A friend of mine gave me Pat Winston's AI textbook when I finished high school. And although, as you know, that's mostly about symbolic AI, I said I wanted to do AI. After that, I did my undergrad in Eastern Europe, in Romania. There was one AI faculty was an expert in expert systems. I tried to stay informed about advances in AI. I tried to follow as much as possible. This was pre World Wide Web. I tried to buy books in AI and around my fourth year of undergrad, there are five years of undergrad in Romania, around my fourth year, I started learning about neural nets, and I implemented Backprop myself single AI neural net. I also implemented some recurring neural nets myself. Really simple with one-two nodes. but I continued work on expert systems because my professor was working in expert systems. My dream in AI was to build a system that is able to solve geometry problems. And as you know, it is still, an open question. So I'm talking here about the synthetic geometry problems, which typically have a logical answer. This was my work. I didn't do any research at that time. I was an undergrad. And then I got in touch with Tom Dietterich, who ended up being my Ph.D. advisor. I contacted him. I said I wanted to do work in machine learning, and I worked with him at Oregon State. So, he was very influential in my career. He was also a person in the mentor who exposed me to all advances in AI and also computer science. I'd like to emphasize that in AI We are computer scientists. Above all applied computer scientists. And, it helps a lot, although, at some points, it may not seem important. It helps a lot to understand the fundamentals of the theory of complexity to have a good understanding of how operating systems work, how computers are architected because then it's gonna be easier for us to build good applied solutions. That's the story of how I got involved in AI. Incidents and experiences. As I said somebody buying me Pat Winston's book in a time where there was no Amazon to search for AI books. Also, I think it was everyone who I mean if you have a passion in an area, you as an undergrad or is an advanced high school student or a junior graduate student should be able to pursue it and start to do a lot of work on your own. As I said I was in Eastern Europe, but I bought myself some neural nets books and some AI books with very little money I got from my scholarship. Basically, it's gonna be a lot of work. Our field is computer science and the product of computer science is code. So remember we tend to see now a lot of people trying to talk about the AI but forgetting that in our product is code. So I recommend that everyone embracing any field of applied computer science to have good coding skills and good software engineering skills. I tried to understand them as much as possible. I'm not seeing myself as a hacker. I'm not in computer science because I really enjoyed the code. I like to code because I want to see a solution in my applied field. That's in short my story.

As everybody will learn as you advance in your career that you don't have the luxury that you have in graduate school to just work on a project or a small set of projects. First, I have to set an AI strategy for Boeing. And, as you know, and as everybody knows, probably in 10 years, it's gonna be clear what the frontiers of AI are and what are the problems in AI. Right now I work a lot in educating others who are not AI people, both engineers, but also managers and executives on what's possible and what's not. Nowadays, a lot of our work and a lot of our results are taken and interpreted and discussed outside by people who are not AI people. Also, it's very easy for somebody who doesn't work in the field to interpret a result in a way so that he or she thinks that some other solution is easy. For example, the steps from detecting objects and images to understand interactions between the object to really understand images. Some people think it's a trivial step. No, we all understand that. We need a totally different set of approaches for that. So you know, just that as an example, I have to explain it to different people, for example, who have problems in computer vision. Also for example, if you have enough data on detecting objects and images, you and I and everybody know that it's a totally different problem than let's say detecting anomalies in images in which you really do not have a lot or maybe you don't have at all images with some classes or object. So, AI Strategy, trying to plan projects and to discuss and to estimate what kind of effort is needed for a project that's quite a bit of my roles. And also, technically, I always keep one or two projects for me to work for hands-on. Now it's somewhere between 20 and 40% of my time. I try to make it as large as possible. I have two projects right now on which I worked myself on and then here's another rule that we researchers in machine learning and AI should have is the problem formulation, which translated in technical terms is actually the problem abstraction. People in the real world don't have machine learning problems. They have optimization problems or predict which airplane will be produced on time or not, that has to be mapped into a machine learning problem or into an optimization problem that we can address. Sometimes people don't care that we have enough data or a lot of data or a little bit. It's up to us to understand how much data we can collect, how much data and what kind of features and what kind of noise they're affected by and how non- stationary the data is and so on. So that problem formulation, I think every student, every senior graduate student and every AI person who steps in the career should be able to sit down with somebody who has minimal knowledge in optimization, machine learning and so on to formulate a viable plan to solve the problem. The reality is that virtually no practical problem is actually just one classifier or one step or one policy that can be learned through reinforcement learning. Our solutions again it's code. We have to architect our solutions. Most of the time now we have in many places, we have learned models and then we have some optimization problems, and then we have a lot of code that do standard things. So this architecting falls into our hands, and that's sometimes if you do not have a good understanding of the tools that are available, that can make or break there. That's the difference between a good group and they're not so good group in addressing problems. We have had people internally and externally, people groups of six trying to address a problem for half a year, and they weren't enabled to. And then once you had somebody who understood the nature of the problem and the nature of the uncertainty in the data and the nature of uncertainty in the real world to address the problem, one person in three months. So, that's my role. And, I have to make this happen. And then we all know that we have to make it happen with the people that we have around.

The preference is driven by technical advantages. Nowadays, if somebody asks you what tools, probably they're thinking we are using PyTorch or Tensorflow or something. And obviously I and my team, we use all these tools. We also write, quite a bit of code and algorithms ourselves. But you know, for deep learning PyTorch lately quite a bit of PyTorch, TensorFlow. I want to point out one thing in AI we use and I'm looking more and more at probabilistic programming. In TensorFlow, you have TensorFlow probability. So if you have an idea on how to model the uncertainty in the task and you can exploit that for example, you know how to model the certainly, but you want the actual probabilities, you can learn those from data. It's actually pretty useful. We started to use probabilistic programming tools also for selecting a hypothesis. So let's say you have two hypotheses learned by the same algorithms and they do the same on your test data. Which one would you prefer for the real world? Here's one thing that I care more and more as you know we teach that. We have training data and test data.