Hello and welcome to the NVIDIA AI podcast.
I'm your host, Noah Kravitz. Tech types like to talk about being ahead of the curve, being first to market, and all kinds of other fancy ways of basically saying, ha-ha, beat you.
Our guest today doesn't need to brag about that kind of stuff, but he got into big data way before most of us had even heard of big data.
Wayne Thompson is a 25-year veteran of analytics software developer SAS, where he's now the chief data scientist.
Wayne's here to talk about the big data, big models, and big computations driving deep learning and to give us some perspective about what makes today's deep learning different and truly new.
Wayne, thank you so much for joining the podcast.
Glad to be here. So let's just start with an open-ended question.
What does the chief data scientist at SAS do?
Well, it's the best job on the planet, first of all.
I mean, it's really a good gig. And I work very closely with customers and R&D.
And what we're trying to do is really infuse artificial intelligence into all of our products.
We're not trying to necessarily build an AI product.
And an example of that might be putting in AI to make suggestions for business intelligence, and we're also trying to push that down into our solutions and make it open for customers who aspire to build their own applications.
And it's been a good gig. And I've been doing that, like you said, for 25 years.
I did some product management and so forth.
So I'm imagining that you work with a pretty broad range of clients across different industries.
Looking back, I mean, it's a big question, but looking back over those 25 years and all the different clients and the way that data and the use of data has evolved, What jumps out at you as far as watershed moments or projects that maybe looking back foreshadowed where we're at now, where we're going?
Yeah, that's a great question. Because in machine learning, supervised learning or predictive modeling is kind of where you get your bang for your buck.
And That's kind of bread and butter stuff.
And 95% of what I see our clients do, they're trying to predict fraud or...
Purchase propensity or some type of classification problem.
Right. What really amazed me is early in my career, we also got involved a lot in text analytics. and being able to blend what customers are saying into the predictive models.
In this text analytics at first, I was kind of amazed at how well just counting and creating, how often a term appears in a document across all documents and then using stuff like a bag of words or singular value decomposition. and just how well that stuff would cluster customer comments how well you could derive sentiment And so forth.
And that really, I thought, was kind of earth shattering to me how well that worked.
And now we've got recurrent neural networks, which are even doing better.
So it's crazy. Right. I was going to say, how long ago were you thinking about when you were just doing text analytics?
So we started off with our first flagship data mining product called Enterprise Miner.
And then I'd say about in 2008, we brought in the text miner.
Right. And then after that, you know, people wanted to also manage models.
So it got into It wasn't as cool and sexy as just building models, but how do you maintain models like assets, just like people in your organization?
Right. How do you refresh those? And how do you get IT working together with the data scientist?
And so we built Model Manager. And then the key for us, too, is with a company like NVIDIA, we need to bring the analytics to where the data is.
And speed is important, but the quality of the models is also huge.
And so being able to do in-memory computations and fit the data very nimbly, very quick, almost like it's up on a magic carpet, That's the way a data scientist likes to work.
And, you know, that's what I'm doing now.
A lot of interactive, bursty type analysis. how much would you say the rate of the technology itself has not changed, but accelerated?
I've, Talked to a lot of folks over the past couple of months who are really talking about the past couple of years as being this huge accelerating point for AI in particular and using GPUs and deep learning and that kind of stuff.
You know, kind of looking back over a broader time or in your 25 years we're talking about, Is this current period really kind of a rapid acceleration pace, or is it more of kind of a... an ebb and flow thing.
And right now we're on an upswing. I think it's still one of the major technology shifts.
You know, prior to this, we had mobile. Prior to that, we had the internet.
You know, we had the computer chip. This is major, man.
People are saying that, you know, this is a $36 billion business by 2020.
So I'm pretty stoked about it. So also, I think that let's not forget bread and butter machine learning, logistic regression, decision trees, classical machine learning methods are really good.
Now we're doing things I call data scientist arborist.
They actually build not one decision tree, but thousands of them and ensemble them.
Right. And gradient boosting is still rocks in terms of classification.
But now when you get into things like computer vision, The machine can see better than you.
Convolutional networks rule. Deep learning rules there.
If you think about speech to text, 10% of all text... now probably is done through speech.
These recurrent neural networks are just killer with regards to being able to do that.
So those... Those two application areas specifically are wonderful for deep learning.
So when you're working with clients specifically, Are you providing them as a goal to provide them with recommendations and solutions?
Or are you looking to, you talked about infusing AI into the software.
Are you looking to give them smarter, easier to use tools that they can use on your own?
What is a kind of example client? That's an awesome question, really, because we try to cover the full gamut.
You know, we're a tools provider, first and foremost.
We do technology and we provide. a full suite of machine learning tools that you can use to program and roll your own code and It's very open, very extensible, including REST APIs.
And you can write Python or you can write SAS.
And you can build your own AI and plug it into your systems.
But we're also trying to do things like customer intelligence, CI.
There we're looking at things like the ability to understand customer journey. and use things like reinforcement learners.
And from those, not only just look at the short-term reward, associate it with these clients, but really get involved with the customer.
Treat the customer like a guest and help retailers like Target Understand the long-term value from these customers and do that with AI.
So what would bringing AI or infusing the customer journey with AI, what might that look like?
Well, you know, customers change and there's always a huge temporal effect to that as well.
So timing is everything. And, you know, in addition to just standard AB testing and recommendation engines using collaborative filtering and so forth. you want to optimize the experience with your customer.
Sure. And just like beating the game of Go, there's lots of different moves that customer can make.
And so you need to be able to respond to those with the right policies and rewards.
And the one thing about this, though, requires is a ton of data.
Yes. It kind of goes, you need a bunch of data to do deep learning.
You need even more data to do unsupervised learning.
And when you get to reinforcement learning, you better be working. with major, major firms that have lots of transactional history about their clients.
Right. So there's an idea out there, and I think it might be attributed to a blog that your name's on, that there are three key reasons right now why deep learning is popular. and they all start with a B. There's big data, big models, and big computations.
So we were just talking a second ago about large companies that have lots of transactional data.
How can you help businesses that aren't so big get their hands around these three Bs, the data, the models, and the computations, so that they can put deep learning to work?
That's cool because, you know, I really feel like at SAS we've been doing this a while and our customers are already very successful with machine learning. and very automated machine learning.
And it's the mid-market, the smaller customers that I'm excited about working with.
And it's really practical in my mind. The first thing is start with a fundamental good business problem.
Work with the business folks. If you're working with scientists, work with the researchers.
And you've got to have data to solve that problem.
Right. And that's one of the issues I find that many companies are data poor.
So it takes time. Focus probably on supervised learning because there you have a label.
You have an outcome. And it's easier to kind of demonstrate the value of being able to apply predictive modeling for inferencing, either in real time or batch.
And the business unit can see that and see the value quickly.
The other thing is you've got to get the technology stack right.
Start small. Don't go crazy. Right. And even start with some basic models, too.
You know, like I talked about logistic regression, decision trees, principal components, things of that nature.
But be careful because you're going to grow.
You know, once you... Get your product built and get this model deployed.
You're going to have lots more users. You're going to get a lot more data.
Right. And therefore, you're going to grow that platform.
And that's where you're gonna start looking at Nvidia.
And you're gonna say, hey, maybe I do need something like a Volta.
Right, right. With TensorCores to build my models.
And you're going to use a mixture of software.
So make sure, like, when you invest in SaaS and something like Visual... statistics to start with that whatever platform you choose is open and can accommodate open source and being able to extend and tie into operations.
We're talking with Wayne Thompson. Wayne is a 25-year veteran of SAS, the analytics software developer, and he's now the chief data scientist there.
Let's talk about you for a second and go back in time a little.
Before you joined SAS, I know, but we go back before SAS.
How did you get into all this stuff, into data science?
What was your background prior to where you are now?
Well, my family actually had a farm, and we raised cattle.
Okay. Where did you grow up? In Greenbrier County, West Virginia, where White Sulphur Springs and—
The hotel is there and it's a beautiful Shenandoah Valley, rolling hills with limestone outcrops.
It's absolutely gorgeous. Right. And I always wanted to get my hands dirty.
I wanted to feed the world and learn agriculture and went to school in Tennessee and learn how to do grazing trials and collect data not only about the animals, but the forages, the climate, the soil, and much like Aristotle, who I think was the first data scientist,
Because he studied empiricism and learned by observation.
That's how I got the bug. And I started, by the way, oh, well, I'm going to take this stack class.
And now I'm going to take this multivariate class.
And you know, learn some code and be able to write perhaps in SAS and be able to write in let's say SQL.
And so to be a data scientist, I think there's a lot of us that come from a lot of different fields and sometimes they make the best data scientist.
Sure. You mentioned agriculture, you know, your background on the farm.
Are you doing any work now at SAS with agricultural companies using deep learning machine learning?
Absolutely. I mean, if you look at this, the USDA uses a lot of SAS for geoanalysis and estimating crop yields and things of that nature to also plan economically.
A lot of our customers definitely use us in being able to understand genetic engineering of new types of plants and so forth.
One particular organization is trying to look at removing this fungus from this plant.
Right. And the issue with this fungus is it kind of, first of all, it provides disease resistance and drought tolerance to the plant, but it makes the animal very sick right so it's you know kind of a genetic problem we're trying to engineer the right type of forage that can survive in these conditions yet be profitable to the farmer.
That's one very specific example. I was talking to somebody earlier who also kind of a veteran of the industry and is his varied background himself, but is now working in deep learning.
And he kind of offered up almost like a advice to the young data scientist or deep learner to try to get as broad of an education background as you can before getting too specifically, not his words, but boxed in to, you know, doing one very high-level task very well that gets you a good career right now.
But if the ground shifts in a few years, and inevitably it will one way or the other, you might be kind of stuck.
Do you have any advice to somebody starting out in the field, listening to the podcast who maybe is in school or looking around for their first job out of school, working with data, working with deep learning? anything looking back that some wisdom you can pass on.
Yeah, that's another good question. First of all, I'm just really – I'm really kind of stoked about the talent of – young data scientists today coming out of college.
You know, because we didn't blend in that computer science experience.
You kind of had to pick that up on your own.
And so now I'm seeing really at least two types of data scientists, those that like to build product, data product, And focus on that.
Get a Raspberry Pi and learn how to build something.
Write code and develop a little app. Go out and write some Python and create a chatbot.
And then the other piece is the blood also, the blood type I look at is analytics.
If you're going to build these products and use machine learning, you need to look at a lot of these offerings out there for courses, SaaS offers, you know.
Free programs, in fact, and free software and the other courses like at Coursera and stuff like that really rock too.
So there's – learn the analytics, and then strive to kind of blend both together.
That's what's going to make you strong where you're type AB data scientist.
Right, right, right. And there's not enough of them out there.
But yet they're coming out of school with a lot more of that training than they did five years ago. which is encouraging.
Absolutely. I mean, definitely. So there's a lot of things you're stoked about, which is excellent.
Optimism is the word of the day today. Looking forward, technology-wise, applications of the technology-wise, anything, just in the data itself even, what are you excited about?
I think that we need to start thinking about unsupervised learning more.
The question about can AI reason? Does it have a conscious?
I'm not going to be totally down on that either.
I think that eventually, you know, with unsupervised learning, that's the way a child learns.
Right. You probably taught your children things.
I tried. You tried. Of course you did. But they learned a lot of things on their own purely by observation just like Aristotle again.
And so I could teach a deep learning model that that's a table.
And give it a bunch of labels and through back propagation, it can look at that table thousands of times.
And yes, it can tell you a table. It's a table tomorrow.
Right. But it can't tell you that that's also something if I stand on, I fall off, I get hurt.
Maybe it's a hiding place. I think that what's going to happen, the hot area is being able to simulate these environments. for unsupervised learning so that machines can truly learn like humans.
And I'm not sure if we're 20, 40, 50, 60 years away from that.
But that's where the action is. And I also think these adversarial networks are cool.
When I talk about supervised learning, that's a discriminative model.
You're trying to discriminate between two classes, fraud or not, for example.
Right. Whereas in these adversarial models, you're kind of looking more at multiple competing models, and you're also looking at, hey, given this label, This is cancerous.
What about this can I learn? What are the features that really identify that?
And again, that involves being able to simulate the real world condition.
You know, if you're trying to build an autonomous car, you just can't keep running them off the road.
That gets pretty expensive. Right. That's where some of the action's at.
Keep in mind, machine learning is supervised, unsupervised, semi-supervised.
Again, we don't have enough labels. And then reinforcement learners.
The last three, just beginning to touch on.
Just beginning. So you mentioned some of the tools and courses that SAS makes available.
If people want to find out more about those, about your work, about working with SAS, jobs at SAS, anything related, where can they go online?
Yeah, just go to www.sas.com. And also we have lots of programs at universities to where you can get a master's degree, for example, in analytics.
And they're very practical, like 18-month programs to where maybe you've already gone to college and you've got like your BS degree, but you want to get involved in analytics.
Sure. We've got a killer program for that.
Most of those students are coming out getting hired, like 99.9% of them.
Right. And they're making good dough. Yeah.
Can't beat that. No. Wayne Thompson, thank you so much for your time.
You're a busy man. You're up to a lot of fascinating things and look forward to. following your career and what SAS is up to over the next couple of years.
Thank you so much. My pleasure. Thank you.
