Utilizing Tech - Season 7: AI Data Infrastructure Presented by Solidigm - 3x02: Using AI to Assess Risk with Mike O'Malley of SenecaGlobal

Episode Date: September 14, 2021

Machine learning excels at finding needles in haystacks, even unexpected ones, and this helps organizations to assess risks. In this first episode of season 3, Utilizing AI hosts Stephen Foskett and Chris Grundemann discuss risk analysis with Mike O'Malley of SenecaGlobal. ML is extremely good at detecting outliers and adapting to changing patterns, and this can yield excellent results in applications like financial pattern recognition. But AI lacks real understanding, and this can limit the use cases for ML. Three Questions Are there any jobs that will be completely eliminated by AI in the next five years? What’s the strangest or most amusing application for ML that you have encountered? Can you think of any fields that have not yet been touched by AI? Guests and Hosts Mike O'Malley, SVP at SenecaGlobal. Connect with Mike on LinkedIn.  Chris Grundemann, Gigaom Analyst and Managing Director at Grundemann Technology Solutions. Connect with Chris on ChrisGrundemann.com on Twitter at @ChrisGrundemann. Stephen Foskett, Publisher of Gestalt IT and Organizer of Tech Field Day. Find Stephen’s writing at GestaltIT.com and on Twitter at @SFoskett.       Date: 9/14/2021 Tags: @senecaglobal, @SFoskett, @ChrisGrundemann

I'm Stephen Foskett. I'm Chris Grundemann. And this is the Utilizing AI podcast. Welcome to the Utilizing AI podcast, where we cover enterprise applications of deep learning, machine learning, and other artificial intelligence topics. Today, we're looking at risk and specifically utilizing AI to assess risk. Chris, this, I think, goes to a lot of the things we've heard previously on the podcast from especially security and analysis, log analysis, that kind of thing, right? Definitely. I think it does make a lot of sense. Risk is something where assessing reams of data can help you to pinpoint different ideas and different potential issues. So it does make sense.

We haven't talked about risk specifically yet, but I get it. Yeah. And it seems like, you know, one of the things that machine learning is particularly good at is spotting, well, surprising patterns, unexpected patterns, and unexpected outliers from those patterns. And so that's one reason, you know, for example, that log analysis works so well using machine learning. So joining us today, we've got Michael Malley from Seneca Global. Mike, do you want to introduce yourself here for a second? Sure. I'm the SVP of Marketing and Sales from Seneca Global. We do turnkey product development for all sorts of different SaaS companies utilizing AI. Excellent. Hey, that's the name of the podcast right there.

There you go. So, Mike, did we get the gist here that assessing risk is all about finding the needles that you didn't know were in your haystack? Yes. And so I think the answer is you did get it right. But what AI gives you is it gives you the ability to scale and not look at just single needles in haystacks, but thousands of needles in haystacks, right? And scale it up to a level that humans couldn't do and be able to do that with two additional benefits. One is without any human

error. And the second piece is without the need of human intervention, whether it's at a risk officer or a security officer, that because of the global talent shortage right now, you wouldn't be able to find. And if you could find them, they're already employed. So we talked a little bit kind of warming this up that it makes sense that risk would be something that we're looking at with AI. But I wonder, you know, from a real world experience perspective, Mike, what are you seeing out there as far as, you know, high level view of risk and AI and how that works out together? Is this like anomaly detection or are we getting proactive about seeing risk on the horizon or, you know, what are you seeing? Yeah, so it really runs the gamut. You know, we're certainly seeing it used for anomaly detection. And we work with a lot of

security companies that are looking at basically picking up, you know, thousands of logs or tens of thousands of logs looking for any type of anomaly, trying to determine is there a bad actor. And then if there is a bad actor, putting some type of policy in place where it will say, you know, if I'm worried about Chris, he's doing something I'm not sure he should be doing. Do I want to block that? Do I want to challenge that? Or do I want to let that through? And so I think we're definitely seeing those types of behaviors. And then, as you said, now what we're moving into kind of on the cutting edge of AI is using it to measure intent. And so this is where instead of just looking at anomalies, I'm now starting to look deeper at what, let's say, let's take a web based example. What are the web pages that Chris is looking at in what type of order? And is that

indicative of the fact that he is a normal user? Or does that suggest maybe he's a bad bot and is up to something bad? So they're getting more sophisticated in that realm as well. Interesting. So it sounds like there's multiple domains where this is taking place in. And I mean, you know, almost different use cases, right? Because there's obviously different types of risk as well, right? And you've already kind of talked about a couple, obviously, you know, but there's many, right? There's potentially physical security risk, there's cybersecurity risk, there's business risk, there's more broadly economic risk. And, you know, I think the list goes on and on, right? Climate, climatological or meteorological, easy for me to say. Meteorological risk, yes. And it can be used in any and all of

those domains. I mean, two domains where we see it being used a lot is in security, without question, as we just talked about, but then also in FinTech. We also see it in financial tech a lot for things like risk-based authentication or mobile payments apps or different things around finance where, again, they're looking at anything from anomaly detection to intent to try to determine what is the intent of this financial transaction. Do I want to allow a transfer to the bank of the Nigerian prince? Do I want to allow that when the normal bank that this person uses is Chase in Chicago? And it's interesting when you talk about this stuff, because, of course, we're anthropomorphizing it. It's not like artificial intelligence is asking any of those questions at all. It's, you know, the machine learning algorithm is just miraculously and mysteriously looking at and seeing, does this fit my pattern or does this not fit my pattern. And I think what we've seen so far, at least in my experience, is that it's phenomenally good at finding things, or it can be phenomenally good at finding things, especially in the examples that you're giving. So like assessing the transactions that are going through financial,

you know, like regular credit card transactions, for example. Machine learning is amazing, amazing at determining which ones are the outliers. Even when people are doing things unexpected, like going on a trip or stopping at a gas station they don't normally stop at, it's amazing that it can pick up those things or correctly pass those things and yet reject things. And I don't think we give it enough credit, to be honest with you. I feel like we look at machine learning and we're constantly noticing the failures and saying, oh, of course this was me. I go to Starbucks every day. How come it rejected this one? But then we don't appreciate the fact that it picked up, oh, that one was in Des Moines. I

see why that one was rejected. Or maybe it was just some mysterious thing. Is that what you see, Mike, that it's really good at this stuff? It's really good at this stuff. And the other thing that I would say is we discount the learning part of machine learning and AI, right? These algorithms get better the longer you run them. And so this idea of it not only understanding the initial pattern, but understanding the changing patterns as you drive cross country from Ohio to Denver to Seattle and buy gas all along the way. It gets extremely smart in seeing how the pattern changes over time and then accepting that as a normal pattern and therefore then passing that or allowing that, right? It gets very, very smart there and seeing how the patterns change over time. And I don't think we appreciate how good it is from a learning perspective. And just to play the devil's advocate there a little bit, I mean, one thing that AI doesn't

have, and Stephen kind of pointed this out with the anthropomorphizing statement, is that it doesn't have any common sense. And what I mean is there's a lot of kind of framing or contextualization that doesn't go on. There's a good example of the folks, I think it was OpenAI actually, that built an AI that won at this online gaming tournament, right? This Dota 2 tournament. And people were like, oh man, like that's pretty amazing. But what had happened is the human programmers had set some context beforehand and they said, okay, well, human gamers look at this tournament in three phases. And so it'll actually, they told it to know to switch tactics

at each of those three phases. And there was some other pieces of context that they provided it to let it understand how things changed. And I just wonder, you know, or I guess I guess that that plays out here as well, that there's still a huge human component to setting these systems up and tuning them properly in order to get those results. Absolutely right. I think, you know, when we look at companies and how they want to differentiate themselves, the key piece that people are differentiating themselves is based on the data scientists who write and tune those algorithms. That's really where a lot of the value is these days, because exactly as you said, the algorithm is only as good as it's been tuned and programmed in the first place. And so a lot of that contextual knowledge can be

built in by the data scientists who design it. And then that's just going to make it that more, you know, that more effectively pre-programmed for what it wants. But I think there's two things it can give you that are very valuable. It can not only tell you about risks and the risks in terms of your user behavior, It also can tell you risks in terms of how your application works and functions, right? And so a great example of that, for example, is Instagram, where when Instagram was originally produced, Instagram could share pictures and chat and share music and videos and all sorts of other different functionality. But then they ran the AI algorithms on it to determine from a product standpoint,

what do people actually use it for, right? And the AI algorithm was what told them that from a normal behavior standpoint, what people really wanted to use it for was sharing pictures. And then they really focused the development on that area and got really, really good at that. So I think there's not only an outward component about understanding how your customers use it, but there's also a component about how the application actually runs itself and how do you make it better, understanding yourself, basically. Yeah, I'd like to follow on that too, because I'm hearing a lot of it tells us, it tells us, in what way do regular real world applications of machine learning,

in what way do they tell us? How do they communicate risk? How do they enumerate it? How do they let us know that this is the better direction than that? What does that really look like in the real world? Yeah. And the answer there is a mathematical model. It basically comes out with a number. And the way it does that is it looks at all of these different inputs. The algorithm runs and basically checks all these different parameters against what it thinks is normal. And then based on whether it comes in, how far it comes in, either in or out of quote unquote normal, it will assign a score. And so it looks across then the entire domain across all of the parameters it's looking at. And if enough parameters are out of kilter, then it will say,

well, this is questionable and we should challenge. Or if almost all of the parameters are out of kilter, then it will say this, we clearly want to block and it's something different. So it's a mathematical model that you can actually tune yourself. So this is when we talk about tuning, Chris, and how you want to tune these algorithms. This is where you can tune it to say anything that's even slightly out of normal, I want to challenge. And anything that's beyond this risk threshold, I definitely want to block. And so that's how it works, practically speaking. Yeah, that that makes sense and then it's up to like so again you're kind of setting the table getting the scores back uh so for example in your instagram example right i'm guessing that what they did is said okay here's all the use cases that we

possibly can think of and score them based on the amount of use they're actually getting across the general user population and and photo sharing came out with the highest score and that was that was the signal exactly they find out that all these other use cases that they're working on, the usage on it is incredibly small and the big spike is all on the foreshare. Yeah. It's really interesting how this becomes just kind of this, this lever for folks to be able to work on larger and larger systems. To me, that's,

that's one of the resounding kind of successes of AI is, is, you know, that that's a lot of data that some person or a bunch of interns would have had to go through in the past, it probably might have taken weeks or months or even years. I think it's, I think it's that fact for sure, right? The fact that that you have this massive amount of data, but I wouldn't discount also the idea of human error, right? As much as we love engineers, and they do great stuff, and I'm an engineer myself, right? Inevitably, we do make mistakes, right? There are such things as bugs, I've heard. I've never seen one, but I've heard they exist, right? And so because of software bugs and things like that, what makes AI and machine learning

algorithms so good is, again, like any other piece of code, they can do these repetitive tasks on massive scale, but doing without getting tired, without getting cranky, without any other piece of code, they can do these repetitive tasks on massive scale, but doing without getting tired, without getting cranky, without needing a cup of coffee, and they never make mistakes, right? And so this is where when you talk to, for example, Department of Defense, which I've done about this sort of stuff, they say, listen, we're interested in AI and automation for everything we possibly can. And those things that we can't,

and humans have to do, then we're going to drill, and then we're going to drill, and then we're going to drill some more, because they know that humans make mistakes. So, I mean, that's an interesting aspect here, is the sort of bug aspect, because, I mean, truly, if you're training a machine learning model with a data set, it's really programming itself. And I think that that's the interesting aspect here of machine learning is that it programs itself based on the data set that you feed it. And indeed, that can lead to fewer bugs and fewer errors, because essentially, it is a black box, but it's a black box that has sort of boxed itself in instead of one that somebody actually like might have typoed something. But on the flip side, as we've discussed countless times over the last 50 plus episodes of Utilizing AI, you also have the chance of introducing bias and problems in

your model. So essentially, if I'm feeding my credit card model, for example, only with examples of people with good credit who behave responsibly, then there's actually a chance that it might miss totally risky, crazy behaviors of people who don't behave responsibly, then there's actually a chance that it might miss totally risky, crazy behaviors of people who don't behave responsibly simply because it matches up to what it's assuming is fine. Or maybe, maybe not, maybe the exact opposite. Maybe it'll start flagging things left and right because that are actually totally normal just because it hasn't experienced them before. Or maybe, you know, maybe you've trained it with only men or only people with white-collar jobs or blue-collar jobs or whatever. I mean, there's all sorts of instances where the black box can actually be wrong for the situation.

How do you communicate that with someone from the Pentagon who wants to have an AI drone that flies in and shoots people? The way you do that or the way I've done that in the past is you really focus on what is the core competency and what is the skill set you want to develop? Because as you said, both of those things are true, right? The AI algorithm is as good as the one that's written, but at the same time, it will learn and get better based the best data scientists that you possibly can to train and write the algorithm and make sure that you're feeding in data without some of those biases, for example, that you mentioned, right? And then turning to people like Seneca Global, for example, a lot of the work we do is with startups that are really just focused on the data science part of it. And then we put the entire SaaS product together for them

so they can focus on their core competency, which is training that algorithm, right? And that's kind of a difference from, think of 80s or 90s startups where they wanted to do all the software development and have their hands in all of those different pots. Now people are realizing at the end of the day, the core competency and the

intellectual property isn't the software, it's just the algorithm. And what are those conversations like? I think that's a really interesting thing to me to imagine. So you're in the room with somebody who's trying to develop an AI application, and how much education do you need to do with these folks about the reality of what they can get out of an AI system? I think that there's a fair amount of education that has to be done. And I think that's done in two ways. One, it's by showing them other AI algorithms that we've done and how they work and how they tune and how they get better over time, right? A lot of these security algorithms, you know, you get them up and you run them for a couple of days to a week, and then they learn all the behavior that they need to learn, right? So a

lot of it is by showing. And then the second thing is, as you said, you know, looking at the mathematical model and looking at the data sets you feed it, having smart people doing that. And that's really when you get into the art of the data scientists, looking at what type of data do you want to feed this algorithm to train it and being smart about that. And then teaching companies that these other things around it aren't things necessarily that you need to do everything yourself. Really focus your core competency on where you're going to be best and where you're going to be best and derive the most value right now is going to be with that algorithm. And then what role does other like, like domain expertise play, right?

Cause I get that you need the data scientists to come up with the right algorithm and then you need the data, but it seems like, you know, there's a potential there for runaway. And Stephen kind of talked about some of that. And I think the protections there are a human environment when you're doing the training to make sure that a sane result is actually happening based on fluid dynamics or whatever that area is, maybe the data scientist is not going to be an expert in because they're an expert in data science. That's, that's absolutely true. And that's where domain experience becomes

critically important, right? I've talked to several clients, for example, in nuclear engineering, where they're doing data science algorithms to look at the water coming out in the reactor, because of course the reactor has to run 24 by seven in order to make the nuclear power plant efficient. So you never want to stop the reactor to inspect them. So you're trying to read all of these signs to determine indirectly, how are they doing? Right. And so they got the data scientists to train that, but that's where, you know, you need to work with domain experts as well, who come in that understand nuclear power and power plants and water filtration systems and how they keep the control routes cool to understand is the result that they're getting a sensible result exactly as you said. Yeah, I'd like to follow in on that too, Chris, because it seems to me that that's the real key, because a domain expert is going to know things about the machine learning model that even a data scientist isn't going to understand.

I mean, they're going to know sort of what parameters to include, what parameters to exclude. They're also going to know which areas are going to require a different type of model. In other words, one example, for example, is highway driving versus city driving or credit card transactions from consumers versus credit card transactions from businesses. I imagine that you might need a completely separate ML risk model to assess business credit card transactions where, you know, a $50,000 charge might be normal for a big, you know, factory or something, whereas for a normal person that would be unheard of, you know, and so you wouldn't want the same model. And I would think that domain expertise really would be the most important thing. Are businesses including them at the table when they're planning these applications?

They are. And often cases, the original idea comes from the business owner or the domain owner, because they're the one that recognizes the problem, right? You take the case of the nuclear power plant, right? It's those nuclear engineers that are monitoring the system that understand the problem, understand that you can't turn off the reactor, but also understand that some of the key indicators of health can be leached out of the control rods into the water and you can determine things from the water. So a lot of that domain expertise often comes from the initial entrepreneur or the initial business owner that has the idea in the first place, and then brings in the data scientists to give a context. Interesting. And then, you know, one thing, we kind of started this out by talking about how AI in general can be really good at anomaly detection and some of this predictive analysis and kind of seeing through the noise of data to spot trends that might be concerning.

Another area that I've definitely noticed personally, but also people around me is, people are really bad at spotting absences. Meaning if you've got this kind of set of, a picture with a certain number of flowers in it, and you're shown another one with one less flower, it's very rare that a person actually spots that. If you move one of them or change one of them

or add another one, it's easier for people to detect. And I'm guessing, but I'm not that um this kind of detection of absences might be a really big part of risk detection right and seeing when things go away um which is something that's really really hard for humans to do that's absolutely true i mean as much as we talk about looking for things that are clearly out of normal the other part of this is looking for for things that are clearly out of normal. The other part of this is looking for things that simply aren't there, right? And so that's where, as you said, from a pattern generation standpoint, computers and AI algorithms are very good at looking at those types of patterns, where humans tend to get fixated just on certain parameters that they view well, right? When you

talk about vision and looking at visual cues and things like that, right? Eyes are great motion detectors. And so we're very good at looking at things that change, but we're not very good at seeing static pictures next to each other because there is no motion there. And that's a good example. It also reminds me, and maybe I'm getting a little off track, but there's that old video of you're supposed to be watching people pass this ball back and forth and count how many times the ball moves it and then there's a gorilla that like a man in a gorilla suit that comes into the scene and leaves and they ask you after you've watched the video if you saw the gorilla and most people are like what right yeah which is something that ai would not have missed i think that is absolutely true yes there we go that's the the takeaway from this conversation ai will spot the gorilla.

So no, but it is true. And I think that that's sort of the counterpoint to this bias situation. So on the one hand, we don't want to programatize and institutionalize our biases based on the data that we feed into the machine. On the other hand, a machine is not necessarily going to replicate some of the biases that we don't pass into it. And I think that that's actually a hopeful and a positive thought about machine learning in that, you know, if you train a system without any biographical information to approve loan applications, you know, it's not going to make the same mistakes that humans may make when assessing those loan applications and saying, oh, I know that neighborhood, we're not approving this one. Or, you know, oh, that name sounds sketchy to me, I'm not going to approve this loan. You know, machine learning would never make that mistake if you properly prepare the data on which it was trained. So I really appreciate that. We do have to move on now to our next segment. We're bringing back three questions for the third season of Utilizing AI,

but with a twist, and that is that we are going to also be inviting questions from our guests and from the audience for future episodes. Since Mike is the first guest here on season three, I think Chris and I are going to have to do all the questions ourselves. But hopefully Mike will have a question for a future guest that you can see in a future episode. So Mike, are you ready for this? I'm ready. All right. So Chris, I posed a few questions in the chat for you. You can pick one or two. I'll kick it off by just asking the first one myself. We talked about jobs and tasks. Mike, are there any jobs that will be completely eliminated by AI in the next five years?

People just won't have that job anymore because the computer's got it now. So I think one of the jobs we could look at, for example, would be accounting and financial closure, right? There are a lot of accounting jobs around closing ledgers at the end of the quarter that could be done via AI, could be done via AI, it could be done via AI at very large scale to be able to look at any anomalies right in the books and close that, especially since it's almost entirely rules-based. So that would be an area I would look at. So Mike, we talked a little bit about this

in the green room beforehand, but I wanted to ask on air, what is the strangest or most amusing application of machine learning that you personally have encountered? So the most interesting one that I've done, because I live here in the Midwest, is with regard to making better milk and cheese. And that is using AI algorithms to work with dairy cows and dairy farms. Because at this point, there's a lot of work being done to look at the health of the cow, look at the nutrition being given to the cow to maximize the milk output and the fat content

in the milk, which makes better milk and better cheese and wonderfully better ice cream. And using those factors, then using AI to determine when the cow should be milked, how often, and also then throughout the day and throughout the year, how to change the food that it's being given in order to promote the maximum amount of output. So that's the most interesting, I would say, and fun one that I've been working on. Well, thanks for that. Okay, one more question. Can you think of any fields that have not been at all touched by AI. So I would have a hard time coming up with something that's not, if it's not commercially available, judging from the startup community right now, I can guarantee you there's somebody working on it. Well, thank

you very much for playing along with our little game, Mike. And I can't wait to hear the questions you have for the next guests on our podcast. If you're a listener and you'd like to contribute a question for our three questions segment, please do find us on Twitter at utilizing underscore AI or send us an email at host at utilizing dash AI.com. Mike, where can we connect with you or find more about your thoughts on AI and other enterprise topics? You can just find me at Seneca Global, the Seneca Global website, or you can look for me on LinkedIn. How about you, Chris? What's new in your world? Well, everything that's new for me is pretty much on my website, chrisgrundeman.com. You can also connect with me on Twitter at Chris Grundeman,

or find me on LinkedIn as well. I like having conversations there. Anything to do with enterprise technology or networking or cybersecurity is fair game. Well, thanks a lot. And as for me, Stephen Foskett, I'm at S Foskett on most social media networks. And listeners of Utilizing AI may be particularly interested in the August 25th episode of the Gestalt IT Rundown. That's our weekly tech news show. Hot Chips was just behind us. And in the August 25th episode, I discuss some of the newest AI computing platforms and chips and some of the news from Hot Chips. So I guarantee that that would be interesting

to our audience. So thank you for listening to Utilizing AI. If you enjoyed this discussion, please do let us know. We love your feedback. And we're so thrilled that you're joining us here for season three. Also, when you're in your favorite podcast application, click subscribe. It's always free. And also maybe leave us a review or a rating. That really helps other people find us. This podcast is brought to you by gestaltit.com, your home for IT coverage from across the enterprise.

For show notes and more information, go to utilizing-ai.com. Thanks for joining us and we'll see you next week.