Thank you. Hello and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. What's in your wallet?
Does that tagline ring a bell? I might be dating myself here, but when I hear the words Capital One, I immediately picture Jennifer Garner Or maybe Samuel L. Jackson pitching credit cards and TV ads, asking what's in my wallet.
But while it's easy to associate Capital One with credit cards and other consumer services, and there's nothing wrong with that, The banking giant is also doing a ton of work with machine learning and AI on fronts ranging from fraud detection to forecasting to customer service.
Here to talk to us about how AI is impacting the finance industry is Nitsan Mikel.
Managing Vice President of Machine Learning at Capital One.
Nitsan, thanks for joining the NVIDIA AI podcast.
Thank you. Glad to be here. Let's start kind of broadly.
Capital One's a bank, a finance company.
Why does a bank need AI and machine learning?
What are you doing kind of day to day and more broadly at Capital One, and how are AI and machine learning playing an important role in your work?
Yeah, that's a good question. It's funny, right, to think of a bank doing AI and machine learning.
I admit. Capital One is the first bank I'm working at.
I came from healthcare and did machine learning there for a long time.
So my first thought when Capital One approached me was, well, it makes sense.
We'll use machine learning to predict lending decisions and other financial forecasting and That would be about it.
So I wasn't terribly enthusiastic only because it seemed pretty specific and pretty myopic.
But with really turned me on to the opportunity and what gets me excited every day now is the fact that we use AI and machine learning pretty much in every facet of our business.
Maybe customer-facing applications like fraud monitoring, automation of internal processes like our call center operations, and there's a ton of work we're doing there. customer experience, digital marketing, basically anything you could think of even image processing and computer vision.
It all comes to play in our application of machine learning.
So can you give us an example without disclosing any trade secrets, so to speak, but an example of one of these applications in a concrete way? use?
Sure. So a big focus we have is leading the next era of banking through breakthrough digital experiences.
And we want those experiences to operate in real time And in order to do that, you really have to have them be powered by machine learning.
There's no way of making that happen otherwise.
So from a customer standpoint, we're leveraging machine learning to help customers become more financially empowered.
We want them to be able to make predictions about upcoming bills, detect irregular expenses that may Reflect something that isn't consistent with what they would normally want to do or that's even a signal of fraud.
Better manage their spending. And of course, enhance their digital experience and their satisfaction with us as a consumer.
Are any of these applications available now to your customers?
Yeah, we have, for example, Eno is our chatbot.
I believe we were the first one to roll that out in a major bank.
That obviously delivers conversational AI and delivers some of the forward-leaning promise that machine learning has been able to deliver on with deep learning.
So it sounds like you're covering a lot of bases from Business side predictions, like you mentioned, fraud detection obviously is a huge problem for consumers and banks alike. lending decisions, those sorts of things, and then getting more into the customer experience.
So you're covering kind of a lot of ground there.
How does the work that you're doing at the company speak more broadly to current trends in the fields of AI and machine learning?
Yes. So if I think about fraud as an example, that is...
What we do is really at the center in terms of the pattern we use.
A lot of the applications of machine learning and AI currently going on as cybersecurity has really had to intensify, as anti-money laundering efforts have had to intensify. anomaly detection from a systems perspective has really kicked up in the last couple of years industry-wide.
I would say that Banking is rapidly catching up with some of the technology giants in terms of how we apply machine learning in fraud-related areas.
More broadly speaking, I think that some of the trends we're observing in terms of GPU use in terms of taking deep learning from what was relatively niche a couple of years ago to pretty much mainstream across the board.
We started out with deep learning only in our voice processing space and have now expanded far beyond that.
Really it we're following the overall methodological trends that are probably true for most industries, but figuring out the best way to create value with the specific assets that we have.
So as you mentioned, there's been this rapid uptick in activity and interest in AI and deep learning across multiple industries.
And, you know, as you said, Before getting into banking and finance, you came from the health industry.
But we've had several guests on the show over the past few months. who've really pointed to the past, oh, say 18 months to two years as being this really fertile period of kind of an accelerated growth pace in A.I.
And they've pointed to hardware advances as being one of, if not the key driver.
And I'm just wondering, given especially that you've kind of spanned two really big industries and what you were just saying about uptick, of use and interest in AI deep learning and banking.
Have you seen the same recent growth in AI broadly?
And then would you also point to hardware and hardware advances as being a key driver?
Or if not, what else have you seen? Absolutely.
I think from my perspective, both in healthcare and financial services, What has been pushing in this direction for the past couple of years is really the need to deliver real time. experiences.
People are now pretty much accustomed to everything being in real time and that's what they want to see. the kind of distributed computing infrastructures that we had previously just couldn't cut it.
So if I think of our fraud prevention experience as an example, we currently are delivering real-time fraud prevention to our customers. starting all the way from interactive alerts So customers can feel protected when they're alerted to potential fraud.
They can easily report fraud transactions.
They can lock their card in real time on the spot.
They can continue spending on valid purchases, again, in real time and receive verification when a new card is on the way.
So we're leveraging structured but also massive unstructured data to detect when a transaction is potentially fraudulent and We've already been able to reduce false positive rates dramatically.
You can imagine that doing all that in real time just requires a different kind of hardware than was previously needed.
Sure. Apart from the day-to-day application, like you mentioned, the consumer experience and fraud protection, obviously.
Apart from these day-to-day applications of AI, are you doing any research in the field in Capital One?
Yeah, so obviously there's not going to be a ton of innovation if there isn't a lot of research.
There's probably a lot more research that ends up not going into straight production because there's learning that doesn't generate a product right away.
No, of course. You're probably thinking of some of the investments that are farther out.
And to that, we definitely have a number of investments going on.
For example, we have the explainable AI effort at the forefront of where we're trying to innovate.
You can imagine as a bank where fair lending is incredibly important to us.
Eliminating bias is incredibly important to us.
Figuring out ways in which we can deliver machine learning in a way that is... unconstrained by opacity, the way a lot of the deep learning, for example, approaches would traditionally be constrained, something we're incredibly invested in.
Our guest today is Nitsan Mikel. He is the Managing Vice President of Machine Learning at Capital One.
And after spending a chunk of his career working with AI and deep learning and data in the health industry, he's made the shift over to banking services.
So you've been with Capital One for roughly a year now.
Is that about right? Yeah, that's about right.
So if we go way back, how did you get into the field of AI and machine learning?
Yeah, so I'm going to date myself the way you did.
I started in this field before it was called machine learning or data science.
At the time, it was just around the time of the Human Genome Project.
It was one of the early applications of high performance computing to massive data sets that were growing really rapidly and there was This realization that the way we were doing things previously is just not going to cut it.
And we had to... develop new tools and new approaches altogether.
So that's how I got into it. I worked for quite some time in the genomic space and then moved on into medical technology because streaming data there is really literally a matter of life and death. some of the decisioning you can make with machine learning just can have such a huge impact.
Imagine that you spent some time around the time of the Human Genome Project working in that space with genomes.
Were you involved in the Human Genome Project itself?
Or if not, what were you doing? And is there kind of a—it sounds like a fascinating area of research and obviously impact— but then also kind of with the growth of high-performance computing.
Is there a story? Is there kind of a 60-second version of what you were doing then and maybe some lessons learned?
Sure. Again, dating myself, it's going to sound very primitive now.
I wasn't on the Human Genome Project itself, but I was doing comparative genomics.
That's where you're comparing long sequences of DNA between individuals, A lot of this has to do with disease detection and genetic mapping.
And then even between species, which is some of the work I did as well.
What you're trying to do is identify anomalous patterns as well as association patterns between long stretches of DNA and some both marked and unmarked labels.
Even though we had access to relatively sophisticated hardware, it was something where we would, you know, send it off for analysis or start the analysis and have to wait for two weeks to get some results back.
So, Honestly, I even played around with neural nets at the time.
It wasn't very promising, but it was just conceptually such a cool idea.
Really, there just wasn't the kind of computing power that would allow you to get very far there.
So most of it was random forest and decision trees.
And even that would take days, if not longer to get any results back.
If you're in that headspace now and then sort of fast-forwarding to 2018, Was there kind of a period or even a moment where you sort of saw, not the light at the end of the tunnel, but almost like a breakthrough to a new tunnel with light at the end of it because of... something that happened in the hardware world or suddenly these two week waiting period, you kind of realize like, oh, this thing is going to change all of that.
Yeah, I wish there was a single light bulb moment.
There isn't quite that, but I would say when really... the move to the cloud became inevitability.
That's when things changed. Now I was in a really regulated environment, still am.
So things were slow. I think we were probably somewhat behind the curve, but Probably around 2010 or so, maybe 2011 is when I started seeing that light at the end of the tunnel. and could really conceive of using some more sophisticated methods with the kind of data we were working with.
And so thinking about the present day and not so much the applications of what you're doing, Get your geek hat on, purely from a technical perspective.
What kinds of challenges are you grappling with?
Is there something that gets you up in the morning that you're really... psych to kind of chip away at and, and, you know, whether it's something you're architecting or a more specific problem you're solving.
Is there anything that kind of gets your geek crank going as our producer likes to say?
Yeah, I actually really like thinking about how problems that various business stakeholders think is really unique and really different. is actually not that unique at all.
Now, I know that sounds like a downer. I find it really cool seeing patterns across problems that indicate a common solution.
Because the last thing we want to do is be building one-off solutions every time And so looking at anomaly detection in the fraud space and then anomaly detection in the customer experience space and seeing that that pattern is the same and we can go after it with the same solution is what gets me up and gets me going.
So taking from there and now looking ahead, let's say the next five years, What do you think some of the forward-looking implications of your work or even sort of AI in the finance and financial service industry kind of more generally you know, what might the implications be?
Where do you think this is headed in the near term?
Yeah, I think the jury is still out in many ways, but At its most basic form, what I think machine learning is doing now, it's collapsing away the distance between the companies and their consumers.
So in our case, the bank and our customers.
And with that distance going away, We're able to really understand the customer and react to what the customer needs in a very immediate way.
And course correct in a really immediate way as well.
And so what that means is the whole relationship between a customer and a bank changes because now we can be a real advocate for them in the most direct possible way.
Well, before... There was just this big gap and a customer was, for lack of a better word, was sort of a number.
Now we can really understand their context.
And the machine learning aspects of that, because I've heard that similar sentiment expressed by companies that do CRM software or other platforms.
From your perspective, is it just that there are so many customers and so much data in the financial space that you really need that? high-performance machine learning to be able to make sense of it?
Or how does the ML play into getting closer to the customer?
Yeah, no, I would say it has much more to do with instrumentation. actually being able to see what the customer is needing in a much more ongoing basis.
But then it's also being able to build machine learning models that take that full account of the customer's history with us so we get to know them better and better over time.
As some of your listeners will know, for example, building a large-scale production-grade time series anomaly detection platform. is really hard.
Um, computationally it's incredibly intensive, but it also scales really unfavorably.
Like the longer the history of a customer, um, or any dataset, the longer the sequence, the harder that problem gets.
And it scales very unfavorably. So now with some of the hardware improvements, whether it's with GPUs or even some of the older distributed computing solutions we're able to do that in much greater scale or at much greater scale and therefore be much closer to the customer.
So it's less about Nothing against CRM, but I feel like that's very 2008.
So apart from the day-to-day applications of AI and machine learning that you just spoke to, I know that research and development R&D is a part of any big technological effort.
I'm sure Capital One has machine learning research going on all the time.
Can you speak to anything that you're doing that, isn't necessarily going to show up in a product tomorrow or maybe never as research goes, but things that you're doing with an eye towards the long haul.
Sure. So broadly speaking, Capital One has machine learning at the very core of its R&D agenda.
Some of the key areas we're working on are accountability, explainability, things like explainable AI, ethics and fairness, governance, of course, security and privacy. are incredibly important as well.
We partner with some of the top universities to further these research areas and critical elements of our business.
And also in some of the machine learning theory that would help advance our ability to interact with regulators and explain what our agenda is.
I can go into a little bit more on the explainability and fairness. if you think that would be interesting to your listeners.
I actually was going to ask, I didn't want to trip over my own tongue and ask you to explain AI explainability, but yeah, that would be great.
Sure. So being in a heavily regulated environment, we want to make sure that we're not just meeting the regulatory requirements, but that we help set the standard for what fair and ethical machine learning deployment looks like, specifically in the financial services.
So there's great potential to leverage data analysis and machine learning to gauge creditworthiness. but we want to make sure that we're balancing the use of this technology some of which is more opaque than earlier modeling efforts, if you think of deep learning as the example.
With the appropriate development application of adjacent algorithmic technology that ensures fair and unbiased outcomes.
So we really want to maintain the highest standards for explainability in an ethical and fair way. as we develop more advanced models for more use cases.
So we want to push the envelope when it comes to the sophistication of our models and our algorithms.
But at the same time, as you push that envelope, it becomes harder and harder to explain the We're investing a lot in research that will help us make sure for ourselves that we're always doing the right thing and also communicate to our customers and our regulators. that we are doing the right thing for our customers at all time.
We do have a Capital One position paper on fair lending at the Fat Trick Workshop on Responsible Recommendations in October.
When it comes to things like lending decisions, say, because that's something I think easily graspable for people.
How much of that decision-making right now is being informed by a machine learning system? versus another perhaps less opaque algorithmic system or even human oversight?
How do those things work together? So Capital One from its very beginning was very sophisticated in our decisioning.
So we've always... leverage data in a way that most banks weren't.
Decision modeling has always been part of our decisioning ecosystem.
But it's very much with a human component and always with oversight that ensures there's nothing wrong. that's going to go awry in that decision framework.
Now, as I mentioned, we're continuously improving what we can deliver from a machine learning perspective, but we're never outpacing our ability to do that in a way that's completely responsible.
For Capital One customers or not, because everybody at this point has a stake in the banking system in one way or another, it's good to hear.
As you said, I think anyone listening can can attest to that sort of tension between the excitement of the system getting more powerful and sophisticated.
And also, you know, that that feeling of even though I built the thing, do I fully understand where it's kind of going now with its decision making?
Yeah, what's also exciting for me is banking hasn't traditionally been at the forefront of technology innovation. in large part because of the regulations which are important.
It's exciting that when it comes to explainability, I think that we're incredibly well positioned to take the lead there. even more so than our technology companies, because it's really at the core of what we do.
And it's incredibly important to us to make sure that that's always at the forefront of every technological innovation that we take.
So for folks listening who might want to find out a little bit more about what Capital One's doing in these fields, AI, machine learning, deep learning, is there somewhere online they can go?
So we do have quite a bit research submissions, etc. at the various conferences.
In addition to that, we have the Capital One blog that always summarizes that information as well as provides additional information that we would be excited for your listeners to come on to.
Excellent. Nitsan, Mikel, thank you so much for making the time to speak with us about what you're doing at Capital One.
Given your background working with data and machine learning and what high-performance computing was called prior to that.
It's very cool to hear your excitement about having hopped from You know, what the service may seem like very different industries, genomics and health care to finance, but getting a chance to sort of see.
There are similar problems, similar approaches, similar outcomes, no matter the industry.
It's very cool to hear about. Thank you.
Thank you so much. Really enjoyed talking to you.
Thank you.
