scala.bythebay.io: Alex Ermolaev, Deep Learning Around Us
Recording: scala.bythebay.io: Alex Ermolaev, Deep Learning Around Us
I you [Applause] hello okay thank you very much for coming so I work for in VT and I work with a lot of software vendors who deploy or develop deep learning solutions for different applications so deep learning is a new technology which evolved in the last few years only in a IT industry we have wave of innovation we had mainframes and the 70s 10 minutes and the 1980s and PCs and the 90s then internet and the latest evolution we have is ai ai evolution and the plane is a big part of it so this revolution started only a few years ago there was a paper in 2010 by Alex krasovsky aliases Katrina and Geoffrey Hinton which demonstrated how to achieve beat bet image recognition using deep learnings and traditional machine learning methods and this particular paper and this particular results just blown everybody else out of the water and the reason is all all of the work prior to that particular day was done using traditional machine learning methods where people look images they construct feature vectors they do tons of work on future engineering literally people spend the entire careers 20 30 years doing finchy engineers trying to recognize the image and then this particular approach there was no future engineering at all so the model itself was able to recognize what's in the picture and you never told the model what to look for so that was the first results and they developed better results than everybody else had in the industry and since then the entire computer vision industry switch to deep learning and nobody else does anything else because everything else is a superior approach so all of it of course is run on GPU so since 2012 we had several developments first Andrew Aang and his group develop a way to develop scalable solutions based on GPU and they did it in 2012 so this paper is actually 2015 you can look it up and see how people were able to distribute the work load and deliver solve complex models on on a GPU in a class WBU in 2015 so the slides in the middle is very important one so when you do feature engineering and you do image recognition at some point you run you stop improving significantly you know you kind of like you try to capture every particular H case and you know etc etc at some point your model just stops improving in deep learning there is no that kind of barrier so in deployment model so for example in a traditional machine learning model let's say you train the model to recognize computer face oh sorry human face so but then picture changes let's say you don't have twice you know if I picture like this as a human is still see human face but machine algorithm stop working machine journal getting stopped working because one of the ice is missing from detection in machine learning approach there is no limitations like that so all you need to do is give the computer more images more processing power deeper network what jeep means multi-layered add more layers and it will continue to improve so what you see with deep learning results is with machine learning results you kind of leigh richey limit with deployment you continue to improve and improve and improve so right now deep learning are able to understand images better than humans don't so about the year ago the computer models that labeling the images better than average humans and and there are slide looks at the speech recognition and the speech recognition is the same at some point as you improve the model the computer models become better and understanding human beach as that human stuff okay so as a results of you know this type of achievements we have a tons of deploying implementations everywhere the first chart is deep learning models at Google Google made them you know my safe investments in this area and now right now they have more than 2,000 different projects that at Google to apply deployment to particular things apply it to like everything that moves right in the middle it's deploying frameworks most of them being open source you can go and download it from from github so those are the main frameworks that people use right now but if you would like to know how many frameworks out there let us counts we have us like 67 so the innovation continues people create their own approaches they write down frameworks and it continued to grow and the last slide is number of companies involved engage with and vga on different deep learning projects and 2013 we had 100 and it was a big achievement now last year we had 3,500 so that's kind of the the trends that we see so this is again talking a little bit more about why deep learning works so if you have you know if you keep aging my data you have a bigger which means deeper model which allows you to do more nuanced decision-making and more computation you get better results so in this particular case Microsoft went from eight layers for image recognition to 152 layers so more nuanced models and you know they that additional GPU capacity they increase the performance from sixteen percent every 2-3 and a half right and the second example so it's a 16 times bigger model gives you several times better results and the speech recognition example from by dou on the right is very similar so it's a ledger data sample they went from 7,000 hours to 12 and the larger GPU cluster and they were able to get better results in terms of the error in speech recognition and so how does this work so in zip line you've all there are two stages there is training and there is inference when you train the model it's very computationally intensive that you know task you have a GPU cluster and you react alkylate those parameters over and over again you have millions of neurons and recalculate them thousands of times until the merchants and the model converge to the to the good results and we see applications everywhere so one of the application is the satellite images so in this particular case this is an ad nasa project called jeep jeep something if I good so what they did is they look at the satellite images Enrique and try to train the system to recognize carbon impact so if you have around big cities or whatever you can actually see where the car bonds are cumulated other people use it the same technology for other purposes for example I know one of the Wall Street companies they look at the number of cars at the parking lot and if the number of cars in the parking lot of the store declines in this Elvis talk and if it's groans and the baddest so you can you can do something very practical as it as well the middle example is from a Pinterest so what they do is a look at a similar pins and if you have a similar pins and you can recommend something to purchase so you can you know suggest people would look at right very much a similarity search and the last example is a company which used to palani on tractors to recognize what needs to be sprayed so right now when the tractor goes around and sprays you know pesticide on lettuce it sprays everywhere when it's needed not so if you can actually see where the weeds are apparently you can reduce the pesticides by ninety percent so you save ninety percent of the cost on mr. size plus the consumer get ten times healthier letters because it's not sprayed with crap you know twenty-four seven the one of the big example for deeper learning is health care so health care is one of the industry which you wouldn't expect to be early adopters but you know in case of deployments it's taking off the first example is months in a hospital so what they do is they look at the electronic health records and this particular sample was 700,000 patients and if you look at the electronic health records apparently you can see when the particular Jesus approaching rights like you know if you have if your test results start changing over time and so you can predict when you're going to have a chronic conditions couple years from now and it's you know it's just math right you can predict those things but in a current health care system when you have you know dr. only see you occasionally and she doesn't have a time to look at your history you know the state is not being utilized with diplo any model you can look through this 700,000 patients with you know tons of different results and predict where the problems are the second is out this but pathology so just looking at the images and look what the problem is I think this was a prostate cancer from us not a naval hospital so this is image problem so the computer can look at the images and see whether there is a problem there or not and the last one is the example from Massachusetts on old for a geology what they did is as a look at the images of bones and see the child a child has a problem with growth development etc so looking at the bone density so we talked about training so when you train the model you need to have a lot of computational capacity and then there is inference so once the model is trained to recognize particular thing you train you know it can be used applied to predict or to recognize the same problem when it's visible again it many cases is done in data center in this particular example our East is a company which applies artistic style to your own photo so you upload your photo or video and we'll change it to make it more interesting so this type of processing is done at the data center so we have a lot of companies using deep learning obviously a lot of companies you already know have deep learning used for at least one application and of course in many typical companies you might have dozens or hundreds potential applications as a Google example indicated so we also have a lot of startups the last count we have about 1500 startups in our early adopter program Colin sumption and for example deep instinct is a company which is using deeply onion for malware detection so that's a standard antivirus stuff that you guys all familiar with so deep learning is better at detecting malware because it's more able to capture mutations the MOBA keeps mutating and if it's exactly the same value can capture it but it changes few lines of code suddenly looks different you might miss it so deep learning is better at capturing mutation in addition it is able to capture its able to capture the threads which are not known yet so and there are different ways to do it so deep learning can be applied to the looking at actual bits you can look at the behavior of the code or you can look at the metadata it's like weights came from what that's called you know the file size all those kind of things so there are three different particular models that can be applied and so the mall where might become a solvable problem so you don't get viruses on your PC once all this work is completed so deep instinct and startup which start working on this problem a few years ago we now have quite a few companies trying to do the same now it's a significant amount of work so when you start working with deep longing you just do the first implementation you get you know okay results maybe sixty seventy percent detection and then you start you know developing your model adding additional intelligence and getting Mon your hands and you can get ninety-eight ninety-nine percent detection of the known Red's of course detecting unknown threats is even more interesting so the genomics genomics is a great example for deploying applications because you have massive amount of data the data is very simple in nature but the amount of data is so huge it's very difficult to process it so the question is how do you find patterns and the data and predict particular thing so it's used in many different meats are many different genomics companies who use deployment to predict you know outcome of trials visualization medicine is coming at some point you know who should be treated with what particular disease and etc for self-driving cars so it's a big business for us so if you listen to what Fords and GM saying they're like in three years every car is going to be a self-driving car which is kind of difficult to believe but all of them I making massive investments so the biggest problem was self cars perception problem which is the problem we're solving so the kind needs to be able to look at the streets and within milliseconds recognize if there is a tree in front of it or not right should I stop I should continue going so it's a mostly perception problem so all of the working with every single large car manufacturer to prevent those systems and the systems will probably be implemented in stages because there are five different ways five different degrees to which you allow computer to drive for you but you know eventually within a generation or something if you can think about the next generation will not know how to drive at all because the technology will be able to do it better than the humans can and the last example is a prisma it's a another example of a company which used the planning to apply artistic style to your own pictures and it's a company which went public in the u.s. about three months ago in its and it was exploding like crazy so it's another example of tipping application so we talked about training we talked about inference done in data center and then finally device so the Kai's device for us so Kai's diva is a device and on the device you need to be able to quickly make a decision whether go stop left right whatever it needs to be done so you need to collect all the information from sensors quickly make Randy planning continuously run the planning inference and make decisions whether to go forward or not we also have other types of devices so people building deep learning into the consumer cameras so you actually can understand things like simple example like if you're in a company like Twitter you don't want people you know when somebody swiped the card you want that person to come in but not other people come in with him so like recognize who is coming in or maybe you don't need to swipe anything the camera will recognize who you are and let you in and not allow other people in so this has a lot of application and basically recognizing people faces of using cameras you see a lot of devices from you know Google Alexa from amazon all those devices use deep learning to help people manage their homes drones a lot of drones are building small version of GPU into the device in order to make intelligent drones so the ground doesn't hit every tree on only hits doesn't hit any hopefully the same as a little robot at the end which delivers groceries to people so it's a small robot but needs to know where to go when not to go needs to be able to navigate the streets and not hit obstacles people or cars right so it mean still needs to have an intelligence capability so besides deeply on you also have several [Music] analytics applications so the analytics space keeps evolving traditionally we have a lot of business process companies and you think about traditional data management then web digital and now we have a massive amount of information which is it's IOT or video cetera so traditionally the stack look like this and with a deep learning we have a with deep longing and gpus we have a lot of companies which are able to take the power of GPU and apply to solve analytics problems so in this particular example we are showing mumsy's one of our database vendors just a relational database the same where you access it using jdbc odbc but unlike the traditional damage that you'll see from an average vendor the normal vendor will show three days of data because Maddie runs on GPUs a show two years worth of data so that's kind of the two orders of magnitude improvement in the amount of processing capacity you have so we have a partnership a spark two weeks ago spark announced availability of GPUs and spark costumes so if you go to spark today I know you guys have skul people you can you know click the button and on the GPU and start running ug planning jobs within spark on data bricks so we have tons of applications which are growing rapidly we have a IC service provided by several companies so if you I don't know if you're familiar with Microsoft Azure machine learning google doing something similar IBM cloud amazone AI tons of companies offering a machinery needed service there are tons of company using GPU offering GPU in the cloud if you have your own problems that you want to solve so all of the big vendors as your amazon have GPUs classes available now within the last humans and finally we have analytic capabilities to solve the problem which will not solve before so that's the end of my presentation and I think we have a couple of minutes for questions one minute okay so the question was what are the evil and dangerous applications so if you there a certain there is one area of deep learning which called adversarial networks so if you are in adversarial situation for example if it's a cyber or fraud so people attacking you becoming more sophisticated and have a more sophisticated technology is probably using the planning and you want to defend yourself so you have to develop more sophisticated capabilities also using the planning otherwise you're going to be left behind so in adversarial situations like a fraud or cyber you generally need to have you know capabilities to defend yourself but in general I don't believe in singularity so so I think all of this stuff that we're showing it's just math right it's not like conscious systems let's try to take over the world it's just math right so like you have a complex problem and you have pattern recognition technology which can solve those problems and suddenly you know the machine can see you know whether to go or not so that's a big improvement but it's not it's nothing else it's beyond that thank you very much [Applause] you