very you know like low-level features You know it try try to you know like immediately detect your edges And then you know like it tries to put some features in there and then you know It try to capture our face and by actually creating this Abstracted layers it can actually keep a lot of information in there so What happens typically in a in a you know like when you develop an application is you build a neural network and You keep training with the data. So the best example, you know like everybody says in deep neural network is the cat It's not identifying a cat, you know how it came nobody knows It's like hello world, you know like why do you first the first program is a hello world. You don't know why is not so So but imagine you are feeding a million cat images here So when you keep actually throwing that that image there then it kind of creates an abstraction saying oh This is how a cat looks like this is exactly how we learn when we grow up So the first time you know like you we get this nice, you know like book from our parents saying hey This is a cat they don't say a cat has you know like two years, you know like it has four legs and nobody tells us So we kind of start creating an abstraction and and we learn by examples You know like after a hundred time we say hey, this is a cat and then we slowly you know We start learning you know like we learn the context, you know like and then there goes on So the key part here is learning and also learning by example So that's the key part. So that's how the deep neural neural networks actually work and Oops, I think I messed it up. Can we go back a few slides go to the big data slide? Yeah, there you go. Yeah Thank you so so the third the We talked about computing power. We talked about algorithmic advances. So the next one is The the big data explosion. So there is a lot of data today So the day typically when you talk about big data, there are three elements the volume of the data The variety of the data and the velocity of the data. Look at these numbers 44 trillion gigabytes of data by 2020 three point thirty one point twenty five million messages 7200 hours of video hundreds of sensors in cars 500 million tweets six billion devices these are like huge numbers and That's the volume when it comes to variety you have now images you have video you have you know like text You know it's so much of the variety in the data and also the velocity when these hundred sensors are pumping up It's pumping up very fast and also look at All of these numbers are in minutes in hours in days, you know, that's a huge amount of data coming So the pace of the data is also very high. So so the real, you know like Challenges, how do you actually make sense of this data? But at the same time there is an opportunity because these deep neural networks they get better with the data So it's really a challenge, you know, like this one challenge is actually creating an opportunity. I mean, that's how nature is isn't that so Okay, so the I briefly mentioned in the beginning that There are two common fears when we talk about yeah, isn't that one is a replacing Humans by automating tasks that we handle well today and also Yeah, exceeding human capabilities and taking over the world. So those are the two common, you know, like fears So this is a very good report. You know, like if you guys are interested in that, you know, I highly recommend to read this You know, it's a very easy to read report. So this came out in October of 2016 this was published by a committee of from National Science and Technology Council They advise the US president on what is coming up in the future how it has implications, you know, like for our society Whether it's you know, like employment or you know, like environment, you know, what not, you know So it's a typical body that advises the president So basically the what the report notes is that the best way to attack some of the jobs in the future Is actually to pay humans with machines in situations where the human partner cannot compensate for weakness in the computer or vice versa So let's see if we can actually work together, you know, like with the humans with artificially intelligence-powered computers to to kind of, you know, like better the outcome Would this work? So let me Work to an example So the the dilemma here is, you know, is that humans are AI, you know, so which one to pick so this is a research done by Some of the MIT scientists and also Harvard Medical School folks and what basically Is they're trying to address a challenge. So the challenge is the detection of breast cancer in from the images of Lymph node biopsies, you know, so they take these, you know, biopsies, you know they do the imaging and then you want to see if there's a tumor or not and When they did that with artificial intelligence to just classify whether there is a tumor or not It's a classification problem. The computers were doing about ninety two five point ninety two point five percent accurate So that's the result from the computer But when a pathologist actually looks at the same image, he was actually giving a ninety six point six percent accuracy on that And then on the next category is the localization accuracy That means is you take the image you need to identify in which part of the image is the tumor in so you need to actually You know identify the tumor So again a is doing about seventy point five percent Whereas the pathologist is doing about seventy three point three percent. They're getting close You know a is getting close, you know to the to the pathologist level But even the most interesting part is when the pathologist was assisted by AI The whole accuracy improved to ninety nine point five percent. That's a damn good accuracy I mean that's actually eighty five percent reduction in human error rate So How good is ninety six point six percent in some cases? It's not you don't want to be diagnosed You know you could be that three point five percent, you know like you may fall into there You may be wrong diagnosed wrongly treated that has a lot of implications on human lives But but look at that accuracy ninety nine point five percent. That's that's very good accuracy so I mean we could work together, you know, so you can actually bring the humans and Artificial intelligence to basically say that you can augment your intelligence by using artificial intelligence So that's the that's the message. So it makes sense. It makes, you know complete sense So we talked about healthcare how we can improve healthcare. So let's look at a social issue I'm not sure this is the right place to talk about it, you know because we are in a bar, okay but this is coming from National in shoot on alcohol abuse and alcoholism and What it is saying is underage drinking is a serious problem in the United States It's actually not just in the United States is throughout the world, you know, like even in developing countries, you know, we see this I'm originally from India, you know, like I see this problem, too, you know, it's not just one country's problem and if you see the data that was actually shown By them it's saying the eighth graders are consuming ninety nine points on percent of them are consuming alcohol At some point twenty one point five percent in tenth grade and thirty five point three percent in twelve three So they don't drink frequently But once when they drink the drink in excess and and they have, you know, they pose serious health and safety risks So this is a this is a serious problem So let's see if we can apply the artificial intelligence to identify or solve this problem So I don't know like if you if any of you have a chance to check out that the demo that Lenovo is showing there It's basically a demo to identify your age So keep in mind that, you know, that's a work in progress. So we only fed 20,000 images to that, you know, so it's not complete demo yet, but we are just showing the early potential for that So basically what it is that it can actually estimate Your age based on the features, you know, like look at, you know, like how is your face is, you know, like, you know, based on that it kind of guesses your age and And and basically you can use that to identify, you know Persons like that are under 21 years in a in a major video. So that's one example of, you know, like identifying that So the second one is Intel is actually demonstrating a demo right there So what this does is it looks at a scene it localizes things and then immediately it identifies people it identifies objects like bottles So again, you can actually train the system to identify the alcoholic beverage, you know, like containers, you know Like the the way beer comes, you know, like they're only a finite number of shapes So you can easily use them to identify those kinds of uses and then there is a startup That's an MIT startup in Boston. It's called Netra. What Netra does is it uses computer vision to identify Brands and pictures and videos. So when you show this picture, it's actually picking up that Ray-Ban And saying it's 100% accurate. That's a Ray-Ban, you know, like so, so imagine that so you can actually again Train this to identify popular brands of alcohol, you know, whether it's a Stella or in a Budweiser and all like But like so that's how you can actually identify those brands So imagine you can actually put all of these three these three together. So age identification Container detection and brand recognition and if you build a tool, this can actually help some of the enforcement You know, like agencies, especially public places, you know, in large crowds and large public places to identify underage drinking So it's just an example, you know, like how how, you know, all of these technologies can be, you know, put together to solve some of the problems so Going back now, you know, like we look at a health care We look at, you know, how it can address a social challenge But let's also look at how it can deliver the overall benefit to the society, you know, like at large So this is where we need to review the history, you know So if you review the history and if you look at really track the GDP per capita And it stayed almost the same, you know, like until, you know, like 17, you know late 1717 hundreds But then it just ticked off. It took the hot stick, you know, how to stick pattern So it is actually the steam engine that was invented in 1775 was actually credited for this Industrial revolution that gave us, you know, like massive productivity and human prosperity by basically replacing the muscle power With energy from natural resources. So this is where, I mean, you you can basically have as much energy as you want wherever you want Rather than imagine, you know, like you need to take thousand people to do some project somewhere Rather than, you know, like you take a mission there and like do it, you know So that that power actually gave a lot of ways to improve our lives and and also it Helped the the economy. I'm not saying it doesn't have, you know, bad implications again at the time There's a lot of people like who are doing human labor work and they were displaced But then they kind of, you know, transformed and found work somewhere else So a lot of muscle work now we actually move to knowledge work and the next generation could be, you know Something else knowledge work powered by AI, you know, so that could be a trend, you know, like so going into the future so The the key thing there is it's actually the the the steam engine took the natural resources and actually deliver the the value So today We, you know humans, we created a big resource here, you know, like in the digital age It's called the big data. So that's the one that I was talking about So you are creating lots of lots of data. This is a man-made resource. So is there a value in that? So if you see the the the data, you know, I can and how you can actually extract value you can put into four different buckets So there are descriptive analytics so you can take all of the data or you can take a sub-sample of the data You can see hey, how did something happen? You know, like what happened, you know So you can always see it. This is what typically you do in monitoring Isn't that so you keep in all like monitoring something and say hey, you know, like it's breaking something and Then there is a diagnostic analytics where you take the data and you try to create a hypothesis saying Why did it happen? You know, so you're trying to come up with you know, like ways to find out why why it happened But actually compared to that there are other analytics that can even provide more value But they are very difficult, you know, so it's very difficult to do that kind of you know analysis Those two are predictive analytics and prescriptly analytics. So in predictive analytics You want to predict, you know ahead of the data. So you don't have the data But you take the what what you have today, but you want to be able to predict what's going to happen tomorrow And that's the predictive analytics and the and the prescriptive analytics is it should be actually sorry There should be a prescriptive analytics What prescriptive analytics is you want to actually change the outcome So you want to do some kind of scenario analysis and see how can I force that scenario if I know what the outcome is So this is just you know, like analyzing, you know, this is able to predict something But then this is a you know able to actually predict the input conditions that can you know force an output or you know Like an outcome so that's where there is a high value and that's where the high productivity is and Ironically we're collecting a lot of data and 90% of the data is is never analyzed Because there are no good methods to do this kind of high value analytics in in every scenario So Can I get us there, you know, like so that's the fundamental question So if you if you really see The two axes that I'm plotting is the volume velocity and a variety of the data So there's a lot of data coming in and you want to somehow create value from that, you know So that's what I'm showing on the y-axis You could do a lot of traditional programming but the traditional programming, you know, for example, especially if you take unstructured data if you want to Analyze images if you want to analyze video, you know, like if you want to analyze all of this It's very difficult to program everything in a traditional programming manner But that traditional programming is mainly driven by humans, you know, so this is where humans kids are adequate So to some extent we are good, you know, so that's how we've been dealing with computers is not so Primarily to deal with you know, like numeric data, you know, like tons of tons of numeric data And especially if you know what to program if you can actually codify the logic Into rules this really works. But after then, you know, like you saturate, you know, like you cannot deliver more value with that This is when then you move to a classical machine learning, you know Machine learning is a kind of it's a broader umbrella. That's where the deep learning fits But still in machine learning it still needs some human input because you have to define the high-level features and ask the Machine to learn low-level features in order to, you know, like make the estimation So but again the classical machine learning it doesn't scale very well You know at some point, you know, it also saturates and then this the deep learning that is based on the neural networks The the very deep neural network. That's when they can come and actually, you know, like help us with creating, you know High value so they can analyze video. They can analyze image. They can analyze text, you know, what not and also Look at here. The human skills are not adequate if I give you a pet of one pet of love of video. Can we analyze it? You cannot I mean we still we cannot I mean analyze that kind of you know Like data that amounts and also We cannot look at patterns, you know, especially as the data is going, you know, like bigger and bigger you got to You know look into 20 variables 30 variables. So you cannot actually extract patterns there. So let AI handle this So this is where the AI can help us in the digital age to get higher productivity like how steam engine did this During the industrial, you know revolution So that this is really in need of AI to extract high value from from data So I just want to give you one quick example, you know, like how Lenovo is using machine learning You know to to extract value So we we get lots of this social media push, you know, like so somebody talks about our products Saying hey, you know, this is good. This is bad. You know, I like this. I didn't like this There are lots of images, you know, like videos and all I can and those kinds of Data, but they're very difficult to analyze it. So this is where we use it actually machine learning to understand what customers are Talking about our product took that feedback and actually build this yoga product, which is very successful The Lenovo yoga brand is very successful. So that's what we created so this is again where Humans cannot actually take these hundreds and hundreds of hours of video or you know, thousands of these images and can synthesize And come up with some kind of you know, like, you know, very simple things that you know We would be able to understand, you know, that's what the computers and the AI is doing and helping us with So another thing is yeah is going to impact every industry nearly every industry You know, so I'm just highlighting some examples here So the investment industry the healthcare and all like as I gave you an example New automation self-driving cars, you know, like marketing Oil and gas manufacturing security or defense media. So there are lots of lots of you know, like Industries that can get impacted by by and and again There's always opportunities to leverage the artificial intelligence in the right way in all of these So I'm just you know, like kind of you know, switching the topic and say how do you build an AI system? If you want to really build an AI system, isn't that so you you have to store a lot of data You know, the data might be coming from somewhere, you know, like as a social media or you know some from your customers in enterprises So you collect all of that all of the data in a big data system This is where you store you prepare the data and you feed the data and then this is where the neural network actually gets trained You know, so this is where you need a lot of computing power. That's why I was actually talking about the importance of the computing power So it's even common to see a lot of these Supercomputing centers are now moving towards AI direction that they voting towards AI because they know how to do this problem You know, like this high-performance computing kind of problem Once you train a neural network, then you want to deploy the neural network So this is where you can deploy it in mobile phones or you know, like smart, you know, like Devices or you know, like your PCs or you know, like laptops, you know So that's where you deploy this and then so for each of them You need like, you know different kinds of you know, like technologies So for example, the big data systems are typically run by Intel Xeon processor and Intel is now planning to release this Xeon 5 which is ideal for machine learning and Also the alter of the FPG S and Intel Xeon for inference So that's how you can actually build a system by taking these technologies that Intel is developing and and we do actually provide those systems to You know, like AI customers or AI users So our mission at Lenovo is to democratize AI. So how we are doing we are actually taking this Three-pronged strategy. So we are building an innovation center. The purpose of the innovation center is to demonstrate what AI could do, you know, like to you to you know, like so this is where we want to Take and you know, like this image diagnosis kind of thing You know, so we want to actually create some of those applications to demonstrate the the value of AI And we also want to create a development platform where our customers can immediately take Take it and immediately start building an application Rather than worrying about what kind of hardware I need to bring what kind of software I need to bring and all other stuff and then finally the Deployment so this is where the we are kind of planning to deliver Scalable systems because in enterprise context if you are to deal with you know, catapflops of data and you know Like you want an entry and solution. So this is where we want to deliver an entry and solution You know system. So that's what the the the What Leno is actually doing to democratize AI, but thank you all. Thank you for your attention Shate of time any questions So this is this is one I came across but definitely, you know, there are others other areas One is again, they're trying to they're all around this image image-based ones because if you see a lot of AI today It's all around image based AI. So they did this in healthcare They're doing this in oil and gas. So in oil and gas when they are doing this seismic processing They collect a lot of data, you know, like so they send this sonar signals and then they do the imaging is not so again This is where there is somebody who is actually looking at these images and try to estimate the oil Reserves there what says AI is doing that and there is by looking at these kinds of examples. They're looking at you know I can these be Together is one. So the second one is again in healthcare mass general hospitals is doing a lot of work You know, so they have billions of images in their repositories again They are actually training all these radiology images with AI Systems and the whole idea is before even a radiologist come and see that It the AI is able to predict something and then it's going to help the radiologist So the radiologist doesn't have to spend 15 minutes there. So you can actually spend only five minutes to verify the diagnosis and also some of the information that's again giving is actually helping, you know So it's a very similar to what you know, like the the other example that I quoted But it's kind of but they're trying to bring that into production You know, so the the example that I gave is more kind of research It's a research paper from MIT and Harvard But actually what mass general is doing is trying to put this into production to bring it to patient care But oil and gas is looking at this Perhaps, you know, there are other examples, but there's something I could come up other than sorry So Imaging is one but I see a lot of analytics, you know, so in a lot of big big data analytics, isn't that? So today a lot of those analytics are based on statistical packages So this is where again they're moving not directly to deep learning but actually machine learning at least So the next steps seem to be machine learning So so imagine before you want to You want to give a loan to somebody, you know, like so the three or four parameters The financial companies look to underwrite your loan is, you know, basically your salary or age, you know, like and all of them But now imagine if you want to look at 20 variables to get better accuracy Okay, this statistical page, you know package, they cannot correlate among 20 variables So this is where you got to do this predictive analytics using machine learning So what machine learning does is it now it takes us all these 20 variables and try to correlate, you know Like basically that's what it is doing and underneath So that means you are putting fine granularity into your data So Enterprises are very interested in doing that kind of work, you know, how can I apply The the machine learning as a next step taking from the traditional statistical techniques And the second the second Trend is what I'm seeing is You have typically the structured data that was living in databases And then now we have this image data That's what a lot of companies are doing is they're trying to bring them together So they're aggregating it, you know, so for example, you as a customer, you know, so You are in my database, you know, I'm an enterprise, you know, like I sold you something but again You may be posting something in the social, you know, like media The one example I gave you is like there's no linkage between, you know, like the customer and the on the data We're just looking at the data in it But now I want to actually really understand your preferences So I want to treat you as a segment of one basically I want to see, you know, what you want to buy, you know, what you know, what your preferences are So this is where it's actually trying to bring these two together, you know It's not just looking at the image data by itself, you know So those are some of the emerging trends That's where enterprise are finding value Yeah, of course So the two areas that we see in higher education is in two different areas One is the academia is trying to push the neural network research itself, you know So they want to actually do better algorithms So out of those three, you know, like if you talk about computing power data And algorithmic advances, they're making a lot of algorithmic advances So that's one area The second area is a lot of enterprises, they cannot adapt to some of the methods Because they don't know, you know, how to do this This is where, again, the academia, you know, like is taking And trying to kind of show some way of doing it, isn't it? So they want to show, hey, this is possible And then the next step is, you know, like for the enterprises to pick it up And see, you know, like how they can use the application So those are the two areas, you know, algorithmic advances And application advances So that's what the academia is doing Any questions? Okay So, well, thank you very much, you know, for listening to my talk And appreciate the time I hope, you know, like you enjoy the rest of the day Thank you