Mohsen Rezayat - Adoption, Value, and Ethics of AI in Software
Show notes
Today’s genius mind is Mohsen Rezayat, Chief Solutions Architect at Siemens Digital Industries Software. Tune in to this episode to learn how knowledge graphs are used for anomaly detection, how Siemens could benefit from publishing an open-source API for industrial AI, and why a well-known scientist stopped his computer vision research due to ethical concerns.
Transcript
You know, why can't we have human-centric AI models for automation and be dedicated to creating an ecosystem with no losers, benefiting everyone from executives and shareholders to workers and consumers. Welcome to a brand new episode of our podcast, Human and AI Mind, Machines and the Grave in the sand. Thanks for tuning in again to geek out with us over the fascinating field of AI and machine learning and the role of humans in it. Today, we're super happy to have Mozen Rezayed with us. Mozen is the Chief Solutions architect at Siemens Digital-Industries Software and an adjunct professor at the University of Cincinnati. So Mosen, welcome to the show and thanks so much for taking some time to join this little podcast. How are you and where do we catch you today? Hi, uh, uh, everyone, it's a pleasure to be here with you. Um, I, um, live in, uh, Cincinnati, Ohio in the United States right now, but, uh, my path to get here is, uh, interesting in that I was born in Iran and came to the States when I was, uh, 17. I attended the University of California, the Irvine campus, and then, uh, my studies at the University of Kentucky with a PhD in engineering mechanics. I was then hired by a company called the Structural Dynamics Research Corporation, which was eventually acquired by Siemens. So I've worked within the same organization for over 36 years and now lead the advanced Solutions Group with the mandate of doing R&D across the entire portfolio. Of uh Siemens Joe Industry Software or the IISW as we call it. I'm also an adjunct full professor at the University of Cincinnati, as you mentioned, and started a nonprofit organization called uh Omi USA about 12 years ago to raise the income level of poor farmers by introducing affordable and Sustainable solutions. So that's sort of a brief version of how I got here, and I truly believe that we need more conversations about the role and circumstances of humans in everything we do, including. So again, thanks for having me. I do it. That's, that's people, passion, purpose. They're all in one. I can't believe it, dude. That's super awesome already, right? And I guess we have in, in, in, in Germany we have a saying called ein urgestein. That's typical, but sounds harsh, but it actually means an English institution, right? You are there. In the PLM space, right, product life cycle management space and also software space and also innovation space for, for quite a long, right? How do you see if you reflect a bit on the past, right? And you know, in the past years, how, how do you see the, the software business and the industry? How has that involved? And are there, do you identify certain major milestones, um, along this path? Um, thanks, Ali. You're, you're too kind. One, observation I would like to share with you in my journey with Siemens, uh, is that the rate of change, which, uh, historically was linear. It is now exponential, uh, in my opinion. Reasons for this include exponential growth in the amount of computing power and available data for creating amazing AI models, doing complex simulations, creating very large, realistic 3D models called digital twins, and many more examples. Now, a couple of indicators for this in our world today that would hopefully make believers out of the listeners is that it's sort of amazing to me to think that China can identify any of their 1.4 billion citizens in about two seconds using AI-based facial recognition. Intel has more than 7 billion transistors on a single chip. You know, it took like 30 years to go from 2 transistors in the 1950s to 20,000 transistors, but then all of a sudden we jumped to 7 billion. What is equally amazing to me is that the rate of adoption of technology is also changing. You probably know that televisions were invented in 1925, but it took 25 years or 25% of Americans to own a TV. Now, compare that to smartphones that were invented in 2007, and it only took 2 years for 25% of Americans to own a smartphone. So this exponential rate of change and the rate of adoption of technology tells us that our development, operation, and business models. Need to evolve as well because customers now expect rapid change, and that extends to the use of software in industry. For example, our customers are demanding even larger scale visualization, collaboration, and simulation that's given to them seamlessly across multiple geographic sites and applications. We also see that software as a service. The so-called SASS is becoming the delivery method mainly for ease of access, scalability, and integration reasons. These trends of rapid change in development, operation, and business models must continue in order for us, Siemens in general and Siemens, the IISW in particular, to successfully answer the demands of the market. Very well put. Yeah, super interesting and impressive facts and yeah, and valid points. Most of you, you have to know about me, I, I love quotes and there's a very nice quote from Peter F. Drucker who said, if you want something new, you have to stop doing something old. So challenging the status quo and exploring the new is something I personally really love. And I would be super curious to know, is there any quote that you are in particular interested in or that, and that you would like to share with the audience? Yeah, this, this answer is, is easy. Um, my uh idol is Albert Einstein, and he has many quotes. I actually, I can show you around my office. I have them on the walls, uh, the quotes from him that I, I really like many of those, but the one that always stays with me is, if you can't explain it simply, you don't understand it well enough. So, whether I'm teaching my students at the university or when I try to explain something new. And exciting to colleagues, I try my best to explain it as simply and plainly as I can without using, you know, complex concepts and big words. By the way, I also like the code of go something like this. If we knew what we were doing. It wouldn't be called research. Quite often to the people around me because they always want me to define everything for them, and I always quote this from Einstein to them. Yeah, I will use it directly tomorrow if somebody's questioning again to research and apply to research. But speaking of software, right, software in current times, you know, I love this, this, it has this open source movement actually, entire young, young generation is pushing their ideas, you know, as fast as something want to. Hub on Archive X, right? If even before conferences, even if you thought like conferencing, you know, research and a lot of research conferencing, you know, they already uploaded this stuff along with code, along with, you know, it's beauty of, of sharing, and this sharing fueled from my perspective, significantly also the, the progress in machine learning. How do you see the role of open source within a corporation and maybe also along the line of the, the PM and in DI software industry? I really, truly believe that one of the drivers for the exponential rate of change I was talking about in the last two decades is the open source software. You know, my own team couldn't do the amount of feasibility studies that we have done in the last 3 years if open source AI NML software like Tensorflow and PyTorch were not available to us. Now, at Siemens, we have practiced this open source concept, but we need to do more, especially with our PLM software at the ISW, if and when it could benefit us as well as our customers. For instance, let's look at JT, which is the de facto industry 3D model format we created, got through the ISOL certification. And then offered to the entire CATD industry. That was great. However, I, I think it would have been even greater if we had also open sourced the JT toolkit. This is the C+ API library that enables read-write access to all data within the JT. Open source and offered it for free, similar to what other formats like UST have done. The result would have been a much larger adoption of this format beyond our immediate industry, such as in game engines like Unity 3D. This would have opened many more doors, in my opinion, to us and to our customers. We now, I believe, need to do the same with the industrial AI related tools we created, similar to what Fallo and Pytorch have done. Because at the end, Siemens will win since we understand better than anyone else the needs of the industry. Of course, shouldn't and won't open everything up because intellectual property needs to be protected. I understand that. But I'm certain we can figure out a strategic way of accomplishing a win-win situation. You know, if Google and Facebook can figure out in a B2C context, how to open source what is central to their environment and AI environment, I'm sure we at Siemens at the IISW can do it in a B2B context for the PLM community. That's, that's nicely put. That's actually, you know, that's nicely put. I think we have a long, a bit of a journey still ahead, isn't it? We're going there, yeah, absolutely. There's so many potentials for synergies actually. But speaking of, uh, potentials, where do you see the role of AI and machine learning in the Siemens software and PLM? Can you maybe share some use cases there, maybe even 3 or more that you are particularly passionate about. I think this goes to really the topic of this podcast. I think generally speaking, in, in PLM AI enabled machines and software are augmenting human workers' capabilities. AI enabled systems do what they can do best, which includes like analyzing huge data sets, generating lots of design alternatives. Answering questions about those designs, performing repetitive tasks, handling routine cases. On the other hand, humans do what they do best, which includes resolving ambiguous information, performing delicate tasks, and we'll talk about this hopefully later. Choosing among many alternatives based on aesthetics and or marketing goals, selling what is created and dealing with dissatisfied customers. Now, specific to DISW, we actually have many examples of such AI enabled software from Mendix and MindSphere applications to Team Center and X products. To calibrate, to process, simulate, and MM software, and I'm sure I'm forgetting other products that have some sort of AI capability at the ISW, but I want to give you more details about 3 capabilities in these products that I'm personally familiar with. The first one, which is actually in the hands of our customers today. It's called adaptive UI in our NX products. This is our modeling system, 3D modeling system. Very complicated product, many hundreds if not thousands of menus and submenus, and what we have been able to accomplish is with a high degree of accuracy, be able to predict what the next command that the user is going to need and bring it to them rather than making them go and try to find it. And again, the customers just love this capability, and we use Markov chains, and now we're kind of thinking about even evolving that perhaps to a deep learning model for this, but the point is that even something that seems very rudimentary could become a huge success. The next one that we are working on right now is anomaly detection and repair using knowledge graphs, which I hope you guys ask me a question about that because I want to explain it more, but we use knowledge graphs to integrate our Teamcenter MindSphere and an ERP system from a partner together, and we actually can now in a single environment in a browser. Allow a user to monitor an entire factory. Detect when there are a broken part or a machine that's not behaving properly. Get from the sensors the part that is causing the problem. Get information and metadata about that part. Order that part. Then once it's delivered through that same browser, get instructions for disassembly and assembly of the part. And then continue monitoring. So this is really tremendous, and it's all enabled by AI and knowledge graphs. And the third example is uh digital assistants, you know, these are now in a lot of our products being, being developed, but the, the one that I'm actually directly involved in is with our mentor graphics colleagues in a product called Capital where harness manufacturing is complex products and for the foreseeable future, it's not going to be automated because putting these wires into these small pins. At least for as long as we could see it needs to be done by humans, and humans, of course, make mistakes, and mistakes are very costly to try to find out a mistake in a harness which is bundles of wires in your cars or planes or ships after the fact is a daunting task, and we were told that if we could even save 1%, reduce the number of errors by 1%. That would save companies up to $10 million. So we are now using head-mounted devices. It could be a Google Glass or HoloLens, and we are using the camera to not only bring the work instructions to the users, but also monitor them as they follow those work instructions. And if they make any mistakes or they forget a step. We can point it out to them in real time and let them make the correction without trying to figure it out after the fact, which I say is more like finding a needle in the haystack. So, these are all exciting capabilities that we are, we are developing and all of them are being enabled because of AI. Love it, right? Po, part and super challenging for certain pieces, right? But love it. But I also poke you directly with the knowledge raft question, no question, right? We, because obviously a natural reflects in some, some kindness or, or actually complexity is bombs, right? This is something like bill of materials that say that the management of, of the product structure, right, for the manufacturing stuff. And, and they, these bombs like this bill of materials can be super large, right? Super complex. You know, mentalities, you have, I don't know, images, taxonomies, or, uh, you know, documents attached to that. How do you envision a bit going further than maybe you can elaborate a bit on the role of knowledge graphs for DII software, right? That you untangle a bit this complexity, right? Or you think it's just a hype, uh, or will it play a role? What, what's your view on that? Uh, absolutely. I'm so glad that you asked me about knowledge graphs again because I want to explain. What I have said before and perhaps answer this question as well. You know, generally speaking, knowledge graphs or KGs are a critical tool for incorporating complex knowledge into intelligence systems. Using these KGs, knowledge is encoded in a sort of a graph-based structure where nodes represent real world entities while edges define multiple relations between these entities. And in fact, KGs provide a flexible way to capture, organize, and query a large amount of multi-relational data. The type you mentioned for management of product structure in, in manufacturing in, in your question. But KGs also provide a semantic-based perspective, which is essential for performing deductive reasoning. At TISW we have used KGs to integrate several applications together and then use deductive reasoning for semantic search across these applications. Or to identify anomalies and take action. The example I just gave before on anomaly detection and repair is a good indicator of how KGs can help connect data from disparate systems together and quickly extract intelligence from their conceptual relationships and mappings. So, shaping AI and machine learning in industry also demands different approaches towards industrial grade, right? So, what is it for you? So what is industrial grade from your perspective and how do you picture trustworthy AI? Well, um. Now you're in my wheelhouse because I can talk about this topic for, uh, for hours, but I try not to do that. I'll, I'll try my best to be brief, but, you know, uh I, I actually did a podcast with a colleague from Google about this topic, but trustworthy AI for industrial applications must be functionally robust, abide by legal mandates and follow ethical principles. I, I talked about these bias issues before, and those could have legal, ethical, and robustness, ramifications. And now, here, I'd just like to touch on another sensitive topic in the short time I have to focus on the ethics issues. Historically, the question of ethics is not raised in computer science and software development as much as It does in medicine, biology, psychology, anthropology, etc. The main reason is that up to now, human subjects were not viewed as much affected by software as these other traits. However, With the advent of AI models that we've been talking about, especially we at the IISW need to pay attention to Ethical issues because first, AI models could create harmful bias. We covered this before. Second, AI models could be weaponized. Third, AI models could allow creation of alternate reality. And fourth, AI models can replace masses of humans as part of automation. This attention to ethics is a huge change, and in my view, software developers are not ready to deal with it on their own. Um, I can give you a perfect example of Joe Redman. He was a tremendous researcher that when he was a graduate student in 2016, he developed and pioneered object recognition algorithm that's called YOLO. You only look once. He revealed not too long ago that he has stopped doing computer vision research altogether because of its military and surveillance applications. That's a shame. Now, some people may protest the intrusion of ideology into engineering and software development when we talk about industrial AI ethics. But we don't want our developers to deal with these issues as a sort of personal moral crisis that they need to deal with on their own. I think a more organized ethics process is needed, and I bring this up constantly during the AIS AI console meetings. I'm, I'm a member of that group. And we will see what happens, but I truly believe that more attention needs to be paid to ethics. Of course, we want to be ethical, but also to remove this burden from developers and engineers that we have so that they know what they should do and how they need to behave to match the best practices that we have. Yeah, a humongous challenge. And actually YOLO, right, you know, every, every, I guess, PhD, you know, or at least interested ML geek in Kubidivision, right, has fine tuned some kind of YOLO, right? At least used it, right? And I found it also quite straight and, and very honest to say, OK, he will not continue and pursue this, you know, because of its implication. And this resonates also, I don't know if you have seen the keynote on, on laws. Nips um conference, uh, December of Charles Isbel, he gave this key, keynote on, you can't escape hyperparametters and latent variables, right? So from a software engineering and argued a bit, uh, really nicely done video, actually, also that, you know, researchers just need to have broad broadened their scope and also their view on the implications their grounding and their contribution have in their application. ations because the most, most researchers are like, I, I just did the glossifier, right? I don't know what they're doing, the bad guys or the good guys, right, with that. But, um, as, you know, these technology come more pervasive and important, it's also, you know, inevitable, you know, that you also consider that something will go wrong or may go wrong, right? And that you need to design for these kinds of aspects and need to, when you contribute um to this field. Also being aware of, you know, this, this application scenarios that's going on. Yeah, Julie, this is, as I said, we can talk about this for hours. I actually struggle with something myself. …