Kevin Grant (IoT Community): Good morning. Good afternoon. Good evening, ladies and gentlemen. Thank you very much for joining us here at this IoT gravity mastermind interview, our second in the series that we’re bringing to you here today. We’re absolutely thrilled to have the legend in Dr. Kirk Borne back with us to host what’s now going to become his customary Mastermind feature with the IoT Community. It is my absolute pleasure and honour to introduce to you today, Dr Gul Ege from Sas. One of our IoT Community Diamond members.
I am going to give her a quick introduction. By way of her background, Dr. Gul Ege is the Senior Global Director for the IoT Research and Development team. She’s leading the advanced analytical components division within SAS. She’s helping customers and stakeholders solve real world problems. Dr Gul Ege has held a number of positions and a lot of different functions over 34 years’ worth of a career. So, they don’t come much more experienced than this, and therefore very fitting of the notion of a mastermind. Without any further ado Dr Kirk Borne I turn the floor and the stage over to you.
Kirk Borne: This is great, thank you very much Kevin. It’s really great to be here today and it’s really great to meet with you Gul face to face. I’ve seen you many times on our Advisory Board calls here. I look forward to having this conversation with you and I look forward to the trip to Cary this summer for those who are going to the IoT Live Slam. And if you are not going there or if you are watching this after the event, watch those too.
We’re just going to jump right into it. First, we’re just going to start with a little bit of background. Just so people can understand where this brilliant person comes from. So, tell us a little bit about your non-work-related background. Where you’re from, where you grew up. What kind of things interested you in school. That kind of thing.
Gul Ege: certainly. I’m from the beautiful city of Istanbul in Turkey. I was born and raised there. I love that city and I have a home away from home in North Carolina for the last I don’t know, 34 years or so. I had a very happy childhood. I went to an American boarding school in Istanbul and had the best of times. I have an older sister but all my other cousins, younger than me or my age are boys. My sister was lady like while I was a tomboy growing up, not to the pleasure of my mom. But anyway, she was always very supportive, always told me there’s nothing you cannot do if you set your mind to it. So that’s always an encouragement I carry around.
Right now, I live in Cary, North Carolina, where the SAS Headquarter is. I am married and have 3 wonderful adult kids and 2 toddler grandchildren right now, and I really thoroughly enjoyed parenting, but grandparenting is more fun, less responsibility. It’s been a great joy in the last few years. That’s kind of who I am. In terms of what I was interested in, just about everything. Very curious about many things when I was in elementary school there were Apollo Missions and I got very much into reading the newspaper and was sure I will turn out to be an astronaut. But in middle school/ high school there’s a lot of sports activities; basketball, tennis, some track running and volleyball for the school team. All the sciences plus math are probably my favourites. Then I went to an engineering school and got all my degrees in Industrial Engineering. My PhD is also on Industrial Engineering and with a minor in Computer Science focusing on Decision Sciences. That’s where the passion to solve business problems started, and hopefully the knowledge came with it. But I learned most of what I learned in Sas.
They hired the cream of the crop. So, one day I finally say, Okay, I stopped by, and they offered me a job. I didn’t even have a resume yet. And although I accepted the job, it was like marrying your first boyfriend or girlfriend. I’m like, I didn’t check anything else – Gul Ege
When I was finishing, getting close to the end of my PhD one of my advisors in EPA. Who, sponsored my research, said you need to supply Sas. They hired the cream of the crop. So, one day I finally say, Okay, I stopped by, and they offered me a job. I didn’t even have a resume yet. And although I accepted the job, it was like marrying your first boyfriend or girlfriend. I’m like, I didn’t check anything else, but I thought I’d stay a couple of years and see how it goes. It’s been a wonderful experience there for 34 years and going, keep going. Yes, we are with the mission of our CEO keeps saying, bring your toughest problem to us and we’ll solve it. And that motivates me. That I think is my passion as well.
Kirk Borne: That’s fantastic story there, I think we have a lot of parallels, except 2 major differences. My degrees are in Astrophysics and Physics and the second major difference, I had absolutely no capability to do anything in sports. I love your passion; I love your curiosity. You know that questioning that scientific attitude and being really thrilled with solving tough problems. I mean, I always felt that was the joy of doing this. It wasn’t the easy problems that were fun. It was the hard problem is that. And I think that I guess that makes us sort of strange compared to most people. But that’s how things get done in the world. My next question was just to ask you a little-known interesting fact about you, but you’ve told us so much already. Is there anything left you could you surprise us with?
Gul Ege: Yes, it’s hard to identify interesting things about yourself because I’ve had myself my whole life. I find many interesting things about other people, but I guess the unknown fact is it’s probably my passion for soccer. I’m an avid soccer fan. I guess since I was a kid watching the Turkish League. I have of course a favourite team, a European League, World Cup, Champions League team. I still like the World Cup though sometimes it shows games in very odd hours. I’m still up, I’m watching.
Kirk Borne: Have you dug into the analytics of all that?
Gul Ege: Yeah, I think the problem with being an engineer you think of yourself as a problem solver. The first time I visited Epcot the whole time I was thinking I wanted to know how they did that. How did they program that? And you kind of ruined the fun of like, just let it be, you know you don’t have to know how they did it. Just enjoy it but once an engineer always an engineer, I guess.
Kirk Borne: Yeah, that’s a good way of thinking about it. I was recently in Las Vegas, and my wife and I saw a magic show and they always warn you in advance. Don’t try to figure it out because, of course, there is a trick. so don’t try to figure out. Just enjoy the show, just enjoy the wonder of it. And not only that but just the wonder of the people who can do those things. I mean that. And I think that’s what you’re describing. When you see a beautiful engineering accomplishment or some scientific discoveries. Someone being able to do that it is inspiring to the rest of us. I really love your attitude towards that. Just walk through life and enjoy it. Don’t try to figure everything out. You have a 9 to 5 job to figure things out the rest of the time just enjoy the world.
So today, we’re really going to be talking about Internet of things, IoT. But before we get directly into that with the things you were doing at Sas that, you know, sort of led you to IoT or that you’re doing before your involvement with IoT. Think about that time before the Internet of things became sort of like the thing. The thing that brought you to this Advisory Board. Think about those times before that. What kind of things were you doing?
Gul Ege: When I first came to Sas I did a lot of the financial analytics kind of stuff. Then I worked for quite a few years in our forecasting products and financial risks, products, initial release. Then I had the wonderful team creating the price optimization and size optimization solutions for the retail industry.
Then I oversaw our manufacturing solutions. I think at that point in being involved in the manufacturing solutions some of the IoT data started coming our way for predictive maintenance and that’s where I think 8 or 9 years ago where the passion, the oh my God! This is going to be good! There will be so many problems in this area because many of the domains are well thought of. They have methodologies. But most of it is before this kind of data was available. So, the data availability at this grain both creates the ability to create new business value but also creates new challenges for analytics.
I like to refer to IoT as sort of insight as a service or forecasting as a service because you get these signals from these sensors that can tell you not only what is going on, so you get the insights, but also if something starts going wrong. You get that warning sort of early because you have a continuous flow of sensor analysis happening – Kirk Borne
Kirk Borne: that to me, that’s just the sort of a very fluid transition from what you were doing before the IoT. I like to refer to IoT as sort of insight as a service or forecasting as a service because you get these signals from these sensors that can tell you not only what is going on, so you get the insights, but also if something starts going wrong. You get that warning sort of early because you have a continuous flow of sensor analysis happening. I think that is what you were doing like in terms of risk analysis, financial analysis, manufacturing, even supply chain, even cybersecurity. Of course, we could talk about the applications of IoT but your story that’s great because it really was just almost like a smooth transition.
Gul Ege: Actually, it really was. I think there was a moment looking at some of the sensor data, because of the many classical statistical procedures, and all assume like normality or some distribution and most of the distributions are not known.
Some of the assumptions of those doesn’t carry over. So that’s why we in a way to expand. I mean, Sas ethical capabilities are vast anyway but for this kind of data these kinds of problems we have a really innovated new methodologies to answer the new questions with this data.
Kirk Borne: You know what? You have a very strong statistical background, and I am going to ask a technical question, hopefully we will grasp its significance. With my background in astrophysics, I was really not drawn to sort of traditional statistics but more non-parametric statistics, knowing that data didn’t have any kind of normal distribution or that type of data and sensor data is like that.
I mean, I think you just want to identify sort of the structure of the data. And when is the structure changing. And how is it different from the structure of some other data. You could characterize your work that way that it’s really looking more now at the non-parametric, that kind of statistics.
Gul Ege: Definitely, the data doesn’t fit to typical data characteristics. Previously it’s multi-modal like your car has shifts, so it won’t behave the same way. So, when you’re looking at what is anomaly perhaps your method needs to capture multi model normals to be able to see what you’re doing. When you’re in fifth gear it is very different than when you’re in neutral gear and the sensor data should be different. Plus, I think, technology wise, we see temperature sensors everywhere. Whether temperature is important or not to the problem being solved, I think it was the earliest ones available. The data is created at the speed that the sensor can create. Yet probably you don’t need sub-second information in most use cases for data. When we look at more digital signals, audio or computer vision, image data, video data also things; we have created a lot in those areas on how to analyze how to keep only the information that’s relevant in the frequency it is needed. Sometimes a straight aggregation of data to 5 min, 10 min might cause you to totally miss what you’re looking for.
But for example, ECG or EKG kind of data, all we need is really the peaks from that information not the rest of it to get something like heart rate variability in healthcare. So many of these methods we’ve identified and also a lot of principal components analysis we have robust principal components because there’s so many variables. And for the question you’re trying to solve, which ones really matter? Because that’s the mystery and I think involving the domain experts in that process for their knowledge is key to success.
We are seeing more people making that journey on the data itself. Because if you don’t get the data corrected and cleaned of the noise and if needed, be reduced, you could be filling clouds without really answering the question you’re looking for. So, we provide a lot of intelligence to our partner customers in that initial exploration.
And so all of a sudden, the file size of my video shrunk enormously because I just saved the changes, not the actual every bit at every second and I think that’s the key dealing with these massive flows, that supply chain and manufacturing and cyber security, and all these other applications – Kirk Borne
Kirk Borne: I think that’s a really key point for people to realize that you do have this high-density high time resolution stream of data. But you don’t need to save every bit. I learned years ago actually in my early days of my astronomy research. I did colliding galaxy research and I made videos, very rudimentary simplistic videos, of these colliding galaxy. The files I created were huge because I saved every bit on the screen but what I realized, what I really need to do is just save the changes from one screen. And so all of a sudden, the file size of my video shrunk enormously because I just saved the changes, not the actual every bit at every second and I think that’s the key dealing with these massive flows, that supply chain and manufacturing and cyber security, and all these other applications. Your EKG example, is a perfect example, because you want to know where the peaks are and you know that sort of in time and things like that is so it’s not like everything that’s happening between the peaks, is there? But it doesn’t mean you need to keep it. I think conversely, if you measured something like your heart rate every 10 minutes that’s not telling you anything.
Gul Ege: Intelligence and data transformation and selection is probably the way we should think about. I love your galaxy example I’d love to see your videos as I was going to be an astronaut. The robust PCA I mentioned is exactly for that purpose on image data to separate the background from the foreground. Whether it is the things that change or whether they are the noise. The background is that you don’t need really most problems, whether we are checking for restricted, you know, if anybody entered this restricted space. The background is the same. So, there’s no reason to save all that in a cloud. I think it really helps the benefits cost ratio when you can make intelligence choices to reduce the data that you need to transfer to the cloud and keep them. We of course provide a lot of this on our event stream processing engine, so it can be done on the edge without too much cost. So only the things that we need will make it to the cloud, or the further and analytics on ESP itself.
Kirk Borne: So, you’re really describing there are a lot of different requirements for these applications and with your space background I think you’ll love this story. With my years at NASA, one of the projects that NASA funded, or I should say they had a program which they funded projects. And the program is called intelligent data understanding the Id. I really love that intelligent data understanding. The genesis of the project was for basically the Mars Rover missions right? Because the bandwidth of transmitting signals in from deep space back to Earth is very small. It’s not like you have a 5G network in the solar system. You can only send back a certain amount of data like when they did the Pluto fly by a couple of years ago. The fly by lasted an hour right, the camera took all these pictures, and it took like 10 months to send all the data back to Earth. This intelligent data understanding was dealing with streaming data and how do you identify the most salient, most important, most significant signals in that data and transmit that. Whether you save all the other bits is another question, right? Whether you have storage capacity to save all the bits. That’s a different question. But what you really need is you need the information rich packets; those are the ones you want to pay attention to.
I think our proof of concept starts with that, the data, the problem we are going to solve and the domain knowledge. That’s where we do the initial exploration with some sample data so we can kind of advice. I like the term you use, what of this data do we really need to do further analytics – Gul Ege
Gul Ege: I think our proof of concept starts with that, the data, the problem we are going to solve and the domain knowledge. That’s where we do the initial exploration with some sample data so we can kind of advice. I like the term you use, what of this data do we really need to do further analytics.
Kirk Borne: So, is there any specific domains where you’re working right now? I mean, are you working any specific domains, or you still like overseeing a lot of different things.
Gul Ege: We get partners and customers in many domains, actually, we look into floods in foundation models for different cities. Predictive, not like, oh, it’s flooded. Yeah, and actually, that brings in some of the physics concepts also into the model right because my house sits on top of a hill so if my house is flooded, all of the city is flooded.
That’s an exciting problem we also look at like utility maintenance and utility industry. A lot of manufacturing of course both for predictive maintenance of the manufacturing equipment because once they go down it is a big issue. Over maintenance is costly and we look at those quality you know, statistical quality control exists forever. Probably we all studied it. But with this data, how to do that where there are multiple variables in the equation of quality control.
How to maybe improve, so that if you can catch a defect happening on the products, you can always take it out and not invest more production time on it. We also have the health area that I mentioned any industry with energy. Whether it’s solar farms, oil and gas or power electric power. Those all have very expensive equipment and how and when to maintain them based on predictions is very important. We’re looking at manufacturing itself. I think the domain was brick manufacturing. Can we optimize?
So, I think we are at climbing in IoT the cycle of analytics. You first need to figure out what is happening. Then you try to predict what will happen in the reasonable future or relevant future for the problem you’re solving – Gul Ege
So, I think we are at climbing in IoT the cycle of analytics. You first need to figure out what is happening. Then you try to predict what will happen in the reasonable future or relevant future for the problem you’re solving. And you either automate on the streaming data, real time, the alerts or decisioning that you need to do or also impact the business value at the end. That’s our full goal. We are climbing. That’s an analytical chain all the way to optimization. I’m very pleased about that. I think in the last 8, 9 years, that I’ve been focusing on this with my team. Things have evolved a lot, 8, 9 years ago we would get request like we collected data we never looked at it can you look at it. And I think we’ve come a long way in terms of even what is an interest in the field, and that’s where we learn where to take both analytical methodologies. I mean, that’s the best way we get our requirements really. That’s the exciting part that creates the excitement.
Kirk Borne: The thing I like about what you just described is the sort of domain specific nature of the predictive modelling. Even the optimization model. In some sense, I sometimes put down time series forecasting when I talk with people that don’t do that because there’s no domain knowledge embedded in just analyzing traditional times and doing traditional time series analysis. Because what is that? It’s basically auto regressive modelling. So, what was the pattern in the data in the past like you think of normal time streams analysis. Is there’s a trend? And there’s a seasonal term, right? And so, you just assume what happened in the past will happen tomorrow. But if your engine never failed in the past, are you saying, my engine will never fail in the future? That’s not realistic.
I don’t think hundreds of years of physics knowledge should be gone down the drain. Takes deep learning problem without understanding the domain itself, deep learning methodology, it’s very appropriate sometimes – Gul Ege
Gul Ege: that’s not realistic in this domain. I mean, I work on time series forecasting and there are many problems in this domain, you know, predictive maintenance purposes like you said. All of these very expensive machineries are very reliable. So even in the past, you might not see any fault when you look at the data. But the point is, how do we use this new data to be able to predict that the performance is degrading. If we think of control limits. It’s getting out of the control limits often and in multi-variate. This is where for example, we developed new methodology because most of the previous quality control is univariate. But you can then take the domain information. Do we want to alert them every time, every second this happens, or do we want to watch if it continues 7 minutes, for example, or 5 mins? But this information will come from the domain, I think, including domain knowledge and domain practice, into the analytics and feature creation. Whether we are future selection is critical to success. It really is. I don’t think hundreds of years of physics knowledge should be gone down the drain. Takes deep learning problem without understanding the domain itself, deep learning methodology, it’s very appropriate sometimes. But my point is, you need to understand the data, the problem, and the domain. And you know I’m an industrial engineer and I think many people in manufacturing quality control would think we have to know how to do this. But it was all prior to this data availability and they can help you shine the insights that they’re getting from this data and improve the business.
Kirk Borne: That story is great because it makes me laugh when I think about something that I was taught years ago as a data analysis and an astronomer, and it makes me almost cry when I think about it, and that is we had this practice it was called 3 Sigma clipping. That is basically any data point which is more than 3 standard deviations from the mean you just assume it’s noise and you just clip it out of the data and then you move on with the analysis. I think to myself that’s maybe where the real discovery was the thing that was anomalous. It was not just noise; it was actually a real discovery. It was a real event, and it just makes me cry to think how many things I flipped out of my data, delete from my data stream because I was taught that way I said, No, no, no, that’s not anomaly detection. I always tell my students, I said, you need to ask yourself 3 questions when you see an anomaly or an outlier in your data set. The first 2 questions can come in either order, but the third question must always come last. So, what are the questions? Well, the first 2, again, you can interchange them. It’s for you ask yourself, is it a data measurement problem? Okay, it’s sensor, a data collection problem. Second question, is it a data processing problem like a software problem? And then if it’s neither of those things, then you can ask yourself the third question, which is my Nobel Prize in waiting. Don’t quickly assume that you have a new scientific discovery. And also, the converse is also true. Don’t just throw the data away. Assuming it’s a data collection problem or a data processing problem because it might actually be a real alert.
Gul Ege: Yes, yes, I think those should be in that exploratory study visited as a discussion with the domain experts to help identify that, especially in real time on a streaming data you don’t want to over alert.
It’s like your alarm kind of snoozing for only 3 mins and coming back to irritate you. So over alerting is a huge problem. And you don’t want to be ignored when it is, you know, really somebody needs to attend to this. You know we can’t under alert, especially if it’s a critical equipment because the expectation is they’ll never fail. And that would come sometimes. So, there’s no way to solve that without domain knowledge from my knowledge and a methodology really to make sure we are not over alerting.
I always tell people that in machine learning I call it the sentinel on your data. And of course, what is sentinel? In traditional language it’s the guard on the guard post, right on the fort who’s watching to see what’s happening outside the fort – Kirk Borne
Kirk Borne: Alarm fatigue. Yeah, the expression is alarm fatigue. I always tell people that in machine learning I call it the sentinel on your data. And of course, what is sentinel? In traditional language it’s the guard on the guard post, right on the fort who’s watching to see what’s happening outside the fort. So, is our adversary attacking us, or what? You don’t want the guard on the guard post to give a lot of false alarms. Then on the other hand you don’t want to have the guard all night long saying, everything’s good, everything’s good. If it just keeps shouting that every 2 s, I mean it’s like that’s going to really get people irritated. I think what you need to know, what’s important to alert on and you need to tolerate some false positives some false negatives, but only in the right ratio, that is, if a false negative might be the costliest thing. If so, if you say the machine is not going to fail and it fails that’s probably a worse scenario than if you say it is, but it doesn’t.
I think domain and respecting the data and choosing the methodology that will not error in either side too much otherwise it wouldn’t be such a complex problem – Gul Ege
Gul Ege: mostly that is true but there is a cost to over alerting as well. Windmills, they’re very expensive to maintain even getting the equipment to them, they’re pretty high up. So, I think that’s where something like support vector data description that we have a multi model normals like in multi-dimensional space. What is normal is all we can model. First because that’s most of the history and unfortunately sometimes the data doesn’t come with maintenance history or failures. And failures don’t happen often, I think usually they over maintain them. But that’s when we have the normal so well-defined things falling out of it, or the border of that sphere, if you want to think about it, that way of normal operation under, for example, wind condition. If there’s no wind you don’t need to be upset that the windmill is not creating too much energy. So yeah, I think domain and respecting the data and choosing the methodology that will not error in either side too much otherwise it wouldn’t be such a complex problem.
Kirk Borne: That’s really good explanation. Now you see you’ve inspired us with a lot of stories. Here I was wondering if there was any sort of emerging innovation or opportunities that you see in the IoT space, something that excites you. You don’t need to give us any Sas secrets.
Gul Ege: I am very pleased that we are also getting a lot of requests like the flood etc., in terms of climate change and more reducing carbon footprint, for example, the optimizing manufacturing or products that would minimize the carbon footprints; so, production optimization in a way. So those environmental studies as more people are willing to put the infrastructure in place. Sometimes the request that comes to us that we tell them where to put the sensors you know. If you put sensors in a river, 10 inches apart, there’s no lag between them whatever. I think that has been a new turn of events. New request coming from more environmental things, city planning things, infrastructure planning. I think the infrastructure bill has helped a lot in this demand area and also getting into more optimization. I mean, areas of helping decisioning not only prediction. It’s exciting. It’s kind of completes the analytical cycle and the more we did it the feeling is even better.
I like to use the word observability in terms of IoT, and I tell people the difference between the monitoring and observability in the IoT space is this, monitoring is what you do, and observability is why you do it – Kirk Borne
Kirk Borne: Excellent, I think what you just described is something that I refer to as observability strategy. And I mentioned the word observability carefully here. In the sort of IT security area that they use that term to as in terms of network observability, that is, sort of an observed network traffic to be aware of. Maybe cyber security incidents and that kind of thing. But I like to use the word observability in terms of IoT, and I tell people the difference between the monitoring and observability in the IoT space is this, monitoring is what you do, and observability is why you do it. And so, an observability strategy comes down to those things you’ve just mentioned. Like where do we place the sensors? What frequency do we collect data from the sensors, whether it’s spatial frequency or time frequency. Why are we doing it? What do we intend to learn from it? What decisions and actions are we planning to make as a consequence of deploying these sensors? So, businesses need to have this discussion. You have to have the dialogue like, why do we need this data? Why do we need to put sensors here?
It’s like having surveillance cameras in the city. If you have a surveillance camera every 5 feet. You’re not really capturing that much new. But if you have a surveillance camera every 5 kilometers that’s not very helpful either. So somewhere between those 2 is the right answer and so you have to have that conversation and there’s no single right answer to this. Obviously, we go back to where we started. It’s all domain.
you need to have a strategy, and I call that the observability strategy, which is the why. Why are we doing this? Why are we collecting this data? Why are we collecting it here at this frequency. What is the question we’re trying to answer, what business problem are we trying to solve? – Kirk Borne
I really love your description of this because it’s something I feel passionate about is don’t just collect data. And don’t just monitor. In other words, just for the sake of monitoring, you need to have a strategy, and I call that the observability strategy, which is the why. Why are we doing this? Why are we collecting this data? Why are we collecting it here at this frequency. What is the question we’re trying to answer, what business problem are we trying to solve?
Never lose the why – Gul Ege
Gul Ege: Never lose the why, I tell our interns or new colleagues. You have to understand the problem first, then come up with a solution. Sometimes people come with here’s my data can you do some deep learning on this? I hesitate. It’s just like, okay what is the context of the data? And what is the value you’re trying to generate? Because like in medicine if your headache is going to go away with Tylenol you don’t need to do a cat scan or whatever else the doctor wants and take heavy duty morphine or something to get rid of your headache.
One shouldn’t start with analytical methodology which I mean, I guess maybe we are spoiled at Sas because we have them all but understanding the problem and the domain and the data is very important before you choose or experiment with different methods to see which one will give the best answer but you know sometimes excitement over one analytical methods or another kind of peaks into the markets. It’s not, I have a hammer everything is a nail kind of approach, at least in Sas we don’t have to do that we can pick the right method.
Sometimes excitement over one analytical methods or another kind of peaks into the markets. It’s not, I have a hammer everything is a nail kind of approach, at least in Sas we don’t have to do that we can pick the right method – Gul Ege
Kirk Borne: On the bookshelf behind me I have my data mining hammer from years ago. Many years ago, when I was still at NASA, I just discovered machine learning and data mining and to me everything was a data mining problem. It became so persistent that a friend gave me this toy hammer, this plastic hammer, and he put a little label on. Kirk’s data mining hammer because to a child with the hammer all the world is a nail.
One of the topics for our AICoE this upcoming June event is explainable AI because if you don’t get domain acceptance on both the predictions and the decision it’s not going to succeed – Gul Ege
Gul Ege: And it’s very capable. Truly some of these methods I mentioned is going to our visual machine learning AI products. I know exactly what you mean because the excitement overcomes. But I think just take a step back. If Tylenol is going to fix it. If a simple solution works don’t complicate it. One of the topics for our AICoE this upcoming June event is explainable AI because if you don’t get domain acceptance on both the predictions and the decision it’s not going to succeed.
Kirk Borne: The data scientist just has to do more than just say, Trust me.
Gul Ege: I think actually Sas does a lot of work on the explainable process. There are regulated industries where it’s not just you need to explain to the end user, but they need to turn around and explain to their regulators why they’re making these decisions based on what. The efforts on that catching bias on decisions. I mean, take insurance, for example, we have recommended engines that could be used in multiple places. But you know the marketing intelligence. What should your news websites recommend to you? This should be dependent on you and whatever advertising they’re getting from their donors, right? so I think that there are other angles to look at this problem of explainable AI. Plus, before you put anything in operation operationalizes real time you need to gain the confidence of course and you need to revisit these models every once in a while. So tracking, I’m taking you in many directions.
Kirk Borne: So, could you give an IoT specific example of this explainable AI.
Gul Ege: This is an example I’m just coming up with it; how safe a driver is. In terms of maybe setting up insurance rates or whatever. Something like that you need to be able to explain this to even the person. I don’t see you stopping at stop lights. You’re like making very sharp turns or going over the speed limit, so much so that they can correct their behavior if they want a more competitive price on their insurance. But not only that, we need to make sure it’s not biased against gender or race. It applies many problems like that is we need to be able to explain. You can’t say you’re a teenage boy and I’ll forever punish you for that. It doesn’t work right? So, some of these things come in IoT through keeping track of our family’s insurance. If you want to put something in your car that they can keep track of your driving, they can offer based on that better insurance rates for you. Oh, that’s one example. There’s probably many more.
Medicine is huge and let’s say, usually, typically now once a year you go to your physical check-up. Get the EKG or whatever similar things, while you’re resting in a bed. But if we want to track it, walking around doing sports and whatnot. For some reason, so like remote monitoring of patients you better be certain of your model. You won’t over alert the doctor or the patients. We did the asthma project with the UNC Medical school, especially in paediatrics with Dr. Hernandez, that helped us, and she is also in our AI center of excellence in IoT community. Usually, patients are not able to detect because the gradual change they adapt. They can’t say I’m getting into an asthma attack. So, we are monitoring their breathing and we can alert that maybe it’s time to take their medicine. I mean you don’t want everybody rushing in the worst-case scenario to the emergency room or worst case is actually dying from an asthma attack where we can detect it. But again, you need to be able to explain that to the doctor and the patient. What’s going on. That’s the adoption.
Kirk Borne: Well, I think this has been a really great story. Maybe we even had more questions, but I think we should be wrapping up pretty quick here. one question I had, which I think you’ve already answered, but you can elaborate a little more if you wish is your involvement in IoT community. I mean, you’ve mentioned the AI center of Excellence AIoT. I guess that’s called AIoT center of excellence.
You are the chair of this CoE.
I guess the wisdom here is probably the call, learn from other people’s mistakes, because you don’t have enough time – Gul Ege
Gul Ege: I am and I’m a blessed with contributors from different domains. These are all practitioners from manufacturing; Ford is present with preventive maintenance, Otis Elevators and the wonderful colleague Pooja and Amazon (Ivan) from Amazon on logistics. We have Dr. Hernandez for medical domain, and we have an energy domain. That’s currently Keith Holdaway from Sas will be leading, the previous leader, Duncan he got a new role in the company, so he is going to transition out. I learned a lot from them. And I’m hoping the community learns a lot from their experience. These are again practitioners out there of analytics and they are addressing IoT problems. It’s not like data comes in you throw a pixie dust on it and make millions. It’s good to know the challenges. I guess the wisdom here is probably the call, learn from other people’s mistakes, because you don’t have enough time. I love that. So, I think it applies in this area which is maturing very quickly, of course. And although the domain might not be the same as yours. What they’re sharing can resonate with everybody in any role participating in these IoT projects.
Kirk Borne: Excellent, I should join your committee because I’ve made a lot of mistakes. That’s the requirement to be on your panel. I think we should be wrapping up. But before we wrap up one last question is, you have to offer any words of wisdom advice, a favorite quote that inspires you or anything of that?
Learn from in personal life and otherwise in IoT learn from other people’s mistakes. So, you make new mistakes, not theirs, and start maybe a couple of steps ahead – Gul Ege
Gul Ege: You and I shared a lot of wisdom. I think one is make sure you include domain from step one. Make sure you question the data before you get into prediction. And the last one, I said, was just the quote; learn from in personal life and otherwise in IoT learn from other people’s mistakes. So, you make new mistakes, not theirs, and start maybe a couple of steps ahead. And I think that’s the goal of The IoT center of excellence is to share these wisdoms to get your business value faster.
Kirk Borne: Fantastic. I think we’re going to change IoT to the Internet of wisdom. So, thank you very much. Thank you very much. Gul today for this wonderful sharing, wonderful stories. Sharing your knowledge with us today. So, it’s been great. Thank you. Thank you. I loved it. I enjoyed it.
Gul Ege: Thank you.
End
# # #
