Innovation Heroes

TRANSCRIPT - Deep Learning and AI Demystified with NVIDIA’s Will Ramey

July 2, 2021

 

Peter 

This episode is brought to you by NVIDIA and its groundbreaking RTX Technology. Visit shi.com/nvidia today to give your team the fastest, most powerful computing tools on the market.

[music plays]

 

Will 

In the movies, they also cast AI as having really flexible problem solving and cognitive abilities, but with superhuman intelligence. That kind of thing makes for good entertainment, but that kind of general AI is just not possible using today's technologies.

 

Peter 

Welcome to SHI's Innovation Heroes, a podcast exploring the people and businesses giving us hope in our drastically disrupted world. I'm your host, Peter Bean.

[upbeat jazz music plays]

 

Peter 

The Hollywood vision we're fed of AI is that sometime in the near future, it'll swoop in and save the day, with computers becoming so smart they'll be able to save us from ourselves. [mechanical whirring] [kids cheering] But in reality, AI and deep learning is already making a big impact. It's an amazing tool for big industry players with money and resources to throw around. And it's also giving us some amazing use cases in smaller industries, too-- when it's used correctly, that is.

 

Uncle Ben 

Remember, with great power comes great responsibility.

 

Peter 

The cultural conception around deep learning and AI makes it out to be something far more mystical and formidable than the reality of this tech. So, this episode, we're setting out to debunk some of those myths. Joining me on this journey is Will Ramey, Senior Director and Global Head of Developer Programs at NVIDIA. Welcome to the show, Will.

 

Will 

Well, thanks for having me, Peter.

 

Peter 

So as a starting point, can you tell our audience a little bit about yourself and your experience with NVIDIA?

Will 

I joined NVIDIA about 17 years ago, after studying computer science in university and then working for Silicon Graphics, and a series of startups, and eventually a games studio. For the first maybe 10 years at NVIDIA, I worked in several areas across the company, really immersing myself in the culture, building relationships, and getting a broad understanding of our business. But then, I served about 5 years as the Product Manager for our CUDA parallel computing platform, which has transformed NVIDIA from a GPU hardware company into more of a computing platforms and solutions company. And many people, including me, at the time, we only knew NVIDIA for the graphics cards that we designed for video games. They're basically parallel processing all the pixels that you need to make beautiful, interactive graphics. But when we introduced CUDA, it opened up a whole new world of applications that could benefit from GPU accelerated computing, including artificial intelligence capabilities, that are now powering advances in healthcare, autonomous vehicles, robotics, retail, even agriculture. Today, I run our global Developer Programs Team, where we support a growing ecosystem of developers, data scientists, and researchers solving some of the most challenging problems you've ever thought of.

 

Peter 

I'm jealous, it sounds like you have an awesome job. [laughs] Something I would love to do.

 

Will 

It is pretty fun.

 

Peter 

 So, look, you get to talk a lot with people about AI and deep learning, right? It's sounds like that's part of your regular job. What are some of the biggest misconceptions and myths that you hear out in your travels, doing your work?

 

Will 

One of the things that I've observed is a lot of people have really only seen AI in the movies. Like, you know, The Matrix, or Terminator, where the AI character has its own consciousness and some kind of internal motivation. And in the movies, they also cast AI as having really flexible problem solving and cognitive abilities like we humans do, but with superhuman intelligence. That kind of thing makes for good entertainment, but that kind of general AI is just not possible using today's technologies. What we have instead is more narrow forms of AI that can be used to perform specific tasks, working with constrained inputs and outputs that we design them to process. So, they're great for, like, building tools, and doing things like helping IT departments detect cyber intrusions, or assisting radiologists in identifying cancer tumors, and even giving customers recommendations for, like, what products to buy, or what movies to watch. And just like any new powerful technology, we're using it to change the world. But we humans still have to decide where and how it's going to be used.

 

Peter 

I find that our human expectations are always far higher than what's actually, you know, possible and practical. I'm curious what some of the top expectations that you encounter are, the people are just dumbfounded to find out they can't do with AI?

 

Will 

Some people-- maybe from the movies and books and things, might think that AI can figure out anything, and that it will eventually solve every problem in the world, right? We have some pretty big problems in front of us, like climate change, or more efficiently managing the global supply chain that gets food and medicine to all the people who need it. But these are really big and complex problems that we humans need to break down into bite-sized chunks before we can design the AI tools that will be able to perform the specific tasks that can help us solve the larger problems. And I think more importantly, we humans need to figure out that there's a problem in the first place. There are some cases where AIs can generate pretty convincing illusions of being creative on their own. But it's really just complex algorithms that have been optimized to perform a specific task. Have you seen examples of AI doing things like composing music in the style of famous composers? They tend to have kind of a style, a way of approaching music, some themes that are consistent throughout their body of work. And we can use AI to study those themes and those styles, and then generate new pieces. It's genuinely new, but it's very heavily influenced and inspired by the examples from which it learned, and so that can seem like magic, because it would take a tremendous amount of practice and education to do that for us humans. But it's something that AI can do as a computer, and so it's not so worried about having to sleep, and eat, and other things; it can just keep chugging away at it until it gets good.

 

Peter 

So, would you say that then this is the unknown definition that so many people get confused of regarding machine learning, and deep learning that we give it parameters and let it create, versus what the movies tell us where it consciously creates from nothing?

 

Will 

So, what we have today is these really, really complex algorithms that have the ability to do things like distinguish between pictures of cats, and pictures of dogs, or all the different kinds of species of animals. But the algorithm itself doesn't have the ability to do anything except that sort of category of telling the difference between different types of things. And then we use data to teach the algorithm about the specific kind of things that we want to be able to tell the difference between, so that then later, when we show in a new picture it's never seen before, it's already kind of dialed in and it knows the difference between a cat, and a dog, and a parrot, and giraffe, and so forth. And so, the AI that we have today-- really, it doesn't work like a human brain. But it has taught us that the biological processes in the human brain are even more complex and difficult to simulate than we previously thought.

 

Peter 

So basically, if we give this technology a set of parameters, we can set it out there to learn...the question is, what will it learn, and then will that produce something valuable? That the way I'm getting it?

 

Will 

Yeah, except that we know what it's going to learn because it can only learn from the examples that we give it.

[music plays]

 

Peter 

I think AI, machine learning, and deep learning are probably the most misunderstood and misconceived statements / technologies in our space right now. So, on the practical side, how should companies go about leveraging AI, machine learning, deep learning, in a realistic way? What advice would you give to someone who doesn't maybe know where to start?

 

Will 

Well, the first thing I would say is get clear on the problem and what specific tasks you want your AI to perform. You know, it's like when you're building a house, and you need to put a screw in the wall-- you use a screwdriver. You wouldn't pound it with a nail. So, you need to think in terms of what are the-- what is the tool? How is it going to work? What do you need in order for it to be effective? And then choose the right tool for the job, whether it's, you know, in your in your business or in your products. It's also helpful to think in terms of guardrails. I think you'll realize pretty quickly that if you're using deep learning or AI for an autonomous vehicle, you know, self-driving car that you're going to put yourself and maybe your kids in. Or for a medical analysis device that's going to help you understand what kind of cancer tumor you have, and what the best treatment recommendation is, you're sort of playing for different stakes than if you are just deciding whether to, you know, open your doggie door, and let the dog come into your house to get some food, or whether you're answering a customer support call. The stakes are just, you know, different. And that means that, just like the software that you have today, you need to think about what level of oversight, and judgment, and what are the acceptable, reasonable outputs that you're going to accept from the system. And when are you going to require additional kind of observation or judgment. There are a couple other things that are really important here, and that is that in order to take advantage of today's AI technology, you have to make sure you have the data, or you can get the data. Because if you if you don't have data-- which represents the experience from which the AI can initially learn how to perform the task you need it to perform-- then you're kind of stuck at square one. But the good news is, there's lots of creative ways you can get that data if you don't already have it.

 

Peter 

Yeah, I was gonna ask you about that. On one of our previous episodes, we were talking about interpretation, live interpretation, language translation. And one of the key barriers, there was simply the data doesn't exist yet for many languages around the world. Like, we've got it all for English, but for quite a few other languages, or dialects of other languages, the data isn't there. So, when you're faced with a problem like this and the data simply doesn't exist, is there a fix for that? Is there another path that you can take?

 

Will 

Well, there's some things you can do. And it's interesting that you mentioned language, because, you know, there are a lot of languages for which there is no written language, it's only spoken, and the number of people who speak it are rapidly dwindling. And so, I'm familiar with the work that some teams of researchers are doing to try and collect audio recordings of as much of that spoken language samples, and translations of it, as possible so they can build up that data. But in general, I think you have maybe three approaches to getting the data that you need to train a neural network for your particular task. The first is collect it yourself. And when you do this, make sure you label it at the source as much as possible, because that's where you have all the context for what was happening when you collected that initial data. The second is use someone else's data. There are great repositories that are publicly available under open source licenses, as well as commercial datasets that you can get access to for a wide range of domains. So that's always an option, and it's worth doing some research. And then the third option is to simulate it or generate it. You know, we do this in our own business where NVIDIA is involved in building technology platforms for autonomous vehicles and robotics. You can use simulations to create data from all kinds of scenarios. Say you have a case where you need to teach a car to drive down a road and stay within the lanes of the road. Well, if you only have data that you've collected from cameras driving on sunny days, then how confident would you be that the car is going to be able to drive well on rainy days when the lights are reflecting differently? Or at night, in the dark, and so forth. And it's really expensive to go out and drive cars on all the different types of roads, in all the different weather conditions, in all the different seasons of the year, in all the different regions of the year. And so, what we can do is we can build a 3-dimensional simulation, kind of like a video game, and then drive a virtual car through that simulation and collect all that virtual data, and then use that to train the neural network. And then you have to kind of migrate that virtual environment, or the training that happened in the virtual environment, into the real world and just confirm that it's all working. But it's a much faster, much less expensive process. And you can use the same approach of simulating to generate the data that you need for cybersecurity, and for retail product recommendations, and for molecular docking to find new drug and medicine compounds. It's a really, really powerful technique.

 

Peter 

This episode is brought to you by NVIDIA and its groundbreaking RTX Technology. One of the things I love most about my job here as an intrepid podcast host is getting to meet with all the pioneers and innovators from around the world. Which brings me to today's sponsor, NVIDIA, who for over 20 years has been breaking ground in the world of computer graphics. They've been giving professionals across industries exactly what they need to do their finest work and bring their boldest ideas to life. And now, NVIDIA is pioneering the world of professional computer graphics all over again. The NVIDIA RTX Technology has been called one of NVIDIA's most significant leaps forward, and it's ushering in a new era of applications that simulate the physical world at unprecedented speeds. With next generation AI, ray tracing, and simulation, NVIDIA RTX Professional Visual Computing Solutions enable unbelievable 3D designs, photorealistic simulations, and stunning visual effects-- faster than ever before. Whether your business works with designers, artists, or scientists, it might be time to check out NVIDIA RTX Professional Solutions and see how much faster and effectively they could be executing on their work. And as businesses are all acclimating to a new way of getting work done, NVIDIA's platform fits with your infrastructure: laptops, workstations, cloud, virtual, at the office, or remote. No matter where your team is creating, you can meet them on their terms with the very best tools and capabilities on the market. Sound like something your team could use? Get in touch with our experts at SHI today to start exploring NVIDIA RTX Professional Solutions and the future of accelerated computing. Visit shi.com/nvidia to get started.

[music plays]

 

Peter 

Clearly, there's a lot of interesting use cases for deep learning and AI that goes beyond the big names we often see in the headlines. But I also love looking forward. With all these other avenues as possibilities, what's next in the world of deep learning?

[music fades out]

So this is all starting to sound really big again, right? Back to my question 10 minutes ago around, you know, larger, deep-pocketed organizations, I want to ask you about some of the use cases-- maybe really successful ones, ones you expect to happen, or ones you've seen happen from smaller businesses in different industries out there-- just to circle back in to the fact that this is not such a massive and heavy task that these businesses can't participate and make a difference with this technology.

 

Will 

There are really two categories of this. One is where we see businesses using deep learning to improve their internal processes. And this we see happening in the areas of, like, logistics optimization and supply chain management. We see it in terms of price prediction, especially where businesses need to model out what they expect the future price of commodities, kind of the raw ingredients that they need to build their products are going to be. And these can be really, really powerful. Another example-- and here's a great example, where some researchers at Google use deep learning to look at all the data, they were collecting about power utilization and workloads in their data center and discovered that they could significantly reduce their operating costs by turning off and then back on machines-- turning them off when they didn't need them. And they published all of this research, and now other companies are able to take advantage of their findings and improve their operational efficiencies without having to invest all of the huge amounts of money that Google had to put in the first place. So that's a case of kind of standardizing publishing, making it easier for the people who come next. The other category of businesses that are taking advantage of this technology is where they're using it inside their own products or services. And so, examples like-- there's a company called Digital Diagnostics, which is using AI to take pictures of human eyeballs and serve as a diagnostic tool to help optometrists diagnose something called diabetic retinopathy. This is a condition that can lead to vision loss and blindness as a side effect of diabetes. Not everybody gets it. There is medication that can treat it. But the medication is only effective at the very earliest stages of the disease when it's the most difficult to detect. And so, by using AI, deep learning neural networks that have been trained to detect the very subtle patterns and changes over time in the eyes of people who have diabetes, they're able to detect this condition and head it off before it leads to vision loss. Another example is in agriculture. There's some really amazing work that's happening in agriculture, and whether that's in vertical farming and optimizing all of the nutrients that need to come in and out to increase crop yields, and things like that. Or whether it's out of doors, in reducing the amount of pesticides that need to be applied to plants in order to, you know, keep them healthy, and have them grow and provide great food for us to eat.

 

Peter 

Wow. Really, the impacts of this are everywhere. I mean, and a lot of them aren't even really noticeable. I mean, I just assumed that humans did those things, and made those decisions. And I have to believe a lot of other people feel the same way. It's really fascinating when you think about where this is headed.

 

Will 

Yeah. And I think it's important to keep in mind that you don't have to build your own AI from scratch. There are tons of vendors and web services that you can leverage to deploy AI where you need it in your business.

 

Peter 

I'm a sucker for futuristic technology, right? It's just my favorite thing to think about, read about, learn about. Is there a time where AI will be the magical thing that Hollywood tells us it is, and will be doing the thinking for us? Or is that just not in the roadmap for this technology?

 

Will 

I think at this point that is still over the horizon. It's something that, you know, researchers will think about and pursue in the coming decades, just because it's, you know, such an exciting idea. But there are so many amazing things we can do today. There's plenty of great things we can do with the technology we already have. The challenge is just identifying all those opportunities and taking advantage of them.

 

Peter 

I'm curious, you know, you work for NVIDIA. And you know, this is the space you're in. What are some of the resources and tools that you have that make this easier for your customers to succeed with?

 

Will 

Probably my favorite is something that we just wrapped up a few weeks ago, and that is our GPU Technology Conference, or GTC. GTC is an amazing event that we used to do in-person all over the world, and hopefully will again soon, but recently have been doing as a virtual event online. And it's an opportunity for-- this time, over 200,000 people registered and joined us for 1,600 talks presented by the experts from NVIDIA and hundreds of other companies, and universities, and government organizations. And we do all this with the goal of sharing the work and learning from each other, kind of cross pollinating our best ideas across dozens of different industry segments. And we record everything, and it's all available for free online. So, if you want to go see what's happening with AI and self-driving cars, or recommender systems for retail and healthcare, or what's the current state of what's happening in robotics, all of these things and more are in this, you know, body of recordings from the conference. It's really amazing, and it's all available from our website at developer.nvidia.com.

 

Peter 

Awesome, we'll put that in the show notes.

 

Will 

You know, 1,600 sessions is kind of a lot to pick and choose from. So, if you wanted to just watch one or two, I would start by recommending the keynote talk by our CEO, which covers a broad range of industry segments, and all of the new, exciting things that we we've just announced. And you'd be in good company because over 13 million people have already watched it. It's been one of the most popular things that we've done recently.

 

Peter 

That's incredible.

 

Will 

And the second thing is a presentation that I did myself called, "Deep Learning Demystified" where it gets into more of a high-level, lightly technical explanation of how deep learning actually works, and how to apply the tools and platforms that NVIDIA and others provide to make it practical to deploy in your business.

 

Peter 

I know what I'm going to be watching later. This is great, I actually cannot wait to dig into this content. So, look, I want to ask you, Will, a personal question. You've mentioned a lot of really awesome stories, ways that this technology can change the world, ways that your organization is contributing to that right now. I want to know what gets you stoked to get up and go to work every day right now. What's driving you personally the most out of all of this wonder that you're involved in?

 

Will 

You know, there are really two things. One is just being part of the revolution, right? I get the opportunity to kind of cultivate and connect with this worldwide community of visionaries who are advancing the adoption of this technology. And I get to talk with a lot of them about the work that they're doing, how they're solving some of these just amazingly complex and challenging problems. And that part of my work is really inspiring. And of course, getting to turn around and then share that with other people is also a lot of fun, because I get to, you know, share the enthusiasm, and pay it forward. The other thing that I'm pretty excited about is this coming revolution in immersive collaboration and virtual reality, which is being made possible by advances in both graphics and simulation technologies that are now both being accelerated by AI. So as these multiple different areas where NVIDIA already has strengths are coming together and getting better, and faster, and more powerful, I think that is going to have a really big impact on the way that products are designed on the way that we collaborate with each other. It's going to really, really step up the way that we are able to communicate with each other. And it's going to take a lot of these workflows that have been kind of built up over the industrial and manufacturing age as a sequence of steps that have really, really long dependencies that, you know, means it takes us years, and years, and years, to create new cars, and airplanes and buses, and, you know, design houses and things like that, and basically design it all in this virtual space collaboratively first. And then from there, very quickly, decompose this thing that we already know is wonderful and we want into all its component parts and assemble it. And I think that's just going to make the kind of work that is happening in these workflows so much more fun, so much more rewarding, both for the people who are creating these products, and buildings, and experiences, but also for all of us who get to enjoy the fruits of their work.

 

Peter 

It does sound really exciting. And as I mentioned at the beginning of our chat, it just makes me really, really want to work with you in what you're doing. It sounds like you have some of the most interesting work of anyone that I get to talk to. It's pretty incredible.

 

Will 

Well, thank you! I'd like that, we should talk more often.

 

Peter 

Well, I think we should! So as we wrap up, I want to give you a chance here to direct people in the right way. So, if there's anyone that's been listening to this show, and they're like, you know what, I really want to talk to Will, Will's team, I want to talk to NVIDIA about some of these things because I think they can make an impact in what I'm doing, where should they go to get in contact with you and your group?

 

Will 

So, if you want to develop your own solutions or have questions about how to do that, the first step is join our free Developer Program. Join this community of millions of, you know, smart, excited, motivated developers, data scientists, and researchers from all over the world. Log in on our community site, we have people from NVIDIA as well as members of the program from all over the world who are happy to talk with you and explore new ideas help you get started. So that's step one. When you're ready to, you know, pick a specific technology, and get started, and try applying it, you're going to find that using deep learning is a little bit different than traditional procedural software development. And so, getting a boost on that through some quick training through our Deep Learning Institute can really help you get started, get you a jumpstart on that. And the Deep Learning Institute is also linked from our developer website. And then the third recommendation would be if you're just looking for more inspiration, more examples, I'd recommend you check out the NVIDIA AI podcast. We've had over 140 episodes, every week interviewing someone new, some new startup founder, or researcher, or vendor who's got some amazing breakthrough capability that you can take advantage of in your business. And there are episodes that explain how it all works, and how it intersects with really every facet of our human endeavor.

 

Peter 

That's awesome. Thank you. Will, I want to thank you for joining me today. This was an awesome conversation. I am totally stoked about what you are working on, and I want to wish you all the best and all the success in continuing to change the world with deep learning. This has been absolutely awesome. Thank you.

 

Will 

Well, thank you, Peter. I really enjoyed it, too.

[music plays]

 

Peter 

I've talked a lot on this show about the amazing things that innovation has brought us in spite of the hardships of the last year. But as much as I continue to be astounded by the new technological advances that crop up seemingly every day, I also like asking about limitations and hurdles left to overcome. This season, we've seen some really astonishing AI capabilities, some shocking advances, but also some smaller, more modest advances. Breakthroughs aren't always accompanied by dramatic fanfare. [horns plays] Sometimes it's about rolling the ball forward ever so slightly to give everyone an equal opportunity to take advantage of some really amazing tech.

[music plays]

 

Peter 

Innovation Heroes is an SHI podcast, with new episodes streaming every second Thursday on Apple, Spotify, Google, and everywhere else. If you like this episode and you want to be our hero, leave us a 5-star review on your podcast listening app of choice. On the next episode of Innovation Heroes, I'll be speaking to Omar Elawi at Tiny Mile about their tiny, pink delivery robot named Geoffrey, and how it's helping to keep Torontonians safe while giving struggling restaurants a new hope. Be our hero; listen and subscribe to Innovation Heroes now. This episode was brought to you by NVIDIA. Learn more about NVIDIA's groundbreaking RTX Technology and products at shi.com/nvidia.

Podbean App

Play this podcast on Podbean App