Interview with Dr. K.P. (Suba) Subbalakshmi

Dr. K.P. (Suba) Subbalakshmi is the Founding Director of the Stevens Institute for Artificial Intelligence and a Professor in the Department of Electrical and Computer Engineering at Stevens. She is a Fellow of the National Academy of Inventors and a Member of the National Academy of Science Engineering and Medicines’ Intelligence Science and Technology Experts Group. She has also co-founded several technology start-up companies.

In this interviews, Dr. Suba talks about the applications of AI in mental health and fraud detection, limitations of AI, and how technical and nontechnical professionals can prepare for an AI-fueled world.

Transcript :

Wei Zheng: Alright, let’s get started. Hello everyone, welcome to the first event of the Women Leaders in Technology Interview Series presented by the Stevens Leadership Portal. This Leadership Portal is a new initiative at our School of Business at Stevens to connect people to lead our stories, cutting edge research, and learning communities through a web-based knowledge portal. Our goal is to have the web portal to go live this April. You can see it in my background; that’s our experimental site.

Our live interview series this semester will feature six top-level women leaders in technology and their interview dates will be shown on the screen at the end of our interview today and we will share more information with you through email.

Today we have the great pleasure to have Dr. Suba Subalakkshmi as our guest. Dr. Suba is the founding director of the Stevens Institute for Artificial Intelligence and is a professor in the Department of Electrical and Computer Engineering at Stevens. She’s also a fellow of the National Academy of Inventors and a member of the National Academy of Science, Engineering and Medicines Intelligence Science and Technology Expert Group. She has also co-founded several technology startup companies.

So thank you so much Dr. Suba for making time for us in your busy schedule. Before we get started, a lot of you are still coming in. Welcome to all of you here. Because we have a lot of participants today, we anticipate about 70 to 80 of you, so for the first part of our interview, you’ll be on mute, but I have collected your questions from the registration form and I’ll be asking Dr. Suba directly. In the second part of today’s event we will have an open Q&A session where you can raise your hand and Yash, our assistant today, will be unmuting you and you can ask your questions directly to Dr. Suba, but you’re welcome to type in your questions in the chat box throughout the event. So without further ado, Dr. Suba, welcome! Could you start by telling us how you became interested in technology in general and eventually what got you interested in AI?

Dr. Suba: Thanks, Wei for this opportunity. It is a pleasure talking to you and everybody else. Would you believe it if I said that my first serious goal was to be a poet? I guess that was my first goal in my childhood. I was always curious about how things worked. My first degree was really physics because to me, it was literally the essence of how the world works. So I was very curious about that and that’s why I started in physics. Then by chance I encountered somebody who did a Master’s in Engineering following her physics, and piqued my interest in that, and then I just lined up going from there. AI is currently in the position where it has the opportunity to be a nice bridge between technology, and have impact, and that’s why I have now started to gravitate towards that.

Wei Zheng: Thank you, so could you start by giving us a little bit of a crash course on AI?  In the news media we read a lot about AI, machine learning, neural networks, and other terms. What is the fundamental difference between AI and other technologies? We also have a question from an audience member who asks: How is AI different from a series of if statements?

Dr. Suba: That’s a good question. So AI is very broad. It’s really literally meant to try and create intelligence that is not in a biological specimen. That’s the very high philosophical moon kind of goal for it, so it is an umbrella term that tries to understand that it’s not just about creating, it’s also about understanding intelligence in that way, understanding how we think, understanding how children learn, it’s the whole thing. Machine learning really refers not to how we do it. Again, machine learning is not equal to AI. I wouldn’t also call it a subset of AI. It’s kind of like these are all things that overlap and machine learning is really about trying to teach a machine to do a specific task or tasks in a certain way. Let’s take the example of a decision, right? One of the things I work on is Alzheimer’s detection and we detect it from a piece of text. What I’m actually trying to design here is a classification system. It looks at this piece of text and tells me whether this is a person with Dementia or Alzheimer’s or not. It’s a sort of classification problem. And these classification boundaries, if they’re all multidimensional problems, meaning you’re thinking of these data points in multiple dimensions spaces. If you can think of this in a gazillion different dimensions, that’s really what we’re trying to do. The data points are like if you take a bunch of soap bubbles and blow soap bubbles in 3D and some of them cluster in one area and some of them cluster in the other area, this is your clustering problem. You’re trying to say this part is a certain thing, and this part is a certain thing, and the boundary between this cluster in the classical sense used to be a plane or a straight line. But when you go into deep learning, that boundary can be a little bit more fancy. So it gives us more degrees of freedom and that’s why it tends to do better than regular machine learning. So what’s the difference between if statements? With if statements, you know really what could happen and you sort of give them a bit of the rule based do this task if you are here, do this if you are here you do that. In machine learning, you’re not quite doing that. What you’re doing is letting the machine learn what’s going on in the data, and then think of clustering it. This clustering idea is not new, it’s been there for a long time. In fact, my PhD thesis was an extension of that for a very special problem. You are looking at classifying classification problems, finding these boundaries, and the cloud clusters are generated by these algorithms. There are if statements in there, it’s not like it’s not there, but it is not only based off of if statements.

Wei Zheng: Thank you for explaining that. So there are two particular areas of AI where you have deep expertise in and the audience has shown interest in both areas. One area is mental health, detecting early Alzheimer’s, and the other area is fake news and the detection of fraud. Could you talk about both areas a little bit more because are very interested in what has been done and what is going on in your current projects?

Dr. Suba: So although those are two different tasks, for a person who is designing an AI engine, they all fall under certain kinds of problems, so the flavors can change and how you actually make these engines will probably change, but the essence is still the same. In this Alzheimer’s detection problem, what I’m really asking myself is because I’m working with text only at this point, and the reason I do that is because it’s non-invasive. You can get people and spoken words. You can translate that. It’s much easier. So when you talk to doctors who are in the clinical setting who need to administer these tests like verbal tests for the people to determine whether they have Alzheimer’s or not, they find that there are a lot of emotional components. People don’t want to take the test because of how they feel it’s exposing their weakness. There is a lot of stigma attached to that and they get angry. Complaints are a problem for them to come back and do the test for people to monitor that becomes very hard, so if you can implement something like this in your Google system or your Alexa or something like that, because you have it in the Google system or you can say make it all as a phone app or something, then it’s not in that threatening kind of situation where they have to go somewhere else and they’re feeling tested. And it’s not like that. It’s a little bit more like my Google is trying to talk to me. It’s a little bit less threatening and therefore complaints are easy, but all of these medical kinds of problems are hard because many other ailments can also have similar problems and symptoms, but the doctors we’ve spoken to have said that even if you guys can design things that work in a population where we suspect that if there is a cognition problem and it’s probably going to be Alzheimer’s then that in itself is very valuable. So in some sense we have that thing to play with.

So fake news is also a bit of a classification problem, and it’s also working on text, we work on text. So we take up an article and then the job we do is design an AI engine which tells us if this thing is fake or not and in both of the cases, me personally I’m not interested in just designing a classification engine, but I’m also interested in designing an interpretable classification engine, so one of the big things that AI always faces is that it’s a black box, we don’t know what’s happening why would you trust it? And especially in things like medicine, that’s a valid question. If you went to a doctor and the doctor said I don’t know the computer told me that you have this disease, so you have this disease, it is not going to work. So if we can point to certain things and then tell the person or the caregiver these are the reasons why we think you have this disease or you have Alzheimer’s or depression or whatever it is, that adds more value. So I’m interested in explainable slash interpretable AI designed specifically for these kinds of problems so that the architecture that we have built is interesting in that it is what we call modular in the sense that we designed it with the future sort of in mind and we don’t just take plain text from the patient or from the news article, but we also add certain features that experts in this area believe you will see as an indicator. What we’ve done is these features can now be sort of added on to our engine and if in the future the experts think that something like a cognition score that you can determine in a certain way can be useful in detecting somebody with Alzheimer’s, we can potentially include that in our architecture potentially without too much change, so those are two things I’m working on at the moment.

Wei Zheng: Could you talk more about the black box idea?  I think that’s the intriguing part about AI making all these decisions. There is expert input in terms of telling maybe the algorithm or some important variables to look for, but then AI makes some other decisions, right? It uses more information to make decisions. So how does AI actually make decisions? And what can we learn from it as humans? Maybe we can learn something from it.

Dr. Suba: So it’s an interesting question and it needs to be sort of like unpacked carefully. The big thing about most of the neural networks today is that they are very complex structures. In a sense, it’s an optimization problem, but with so many variables. So for humans to give an exact understanding of every single step of the AI it is probably not going to work, and it’s probably not going to work in the sense it’s computationally impossible. You could potentially keep track of every parameter that you had and you could throw about a gazillion different charts at people, but that won’t make any sense to any of us. So what we want most of the interpretable AI community generally works on is two ways of doing this. One is to design what we call post hoc explanations so that AI which is designed to explain is not the one that is making the decisions. So you take an AI that functions well for something, you run the algorithm on whatever it is, maybe it’s fake news detection, maybe it’s a mental health issue detector. Once it has given you the results, you would run this thing on top of it and then say these are reasons, but so far this area is very new, so although a lot of people are working on it, we don’t even have a clear idea of how we can rank explaining ability. We don’t even have a metric for it. We don’t even have a full definition because it’s hard. What is it? We have some idea. What we can say is that certain kinds of explanations will make sense to the developer. So the person who is making that AI engine if they realize that XY&Z is the reason why this is giving you this answer, and if that XY&Z is fishy, you know to go back and fix the algorithm. So as a case in point, I can tell you I’m sure you’ve heard of this fairness issue in bias, especially so when somebody designed a soap dispenser system for some airports and that just would not dispense soap for people that are dark skinned, it was able to work well for light skin people, but it wasn’t working well for dark skin people. It was not any intentional thing that happened. It happened because the data on which this thing was trained was on mostly white people, so it just didn’t recognize darker hands. If you can get that insight, if you can get an end before you go and launch this product and find that it fails. If you can do this kind of stuff and see this is the reason why it’s doing, and then you realize, well, hey, why is the color of the skin being an important thing for this AI to design?, then you wake up and realize maybe my problem was in the data so I fixed it. So this is one kind of explanation where you give the developer an idea of what’s happening in your big AI thing, and then you try to fix it. The other is to give the end consumer a kind of understanding of what’s going on. So if you have any engine that’s making decisions on credit, credit cards, or bank loans, or things like that you want to be able to tell the person that your credit score was coming in this region, and something else was happening here and then you just switched your jobs too many times in the past three months. These are actionable intelligence things that this person can understand, right? So depending upon which domain we are talking about and depending on what you want to do with that explanation, what explanation changes and based on that you define you have different ways of approaching this problem.

Wei Zheng: Interesting. I know some of our audience members came here because they wanted to know more about detection of fraud and fake news. Could you share some of your research and knowledge in this area in terms of for example, if we look at articles or Facebook posts and things like that, what are some commonalities that fake news have and do you build an engine to detect the fakeness in those texts or speeches?

Dr. Suba: So even within our so-called machine learning community, the sort of first generation community of machine learning algorithms did this thing called feature engineering, which is a laborious process, takes a lot of time, and what it does is it extracts a bunch of features from the text, and then it tries to figure out which of these is important. Some humans are programmers doing that. Then it then uses that to further design your feature extraction, but as you go and as you start, the nice thing about the deep learning thing is that you can expand the series set of features to really big things and you don’t have to predetermine which ones they are, which one is important. So what we found almost consistently is that emotions are some of the indicators of fake news detection. I am always scared to tell people that if you see XY&Z then those things are indicators of fake news because people have a tendency to think that oh, ok, so if I look at this and I see this, then that’s it, that’s fake news. But remember that as I said before, we are dealing with several features, multiple features at a single time, so it’s not just one value of that feature that will determine it, it is the combination of where each of these features are sitting in that big space that actually determines it. So some of the indicators are how emotional this whole thing is. Oftentimes, these fake news things tend to be attention grabbing. They want you to look at it, and you’re not going to, well, you probably will, but a bland way of writing is probably not going to get that many clicks, and if it does get that many clicks, it’s probably not going to get that many shares, so there’s lots of research that has been done in social networking, which said that stories that have a sort of like you know if it bleeds it leads kind of thing, if it has that element in it, it gets shared a lot more, it gets slightly more so fake news designers tend to use that, exploit it and so kind of ratchet up the emotionality of it, and so on and so forth. So we are looking at those kinds of features, and in our case, we were looking at a bunch of other things like readability score. The readability score really measures a person’s reading ability in 3rd grade level, 4th grade, 8th grade level, and college level. That sort of thing. These kinds of scores also claim a role in determining if a piece of news is fake or not. So there’s a whole slew of such things we can extract. There are many ways in which you can give a numerical score to these kinds of concepts that are more philosophical. You can make it numerical. And then you can feed it into your algorithm. And that’s what we find.

Wei Zheng: Interesting. So the emotionality of the messages and the ease of understanding.

Dr Suba: And sometimes also it relates to the title of the article. So the title might say something, but what you’re seeing there is kind of really not focused on the title. It’s kind of veering away and giving you all sorts of stuff. So how much is it relating to the title? That’s also good-

Wei Zheng: quality of the writing. So how do we tell AI? Do you tell the AI these are the indicators to look for? How accurate can AI be in detecting fake news?

Dr. Suba: Right now, we have about 90 something percent accuracy in very standard data set that people have created and PolitiFact is a good place where humans have taken the time to find what is fake and what’s not fake and actually created a data set like that to see how things are able to detect it.  You had another question in the beginning that I may have missed.

Wei Zheng: It’s okay, how about Alzheimer’s?  How is AI detecting Alzheimer’s, and how accurate can it be?

Dr. Suba: Currently our Alzheimer’s detection rates are at 90% accuracy, which is again from a known data set that’s been around for a bit, and so on and so forth. So this is what we do, which I kind of like. One is something that we call latent features, which means don’t tell anything to say about what to look for. Just find a vector of this text you have in a known way of embedding meaning, a vector representation of your text, use that. But in addition to that, use a few more features that people have thought of as being indicated, like people have done qualitative research on this, so I added that to this and I found that having both together works well in determining Alzheimer’s.

Wei Zheng: How adaptable is AI in novel situations? If AI is able to figure out something in one situation, how transferable is its ability to detect maybe Alzheimer’s or even different fake news in a completely novel situation?

Dr. Suba: So the whole thing is about your data set on which you’re training. If you get a good data set that is truly representative of the general population, then you can build a strong AI engine, but if you don’t have something like that and you have kind of focused on a very specific kind of data set on which you’re training, then it’s not going to happen when it sees something that’s different. For instance, this is something that we should be doing further. So currently our Alzheimer’s detection data set is at 203 hundred people. I think from a cohort as a data scientist, that’s a small number. The number of transcripts will be about 1000. That’s a small number, but for people who normally do this kind of studies in medicine, that’s a large number because finding that many people, tracking, keeping track of them is not an easy job. However, now I have built so many Alzheimer’s research centers and they are collecting data, so in the future what we’d like to do is be able to get that data and be able to do a much more vigorous training of our AI engine on that kind of data and then if that brings us together, if it means that we have to change a few more, we’ll start to learn more when we design, when we see that kind of bigger data, so you’ll see, okay, maybe I need to tweak this year after that, and then the design will evolve with that. So that’ll be the next step. The bottom line is the data. If you cannot get the data, you have to be careful where you play.

Wei Zheng: I see. If AI can be deployed successfully in the large scale to do detections of fake news or Alzheimer’s, what do humans have to do? What do humans still do that AI can’t do yet? And what can AI do that humans can do?

Dr. Suba: Humans can do a lot more like what AI can do today. We can generalize a little bit. I mean, we’re not always good with generalizing, but we can do it. We can do empathy. So far, I don’t know of any AI that does empathy. AI is really more like as it is today. It is more about just getting very good at one task, so we are in the realm of what you call narrow AI, not general AI. So what I like to say there is, your Roomba, the floor cleaner robot, is getting better and better, but it still in my home often shuts itself inside the bathroom and tries to eat up the bathmat so you know it’s not even that good for doing just the floor cleaning. Then there is another small bit of AI that’s in your car that tells you hey, don’t change lanes and you know it does that in every car now. You can’t strap this Roomba on that and then assume that this new engine is now going to be able to not only help you clean the house, but also help an old person cross the street. It’s not going to happen, so there is no transference like that. There is a lot of work going on now in transfer learning, but it’s still very, very small, very narrow at this point.

Wei Zheng: So one of our audience members asked this question, “There seems to be a lot of hype around AI because it’s like AI is taking the place of humans and there are fiction novels and  movies about how AI is controlling, influencing, and deciding what our future would be like. Do you see a lot of hype around AI? And also, what are some real major trends and bottlenecks in AI research and application?

Dr. Suba: So yes, there is always a group of people who think that if we get a really good rumor, that rumor might take over life, but it’s not going anywhere soon for the same reasons I told you that it’s very narrow focused, it doesn’t have the ability to crossover and do all this meta-learning things. We’re not there yet. There’s always hype and now there’s also a tendency to think that AI will solve all the problems of all the world and everything else and that’s also not true. It’s not anywhere near that. So it’s a great technology and it is a nice new way, some people call this the 4th wave of the industry. It’s great, but we need to take it for what it is, which is like it’s next wave of what’s happening. It’s not going to be one of those sci-fi universes just yet.

Wei Zheng: It sounds like you talk a lot about data. If it doesn’t have enough training data, then AI cannot perform as well as we’d like it to be. Is our data a bottleneck? We do have a lot of data, so how do you make sense of them?

Dr. Suba: We don’t, we do, and we don’t. We need to have data that’s clean, that’s labeled, that’s representative. And that is where we don’t, we’re not yet there. We have a lot of data. We also don’t have a lot of data because it’s not often properly labeled, meaning it’s not enough for you to say that I have all this information collected from people, but there is some human that needs to be sitting there reading that thing saying this data belongs to this class and this other data belongs to that class and we don’t have that everywhere. We have a lot now compared to what we did, but we still don’t have everything that we would like. The other challenge will always have is the amount of computation. Yes, we have gotten much better computation nowadays, things can work much faster now, the CPU’s are much faster than they were when I started my career. That is one of the things that’s pushing our sort of evolution in this AI, but energy is not infinite and so on the other hand, we are also talking about global climate change and being careful about energy and so on and so forth. So it is something that we need to be also conscious of when we are developing these things but it’s a hard ask at this point

Wei Zheng: What are some trends or promising directions in AI research and application? A few of our audience members are students who are considering maybe getting to this area. So what are some really productive areas that you would suggest they get into or take a look at first?

Dr. Suba: The AI used for image and image-related things is very well developed and is continuing to do that and so that’s a great place where you can look into. And I’ll tell you natural language processing is a great place that a lot of industries wanted, like the finance industry. Practically every industry likes it. It’s now being used in not just your chat bots, you know question answering systems, but beyond that, like for your customer care and customer retention and it’s just everywhere, then everything is everywhere. So getting into these things are probably, even if you’re not interested in pushing AI further, learning how to use these tools, learning how to figure out what the next best tool is in that thing could make it a very marketable skill for you, because AI in these zones is very much in demand today.

Wei Zheng: Thank you, so let’s shift a little bit to applications or people trying to use AI. So if some organization wants to use AI to do something, maybe it’s to understand their customers better, or maybe it would make better recommendations to their products or other things, or self-driving cars, so these industry applications, if some organization wants to get into it or considering using it, where do they start? What kind of questions should be asking in order to consider whether they should or how to involve AI?

Dr. Suba: So it depends on what the industry wants to do. So for example, if you’re thinking of automating some part of your system, the first question you would ask is what does this mean? I’m assuming they’re not doing R&D there and they’re thinking about getting cards, products or what’s already off the shelf? So the first thing you would ask is what is on the shelf like what’s readily available, and if it is really readily available, what kind of data have these people used to train? Was that really representative of your customer base? And then of course you will be asking the usual things like how good is the accuracy and how good is this method, but the other more insightful questions you would ask is what data did you train it on, and then maybe you would do a pilot test first on your own data and see if it’s working or not. That’s where I would start.

Wei Zheng: Great suggestions with representativeness of the training data. Yeah, that’s critically important. What are some mistakes you have seen or heard of people make when they introduce AI into their work settings?

Dr. Suba: Generally speaking, one is thinking it will solve all your problems, A. And B, thinking that you need air for all problems, or most problems, and it’s a good idea and I think really long and hard about it because some of the more classical learning training techniques can do the job that you need without a big performance sacrifice. And remember, most of what AI at least in the training stage is going to take a lot of computational resources. So do you really want that? And if it turned out just because you want to say you have done AI in your product, do you have to do it? Like I said, it’s a decision or prediction problem, and if those predictions slash the decision area naturally those boundaries are flat or straight lines, your classical techniques will do a great job. You don’t need these other things. So that’s something I find people do and I also see this a lot sometimes with entry level students. You give them a problem and then they realize that this data is a vector, I found something like a code that’s available on GitHub that works for vectors, so I’m going to stick this thing in that and I’ll spit out the answer, hey look, I’ve done it. You don’t know that. Just because it accepts a vector input doesn’t mean it’s the right thing for you or that you have the right kind of vector for that AI. So you need to dig a little bit deeper, you need to understand what it is that it’s doing, not just like a program, but beyond because there are restrictions of what it means. You might be doing something completely absurd that AI was never designed to do you think that, well, I threw an AI on it, so it must be good right? So there is that danger if we don’t do enough due diligence and understand a little bit more.

Wei Zheng: So could you talk a little bit more about what kind of problems would be appropriate to delve into or investigate the usage of AI, and what kind of problems would be better to be avoided or to get into or try to find other tools for?

Dr. Suba: That’s a tough question to answer on a broad basis because it really depends on the nitty gritty of that problem. It’s like trying to think of I want a better data entry system, then maybe you don’t need an AI to do that. But if you have too many customers and you don’t have that many people to take calls all the time, then having a bot is a great idea. At least you filter out some of the easier questions and things. So it’s difficult to say that in a more general way, because industry is so wide.

Wei Zheng: We have the last group of questions before we open it up and they’re about people preparing for the coming of AI. Maybe that’s the hype talk, right? The coming AI or increasing applications of AI in different areas of life. So how do we prepare for that? For people who are in the technical field, not in the technical fields, and for many of our students respectively, what’s your advice?

Dr. Suba: So students are probably easiest for me to answer because I am a professor. So learning about AI a little bit more than what you will see on a magazine is useful because that gives you more of an insight into what AI can do and cannot do and if you’re in one of those engineering fields, then definitely take courses that will support development of AI and machine learning because that will give you an insight later on when you are out there in the job market or if you’re in a job and then your boss tells you , “Hey, I’m looking at this, can you tell me if we need this or not?,” you have the basics, the foundations to evaluate whether first of all, do we need this or not and you also have the ability to say which one is good and which ones are not, so I would say that for those who are students to read to themselves and if it’s professionals and people that are working in a company, you can expect this thing to be everywhere, so again, reading some of the more tech savvy medium kind of posts or a little bit more advanced than that is a good thing. There are so many certificate courses available that are for professionals, those are good. If you think about companies as units and what you can do to prepare for AI, then I guess you’re asking questions like if I don’t jump on this AI bandwagon, am I going to lose out? Is my customer going to go away? Then you have to take a hard look at what the product that we are finally selling and what ways can this AI actually help you do something like add value or a different market share or a different market altogether? It lets you target another market or not. Those are the questions you need to first ask and then again the same questions like do I need an AI to do this or do I just find some of the more classical things that do well? So that’s what I would say from different perspectives. I hope that answers the question.

Wei Zheng: Yes, it makes sense. So there are two parts of the question for professionals. One is the business question: What specific questions do we want AI to help us to solve or another technology that we need it to solve? And the other is the technology itself: How does AI help and what’s the data set based on which the AI is trained and is it appropriate?

Dr. Suba: What are the ethics because I’m sure you’ve all heard of what happened with Google- their top ethicist was actually fired because of some problem in that that didn’t go well, so in fact, if you’re a business, I think it’s almost imperative. And if you have some AI in your company, I think it’s important to have an ethicist working in your organization so that they can ask these tough questions: Did you train it on this kind of data? And because sometimes engineers are so focused on doing things they might not even think about it and in fact I would say going back to your question on what students can do, I guess you can even ask that question on a more meta-level, what can universities do, right? And I think one of the things it could do is make it mandatory for engineers to learn about ethics. Engineers can sometimes get so focused on what they’re doing, the optimization problem, whatever it is they’re doing right now, so they do not think about that, but it has become such intertwined issue that having that background will help the success of that student.

Wei Zheng: Thank you very much. Now we can actually open up for questions from our audience. You can either type your question into the chat box or you can just raise your hand and Yash will unmute you. We do have a question here on how to distinguish between deep fakes and real entities?

Dr. Suba: Deep fakes are a much harder problem simply because it is an AI that’s creating it, so these kinds of problems are literally a cat and mouse game between defense and offense, so as these AI start to get better at creating deep fakes, your AI engine needs to start adapting to catch that, so the way you would do it is start training on the deep fakes and when they do, it’s going to be a tough problem. It’s not going to be that easy, but the interesting thing is that the technology that you need to create these deep fakes are not in the hands of a normal person that wants to do something somehow. It’s more in the state actor/big organization realm, so this is not going to be just be solved by AI. That problem is going to have good old-fashioned reporting, good old-fashioned fact checking.

Wei Zheng: Thank you. We have another question: Could you have an AI engine for training other AI engineers? I think another participant probably asked the same question in the registration form which is: Are there AI engines we can train ourselves on to get started? And this question then has a little bit of a pivot, it says to help reduce any bias data or other human related errors that contribute to the training.

Dr. Suba: So the first question, I’m not sure if you’re if you’re asking me if there is an AI engine that will train an AI engine or is it an AI engine that is training you?

Wei Zheng: I think it’s an AI engine for training other AI engines. Could you have an AI engine for training other AI engines?

Dr. Suba: There’s a lot of research that’s just beginning to happen, and it is still, in a very narrow sense, trying to do a certain thing. That is a possibility, AI engines training people, you can. In fact one of the things we’re thinking about is: Can AI engines help with rehabilitation of people that have injuries? So just like how a physical therapist is trying to help you move in certain ways when you have a certain kind of injury, some collaborators here at Stevens are thinking about it.

Wei Zheng: The second part of the question is to help reduce any bias data or other human related errors that could contribute to the AI engine. I think that’s probably why Jill asked the question does having AI engines train other engines help reduce some of the biased data or other human related errors.

Dr. Suba: That might not like if you have a problem with the data, I don’t think you’re going to be solving that with another AI, so you can explain why and I think it can kind of point you to the fact that it can make you aware of the fact that you have big problems with your data, like you were missing a whole chunk of something. But I don’t think that’s going to take care of the problem of bias being introduced into that. Human error, yeah, that that comes down to how well the data tagging process was. So yeah, we want data that was labeled. We also want to make sure that the labeling process was good and it was done by people that have the expertise to do it, it was done by people who were not just randomly incentivized to do more. Like you know, if you just keep incentivizing people to do more of these tagging, they might push themselves way past their fatigue and then your data quality will drop. So how do you deal with all of that? So this is much more complex. That’s why you need someone that works on ethics and things like that to control those, but there must be processes in place that vet the data that is fed to the engines. So in the end I would say that this is all something we cannot have the machine reduce at the moment. This is all something we humans have to take responsibility for.

Wei Zheng:  Which is interesting as a social scientist when I see these AI errors, they’re actually representing human errors, it is just sort of amplified out for us to see more clearly, so it really takes intentional efforts to ensure fairness and equality in the technologies or engines we build. And we do have another question having to do with this: Is there a guideline or reference source that you would recommend regarding ethical considerations in AI?

Dr. Suba: I don’t remember it off the top of my head, but I know that a couple of years ago some of the big companies, Microsoft and a whole bunch of them formed a coalition, around 2017 or 2016. They formed a coalition for a code of conduct for people who do AI, and so if you find if you try to Google that, you might be able to find that link, otherwise I can look for it and send it to you. So what is the code of ethics for people who are working in AI? There are people are beginning to work on that and think seriously about it.

Wei Zheng: Any other questions from our audience? I’m going to ask one more question from the registration form where somebody asked about your own experience as a woman in this technology field: What has your experience been like? Have you faced the glass ceiling and how have you been navigating this? It’s still a male dominated area, so what’s your experience been like?

Dr. Suba: It’s a male dominated area. Glass ceilings, you know they exist. They exist everywhere. What I’ve heard people say is find advocates, make friends, get a big network going to the point where you are not isolated. It’s not an easy job, but as a species, I think we have a long way to go in that department. But it definitely does exist.

Wei Zheng: You talked about finding advocates and friends. What are some ways people can do that effectively?

Dr. Suba: Reach out. So at the beginning of my career I think I made a mistake in that I am an introvert which means I’m not a person that talks, so I don’t do a whole lot of small talk and it’s not that I don’t like it or anything, it just doesn’t come up naturally, so I have to train myself to be that way, and so I never really reached out that much to people. But I have since learned to do that. And you know, always remember that the world is a really large place and if there are five people that don’t want to talk to you, maybe you just haven’t found the right five people that do want to talk to you. So don’t lose heart and just continue talking and someone is going to see the value in what you do. This is something you can do, but there are always things that will happen that you cannot do, and so another thing I have learned from personal experience is you know you do your best in the true sense of the world, but you’ll end up getting stuck against things that you cannot change. You know what they say about banging your head against cinder block walls, right? It feels great when you stop. So find something else that is that will help realize your potential and keep working in whatever way you can to make it better for the person that’s coming behind you. That’s how I live my life.

Wei Zheng: That’s great advice, find your community. Find a community that lifts you up, yeah. Two more questions. One person asked about specifically about the military sector. You talked about implications for the medical sectos. What do you see as implications in the defense/military sector?

Dr. Suba: There’s lots of places where you can apply AI in the military sector. So for example, all the fake news type things directly play into detecting other government influences in destabilizing the government, in destabilizing the country, so there are a lot of applications for the defense sector. And then if you are not in that type of defense area there are regular control theory type things which are used in a lot of library type systems and things, so there is an applicability and robotics is often used in so many, you know underwater vehicles and things like that.

Wei Zheng: Another question: How is the current implication status of AI application in surgery practice?

Dr. Suba: I don’t think I am qualified to answer that question because I don’t know how that’s going. I guess it will be strongly related to what’s going on in the robotics part. There’s a lot of those reinforcement learning type problems there, but honestly, I haven’t thought about that enough to make a meaningful answer there.

Wei Zheng: Okay, thank you so much, Dr. Suba. I really appreciate the conversation and really appreciate the view of the non-hyped version of AI. I really appreciate knowing that and how we can prepare ourselves in understanding how to differentiate different kinds of AI or different. sales language when it comes for us to differentiate or make some judgments about where to go, what technology to utilize. I really appreciate it. Thank you so much.

Dr. Suba: Thank you so much. It was a pleasure.