Listen: Explainable artificial intelligence could be game-changer for banks

Explainable artificial intelligence (AI) could be the key to making AI more accessible for regulators and banks looking to provide insight into how the technology makes decisions. One of the challenges with AI in financial services is explaining AI’s “black box,” which refers to the limited visibility, even by developers, into an algorithm trained with machine learning (ML). This is one reason AI can include bias and run afoul of regulators.

Subscribe to The Buzz Podcast on iTunes , Spotify , or download the episode .

Explainable AI could change that, said Stuart Tarmy, global director of financial services industry solutions at Aerospike. The company offers a NoSQL, or nontabular, nonrelational database that powers real-time networks for PayPal, Visa and Charles Schwab. Explainable AI is “almost an artificial intelligence system on top of an artificial intelligence system to explain how it's making decisions,” Tarmy tells Bank Automation News in this podcast. Explainable AI offers two advantages: It provides insight into how the AI makes decisions so models can be tuned, and it addresses data privacy regulations, such as the European General Data Protection Regulation or California’s Consumer Protection Act, which both mandate that companies be able to explain to consumers how their data is used and how a decision is made. “If a company is using AI, ML, they don't know exactly how it's being used, or how the decision was made,” Tarmy said. “So, this idea of overlapping explainable AI on top of it gives them that map of how the data is being used.” Explainable AI is cutting-edge tech, Tarmy said, but some companies are already building it. In this podcast, he delves into how banks can leverage this technology. The following is a transcript generated by AI technology that has been lightly edited and may contain errors. Loraine Lawson Good day, my name is Loraine Lawson and I’m an associate editor with Bank Automation News. Recently, I spoke with Stuart Tarmy, global director of financial services industry solutions at Aerospike, a NoSQL database company used by PayPal, Visa and Schwab. We talked about what’s new with artificial intelligence and he explained explainable AI, deep learning and neural networks.Loraine Lawson What are some new use cases for AI? Do you think especially in terms of financial services in automation, or really just anywhere where they're automating?Stuart Tarmy I mean, financial services is where the use of AI and machine learning are exploding in financial services. And it's, it's become even faster and really accelerated because of COVID. And part of that reason is because people because they can't click the branches, they can't be with a financial advisor, face to face, or moving online. So in response to that, a lot of what the financial services firms are doing is they're building recommendations, recommendation engines chatbots. automating the interactions with clients, they're moving work toward Robo investing, where they can do automated investing recommendations for clients without a necessary life financial advisor. They're moving to mobile companies. Are you have new banks or neobanks challenger banks that are entirely being written to is online without branches that the legacy banks now have to compete with? And all this is being driven by AI machine learning. And another dynamic we're seeing is just is just in fraud. So as people can't go into branches, now they're moving online in heavily getting more types. fraud and mostly around card not present fraud. I'm not gonna give you my card, and identity fraud. And so AI and n LR are powering these to become much more efficient and effective out there. And we can we can talk about these in more detail. ButLoraine Lawson yeah, I was just gonna ask, I was just gonna ask when it comes to fighting fraud, how does fraud — not frogs — How does AI play a role? Stuart Tarmy Sure. So when you think about, you know, fraud systems out there, there's different types of fraud, right? There's first identity fraud, which is, is Lorraine who she says she is, right? Or was it a stolen card, and then there's once I've identified that it truly is you, there's other ways you can do fraud, you can have theft, you could be doing things knowing you're going to default or go into bankruptcy. So I'm looking at both identity fraud, and, you know, theft, or type of fraud later on. So sort of a history of, of systems for fraud is the early systems and, and 50% of all systems out there are still this way, are, are what are called rule-based systems. So rule-based systems are when you just build rules, and they can build up these libraries of 1000s of rules that say, you know, if you made a credit card transaction in one state, and you make a transaction in another state, within a 30 minute timeframe, decline it, right, that's a rule. And you just built you can build up these rules. And you know, the leading firm for rule based fraud system, you know, people typically think of Actimize and Actimize as we moving towards adding AI ml into their, their fraud solutions. But the problem with rule based systems is they're hard coded, right, I have to know before the fact what my rule should be. And as someone defrauds me, I'll add more rules. But you're always sort of a day late. And so the fraudsters often know that they they do new things that aren't captured by the rules. So the trend has been to move to AI, artificial intelligence, machine learning based systems that can be more self learning, and can use much more attributes. So AI ml based system, often could look at anywhere from 10 to 100, hundreds of attributes, and can do it very quickly. And self learning, there's things that we really want to dive underneath the covers, there's things called supervised learning, unsupervised learning different techniques. But the idea is it's it's sort of meant to sort of mimic the human brain, how a human would do things. And people have been moving dramatically there. But people things don't stop there. So if I look at a company like PayPal, PayPal has moved very much into AI ML, and they’re clients of Aerospike. The evolution of fraud systems moves even deeper into what they call neural nets, or deep learning systems, where you're truly trying to mimic the idea of how human brain works. And it's not exact by any stretch. In fact, if you talk to a biologist, you know, he'll tell you neural nets, the brain does not work -you know, it's a it's a crude approximation. But what allows you to do is to process much, much more data, and much, much more attributes. So if I talk about AI systems, that I said, they'll process anywhere from, say, a 10, to 100 200 300, attributes, neural net deep learning systems actually will process a million to 10 million attributes, and look at it where everything is interlinked in this huge network, the leading players in that area is PayPal, PayPal has always been considered very good at fraud. They found using these neural networks, that their reduction, false positives and false negatives, and the performance has actually gotten 30% better, which is really incredible, because they were good before. Now the 30% they're claiming better. So, you know, as you look at it as a, as a technology platform in the evolution of fraud systems, you know, neural nets, deep learning is, is to the far right. And there's even one beyond that, which I've talked to people about, called explainable AI. And what explained why AI is, is the, the ability, it's almost an artificial intelligence system on top of an artificial intelligence system to explain how it's making decisions. So one of the problems in a sense of AI systems and neural net systems is you don't necessarily know why they're doing what they're doing, or how they work, you know, the input you put into it, you know, the output that comes out of it. And the decision that's reached, but you don't know how it's how it's, you know, what, it's what's calculated to make that happen. I mean, it's literally could be processing in a deep learning system 10 million attributes in under 20 milliseconds to make that decision. To explain why AI is the idea to overlay on the AI system, as a system to actually explain why decisions are getting made. And there's a couple advantages of that. One advantage is that, the more you know about how it's making decisions, the better you can tune your AI system re AI model. So if I, if I know, you know, XYZ, I may know, I need to add more attributes, or I may say, you know, the attributes in there that really don't affect the outcome. So it's important to do it for that. The other reason is around data privacy. So if you look into the, say, in the European GDPR, General data protection regulation, or California's ccpa, California Consumer Protection Act, and there's other ones that are coming, some of the requirements there is that if a consumer asks you, you know, how is my data being used? Or how do you make a decision, you know, why did you decline my credit card transactions? example? How are you using my data for marketing purposes, you need to be able to tell by by regulation, and if a company is using AI, ML, they systems, they don't know exactly how it's being used, or how the decision was made. And so this idea of overlapping explainable AI on top of it, you know, gives them that, that map of how the data is being used. But that's far out stuff out there. very cutting edge cutting edge stuff. Loraine Lawson Yeah, I wonder, do banks even have access to that level of technology yet? Or is that still, you know, big tech stuff? Stuart Tarmy Well, they do they do in the sense that, you know, if you're a a large large institution, you know, they have many, many computer science people who are, you know, building AI systems, doing work in enormous large places, pay pals example is heavily invested in that. And, and they can be building explainable AI into their systems. What's interesting is the most primitive of the systems which is rule based, is actually the easiest to do explainable AI because you can, you can trace the rules that are buyer, right, so it's in those systems. But what's happening also is that third party vendors are building explainable AI or explanations capabilities into their systems. So if you use a third party I'll give you an example that's advertises says that a billion into it is DataRobots is building explainable AI capabilities into their, you know, their routines and DataRobots, you know, used heavily by companies to, to run and test AI algorithms. If you want to get you know, extraordinarily advanced. I've read things where were DARPA. DARPA stands for Defense Advanced Research Projects, they just say it's the advanced scientific arm, VC type arm of the Pentagon, has Orbis initiatives around explainable AI. And they often fund universities and industry. So there's a big, big efforts going on in this stuff today. Loraine Lawson And since you're in financial services, say how the regulator's respond to this cursor thing, because sometimes the AI can be a bit of a black box, and I understand regulators like things that are very explainable. Stuart Tarmy Right, great question. So, I mean, fraud is, so it's a little less than fraud, and the regulators want to make sure that companies are not being defrauded. But the the bank as an example there they own cost if they're defrauded, right, or fraud letters, levels rise above an acceptable rate. And acceptable rate is usually considered below 80 basis points point 8%, I guess, fraud, overall average is usually about 100 basis points 1%. And it can differ by country and industry you're in. But there are other areas where data privacy as an example, the regulator's care, you know very much in terms of how decisions are being made. You see often the papers, bed bias, right bias, so these systems, they wonder how decisions are being made. There's also, you know, a number of regulations in capital markets and trading as an example. And a lot of these systems are being powered by AI and machine learning. So they get very interested in how trading decisions are being made at our companies, you know, adhering to the regulations and compliance limits for things. So, you know, there's normal audits that occur with regulators with with banks and asset management companies. And that's one of the things they you know, they ask it out and do audits. Loraine Lawson Can you just explain to me a little bit more about neural networks and deep learning and what technologically is going on there. Stuart Tarmy Sure, sure. So a neural net is, it's a branch of, I guess, computer science or AI machine learning. And that's sort of a catch all phrase to say, you know, how can we mimic the human brain? In computer software? Like, how can we actually develop the system is to sort of mimic how the brain does things. And the idea was that the brain is really made up of cells. And if you could look inside somebody's brain, you'll see a cell, either, you know, fires or not, right? It's either either the fires, fires electrical impulse to the next cell, or not, on the cellular level, I guess, the neuron level. And that's, you can sort of mirror that in a computer by, which also works on ones and zeros. And so basically, what they're they're doing, is they're taking a system and saying, so I'll talk about neural net, it's hard to show you without a picture. But you can envision a if you think like a rule based system where there's a tree of rules. And if a rule, is this true, it'll fire to the next, the root of the tree, so to speak. And so there's sort of a yes, no, of what happens there. And it's a one to one, right? If there's a yes, it'll fire to something else. What a neural net will do, this is very crude, we'll say if it's a yes, it can fire to the next node, but it can fire also to 10 other nodes, in a sense. And that's the next line down. And that's what's happening. And then it can be recursive, in the sense that, depending on what the response is, it may want to give a feedback loop to the prior roots in the in the graph. It's very hard to make sense, it's very hard to show without that pictorially, it's much easier. What deep learning means is you just replicate that many times. So you have nodes that have, you know, a yes, no, this results from an algorithm that fires a zero or one into the next node, depending on how it performs. And then it can fire again into a next node down the line. And the next node down the line, and you can just mess these together, you know, infinitely, so to speak very, very deep. And that's why it's called Deep Learning. Loraine Lawson I guess that's why you need something like a no SQL database, you need a big data solution to connect with this is that Stuart Tarmy that's correct. So a couple of things happen. There's two reasons why as example, companies choose Aerospike. One is you need to be able to handle enormous amounts of data am I talking enormous, I'm talking could be 10s of terabytes, hundreds of terabytes, we have clients in in 10 petabytes of data, just enormous more than you can even think about just a norm. I mean, think if you were, you know, Verizon media, or, you know, Schwab, you know, just huge amounts of data. So that's one thing, you have to be able to handle all. And the second thing is, is that people want the response back in real time, they want the answer fast. So for instance, when PayPal does a transaction you want to do is send a PayPal money from you to somebody else, they have to authenticate you and another person very fast, you just expect it to go through, you need the power to be able to do that very fast. You play acetate, what is very fascinating, what is real timing. So the the best in class type companies, just to test valid against example, will make that decision over enormous amounts of data in 20 milliseconds or less, in less than 20 less than 20 milliseconds, which is, you know, to two tenths of a second, they have to decide whether to make that happen. So you need a system, a huge system to make, like Aerospike that's optimized for speed and scale. Loraine Lawson It's way faster than I can make a decision. So when you talk about all these technologies has advanced technologies, what will they mean for banks and automation? You've talked about fraud a little bit, but are there other use cases or what what me? Stuart Tarmy Yeah, yeah, there's use cases all over. So one of the areas I talked to people about saying in capital markets or asset management, mutual fund companies, Wall Street is you know, they typically look at their applications as front office applications, middle office applications, back office applications, and we work with companies in all those areas. So examples of those are a front office application is customer facing. So it might be things like, right giving you recommendations, you know, you you your portfolio looks like XYZ today, and you should just it given interest rates and where the market is today and and your personal family situation. It can it can be programmed to do that automatically. So can do automatic automated recommendations at the one to one level for you you have this investing for quantitative trading and analytics are large you know the analytics is front office for say a mutual fund companies or pension pension money type of things so it'll affect that and it's becoming, you know, it's big today it's becoming even bigger you have middle office applications and middle office applications tend to resolve revolve around risk management compliance type things and AI and ML is being used to monitor all types of risk or market risk liquidity risk it's working to to monitor banking regulations and the banks have so many regulations that they have to deal with so it's becoming very very big there and then back office applications if I again stay with capital markets asset management there's things around you know what's the best way to clear and settle a trade what's the best way to make sure i get best execution price what's the best way what's the best way to make sure that I do the the accounting for it to portfolio accounting for it's a reconciliation I need and reporting so AI machine learning automation this is very heavy investment areas for companies because it makes allows them to do it much better much faster and allows them to reduce costs. Loraine Lawson: You’ve been listening to the Buzz, a Bank Automation News podcast. Thank you for your time, and be sure to visit us at Bank automation news.com for more automation news. You can also follow us on Twitter and LinkedIn. Please don’t hesitate to rate this podcast on your podcast platform of choice.

Listen: Explainable artificial intelligence could be game-changer for banks

Podcast explores how to map AI decisions with Aerospike’s Tarmy

Chase builds out digital and product leadership team, Citi hires new compliance chief

EU moves to ban ‘unacceptable’ uses of AI

Related Posts

Mistral joins Nvidia’s open-model coalition

Ken Griffin says AI ‘almost certainly’ in a bubble right now

Fintech Ramp applies for trust bank charter with US regulator

EU moves to ban ‘unacceptable’ uses of AI

EMERGING FINTECH DIRECTORY

FinAi Podcast

SPONSORED

The Next Generation of Cyberthreats: Powered by AI

Build an Antifragile Strategy to Outperform the Market

How AI and Product Experts Turn Fuzzy Requirements Into Focused Dev-ready Roadmaps

Connect

Welcome Back!

Retrieve your password

Continue Reading with FinAi News Premium - Less than $2/Day

Listen: Explainable artificial intelligence could be game-changer for banks

Podcast explores how to map AI decisions with Aerospike’s Tarmy

Chase builds out digital and product leadership team, Citi hires new compliance chief

EU moves to ban ‘unacceptable’ uses of AI

Related Posts

FinAi Podcast

Connect

Welcome Back!

Retrieve your password

Unlock This Article

Continue Reading with FinAi News Premium - Less than $2/Day