sfspark.org: Will Ramey Lightning Talk
Recording: sfspark.org: Will Ramey Lightning Talk
so as he said I'm will Remy I run the developer programs or rather my teams at Nvidia run our developer programs in including something I'm pretty proud of which is our deep learning Institute which is an initiative through which we have trained over 10,000 people worldwide how to apply this new deep learning approach to artificial intelligence to solve lots of challenging questions next slide anyway so um how many of you are familiar with machine learning in general Lots awesome how many of you have a familiarity with deep learning about half two-thirds of you okay so you know that deep learning is just a relatively new approach to machine learning and the big difference is that uh whereas with earlier approaches to machine learning it required a lot of human e effort to form the hypothesis around what are the features that actually matter uh in your data set and to manually write feature extractors to characterize or to detect those features and then to hang them together into some kind of logical tree structure and then use some example data to tune the weights of those feature extractors or detectors uh to um make sense of input that the neural network hasn't seen before but that with deep learning all of that manual effort which can be tremendously timec consuming and at times error prone uh as well as not being generalizable to other problems is basically Al solved by using much more sophisticated neural networks deeper with more layers and also using huge huge data sets basically collections of experience to train those neural networks so that they automatically figure out what are the features of the characteristics that make a difference a lot of people ask us why deep learning is happening now and what is the role that gpus play in deep learning and so we use this um fairly busy diagram here to explain that the three initial conditions the three ingredients if you will that had to come together to make deep learning practical uh and possible in the ways that we use it today are first the Big Data if you don't have huge amounts of data to capture that experience then you don't have enough examples to teach the Deep neural network how to do the job you needed to do at a reasonable level of accuracy the second is that the algorithms the ideas uh that get captured in all of the different layers of the neural networks had to not only exist because they'd actually been around for a while but they had to be parallelized so that they could process all of this data in a reasonable amount of time and then the third component to that is the GPU parallel accelerator that runs those parallel algorithms to process the Big Data fast enough to make deep learning practical and in the early days of deep learning there were a small number of researchers who survived the previous AI winter that is when all the funding dried up um but we're lucky enough to have uh some some researchers in universities who were um uh able to take advantage of their academic freedom to continue exploring these areas and they figured out that hey now that gpus are programmable we can do cc++ uh Fortran python Etc with this massively paral accelerator let's try applying that to these deep neural networks and they not only published their research and papers but they captured their work in these software Frameworks these higher level abstractions and publish them into the open source community so that all of us could begin to experiment with them and expand and contribute to them and apply them to problems that the original researchers had never even considered and so many of us did and we found that it was useful for lots of different applications and cloud service providers said hey this could be a big business let's package these Frameworks up let's make them available in images let's have um you know webinars and examples and case studies that show people they can take advantage of our Cloud platforms to run deep learning workloads and when they did that rather suddenly a very large number of startups said wait a second I don't have to build my own data center I don't have to develop all of my own software from scratch these open source Frameworks already exist my barriers to entry in terms of cost and number of people and all sorts of other things have basically evaporated or at least very very low now I'm going to go start a business and so they did and thousands and thousands of startups showed up taking advantage of this deep learning and of course we're not really sure which comes first the funding or the startups they seem to feed on each other but billions of dollars of venture capital showed up to fund these startups and now we have this huge ecosystem of startups that are focused on applying deep learning Excel or I'm sorry deep learning uh artificial intelligence to solving all sorts of problems everything from agriculture to medicine to health care to zoology uh protein folding and excuse me everything in between and when you have that much energy going into the startup ecosystem the large companies the uh Fortune 500 the global 1,000 companies they sit up and take notice because those who are more aggressive more optimistic more optimistic more leading in their business strategies see it as a huge opportunity for them to improve their products for them to improve their business and uh for them to improve the ways in which they engage with their customers and for other large companies who are perhaps not leading but seeing that their competitors both large and small are taking advantage of this technology and beginning to reap the rewards they have to start investing as well from a defensive standpoint because if they don't they may no long they may no longer be able to grow or sustain their company so this is what we call the big bang of modern artificial intelligence and it helps to tell the story of why this technology is growing so quickly but next slide I like having my own clicker um invest in that next next time we'll have promise upgrade um but what we're seeing is that this uh deep learning technology is applicable across a very wide range of use cases in fact a very wide range of Industries I've listed just a few of them up here uh because they have lots of fun use cases so how many of you have used like uh Siri or um right Siri or Alexa or any kind of speech recognition right almost all of them certainly all the ones on your phone they record your voice they send the audio really quickly uh up to the cloud it uses some kind of GPU accelerated deep learning to understand do natural language processing do speech to text understand the intent of your text and then send it back to your to your local device um uh Facebook for example has when you upload your pictures right they um try to put the name of well your name all your friends and everyone else they can recognize in the picture and if you had to upload that picture and then wait a few minutes for them to send you an email or some kind of notice to say oh now we think we know who the face is that's not a very delightful user experience they need very very low latency response time in order to deliver a good user experience so they're using a um uh GPU accelerated backend with deep learning to to do that work as well um similarly in medicine examples from um you may not know this but um when when people have someone with diabetes um when that disease progresses one of the things that can happen is they can go blind there's a condition called diabetic retinopathy where their retina starts to Decay and there medicines there's a treatment that can prevent um that condition from progressing or at least slow it down significantly but it's mostly effective only at the very very early stages of the disease now optometrist optometrists have been taking pictures of our eyes for a very very long time there's huge amounts of data and example and um a really good training set around that and so they've you but previously they didn't have a good technology approach that would allow them to use it as a tool in the clinic right it was something you'd have to either send off or it was just too computationally intense and there wasn't a good a good solution now they're able at the the same time the patient is in the clinic take a picture of your eye put it through a deep neural network or an application that uses a deep neural network and with a very high degree of confidence the doctor is able to make a diagnosis there on the spot and prescribe the medicine while it can still be effective in Saving the patient's site so examples of you know low latency response time uh more examples from uh media entertainment uh security and defense autonomous machines I think was mentioned earlier is a big area where we're investing we we're building our own self-driving car we're also partnering with uh all of the major Automotive companies to help them take advantage of uh deep learning and GPU other GPU accelerated Technologies to build their next Generation self-driving cars next slide please so um if you want to talk about gpus and um deep learning in general I'm happy to to talk about questions but the one recommendation I have for you today is if you'd like to get your hands dirty if you'd like to get some hands-on experience working with deep learning across a wide range of the open source deep learning Frameworks please check out our deep learning Institute Training offering we offer uh free and very lowcost uh Hands-On training experiences in the cloud where you can just sign in you get access to a fully configured workstation running in the cloud has all the software you need all all already installed example data sets and exercises that you can walk through to learn things like uh object detection image classification image segmentation um um style transfer some some higher level things like that there are some uh Medical Imaging and uh patient health record analysis for people who are interested in in doing text processing for documents um and uh lot lots of other examples as well as I see a GPU technology conference shirt in the back there um we trained several thousand oh he's not paying attention sorry I called you out just because you're wearing one of my favorite GPU technology conference shirts there uh we trained several thousand people uh this year at the conference itself just last month um going through our deep learning training and got some really great feedback so I hope that you will take advantage of that as well thank [Applause] you oh