Jun 13th 2021 | E056 From PhD In Graph Theory To Google Brain Robotics
Dr. Krzysztof Choromański is a Research Scientist at Google Brain Robotics New York and an Adjunct Assistant Professor at Columbia University. Prior to joining Google, he completed his PhD at Columbia University working on structural graph theory (Ramsey-type results, in particular the Erdos-Hajnal Conjecture). His interests include structural & random graph theory, robotics & reinforcement learning, quasi Monte Carlo methods for machine learning, Riemannian optimization and more recently, attention-based architectures for sequential data with applications in robotics (memorization, lifelong learning), bioinformatics and vision.
The episode was recorded on June 12th, 2021. This material represents the speaker’s personal views and not the views of their current or former employer(s).
Krzysztof’s LinkedIn profile: https://www.linkedin.com/in/krzysztof-choromanski-288bb012/
Natalia Bielczyk 00:10 Hello, everyone. This is yet another episode of Career Talks by Welcome Solutions. And in these meetings, we talk with professionals with fascinating career paths who are willing to share the life hacks with us. And today, I have great pleasure to talk to Dr. Krzysztof Choromański, who is a research scientist at Google Brain Robotics in New York, and an adjunct assistant professor at Columbia University.
Natalia Bielczyk 00:34 Prior to joining Google, he completed his PhD at Columbia University working on structural graph theory. His interests include structural & random graph theory, robotics & reinforcement learning, quasi-Monte Carlo methods for machine learning, Riemannian optimization, and most recently, attention-based architectures for sequential data with applications in robotics, Bioinformatics, and vision.
Natalia Bielczyk 00:59 Thank you so much, Krzysztof, for attending our episode today and for accepting our invitation. Great to have you. Thank you so much for finding time in your packed schedule to be with us today. And I’m very curious about how your life story and how your career story panned out so far. And what were the bottlenecks, and what were the important decisions you had to make in the process? And it all sounds very exciting. I’m dying to know more about your story from the very beginning.
Krzysztof Choromański 01:34 Good morning. It’s a great pleasure to be here. I mean, I would say that in terms of, my story, my research story. It really all started, I would say, like with the IBM Research internship that I had some number of years ago. It was in 2008, so quite some number of years ago. But I was a student at the University of Warsaw in Computer Science and Mathematics Department.
Krzysztof Choromański 02:16 And me and my friends, we got the chance to work as interns at IBM in New York. And that was the first time I saw New York; I fell in love with New York. And I also started thinking very seriously about applying for PhD at universities in the United States. Not necessarily just like in New York. But obviously I had in my back of my head that while they are great universities in New York, Columbia University, NYU.
Krzysztof Choromański 02:52 That was like one of the I would say, like turning points in my career. I applied, I got admitted to a couple of universities in particular, NYU and Columbia University. And one of my friends the dilemma was which one to choose. And the funny anecdote that I can say that not many people know about is that when I was admitted, like in particular at NYU and at Columbia, I was invited to see the campus.
Krzysztof Choromański 03:27 And so, I traveled to New York and on my way back to Poland, I had the seat next to a student at NYU and we started some random conversation. And I explained that I’m just coming back from United States that I’ll start my PhD. It was kind of funny because I told her that I’m still thinking about which university to choose. And she said, ‘I remember it. You should never even think for the moment. Columbia University, it’s Ivy League. This is like the best choice you can make.’
Krzysztof Choromański 04:08 It was kind of funny because she was from NYU. I mean, both universities are really great. I don’t want to say that I listened to her. I was kind of like convinced at some point that Columbia will be a better choice. Even though as I said, it’s hard to say. I’m sure NYU will be also great. But that was an interesting story. And then, the PhD started and I had the pleasure to work with Professor Maria Chudnovski on graph theory.
Krzysztof Choromański 04:39 I met her, even before I started PhD, during my summer internship at IBM. And I would say it was really a great time. I would say, the PhD was too short. I complete it in four years. And my advisor for that, like, I don’t need like more time. But I was like a little upset because I really enjoyed my time as a PhD student. Maybe, you know, not so much, in terms of financial stuff but I really like living in New York.
Krzysztof Choromański 05:18 And that was just a really great time, like when I could focus completely on research. Thinking about, I mean, proving theorems I and stuff. It was a really good time. And obviously, enjoying life in New York outside of the research scope, too. That’s kind of like, you know, really great years in my life. And while I was working on what’s called the eldritch handle conjecture, which is, it still open.
Krzysztof Choromański 05:51 We managed to prove a couple of interesting results, but the conjecture is still wide open. And frankly, speaking, I still work on it, like as a kind of, I would say, like side job. I think that would be like great to prove or disprove it. And so, but that was like really great time. And then, I would say the next important point really, the decision to make which path to choose after completing my PhD.
Advertisement 06:23 In welcome solutions, we help highly educated professionals with navigating towards their dream careers. We offer intensive career rotation courses, combining self-discovery with practical information about the job market. We also work on our own educational materials, such as books and self-navigation manuals. Our new tool, the Odyssey test, will help you discover what your natural way of creating value in the job market is, and which work environment will fit you best. If you’d like to stay in touch with us and receive monthly updates, please subscribe to the newsletter.
Krzysztof Choromański 06:52 Should it be academia? Just to continue working on graph theory. And then there are a couple of options like looking for a postdoc. I was looking for president; I got some offers. The other option was Google. The way it started is, I’ve got internship at Google, in last year of my PhD status. And again, like a very interesting, funny story. Because I got the call from Google from the recruiters.
Krzysztof Choromański 07:26 And it just turns out that one of my colleagues from the University of Warsaw recommended me. That’s kind of like how I got that call. And I got admitted to this internship. Went to San Francisco for a couple of months. I was working in Mountain View, but living in San Francisco. And that was I think, the first time when I saw the West Coast. And so, it was a very good time for sure.
Krzysztof Choromański 08:01 And I really enjoyed working there. I was working on graph theory, too. And this was kind of, just like more applied. And then, as I said, it was a big decision to make whether to go. I mean, just continue academia path, or I’ll think about industry joining Google, for instance. And I decided that Google could be the best option. I think, I kind of like felt that this time, when machine learning will be very big is coming.
Krzysztof Choromański 08:39 And at that time, it was like, I think about 2012. I mean, people were very excited about what ML can do. But it wasn’t like what’s happening right now, it was still kind of like a little bit a pre-revolution. You can kind of sense it in the air that, you know, it will be like a very big field. And I think I kind of understood it like early on that, you know, me going into machine learning can be still able to use all of my skills and also, like graph theory.
Krzysztof Choromański 09:09 But there are lots of other opportunities that opened and it’s kind of like just a very wide area where I can use a bunch of skills. And it just turns out that was exactly what happens. And so within like a few years, I learned machine learning because my PhD wasn’t in machine learning. I learned it at Google with really like amazing colleagues that I had the pleasure and I have a pleasure to work with.
Krzysztof Choromański 09:38 And this is where I am right now at Google Brain Robotics Team. I really enjoy it a lot. Working with robots is something that I wouldn’t like expect a couple of years ago. I mean, definitely not when I joined Google, but it’s really fun. I mean, it’s making all these theorems, like all this mathematics work, really in practice. I mean, when you see that the robot is doing what it’s supposed to do, that the algorithm works. It’s just great satisfaction.
Krzysztof Choromański 10:18 But I keep links with Columbia University, as I said, I had a really amazing time there. As you mentioned, I’m an Adjunct Assistant Professor at Columbia. I really enjoy working with students. You know, I don’t consider myself as a professor who kind of like completed a PhD years ago. Like for me, you know, I’m working with students; with young people. It’s something that I enjoy a lot. And they have a very fresh like point of view, because they don’t know that certain things are impossible. It’s their heart.
Krzysztof Choromański 10:57 And that’s the right attitude in life. And that’s why you know, I keep doing this. I get the, you know, pleasure to work with so many different people. Google is also happy with me being involved, like, and all these efforts at Columbia University. That’s kind of like in a nutshell, you know, my story since 2008, where I went for internship at IBM. and that really changed my life in many different ways.
Natalia Bielczyk 11:28 Fantastic. have so many questions now. The first question would be about like, … Because since I graduated from the same university; we were both students of University of Warsaw. And I’m still in touch with many of my former students. I sometimes wonder, how does it happen that some of the smartest people among the pool of students that we were studying with never even graduated from Masters’.
Natalia Bielczyk 12:02 Not mentioning about PhDs, not mentioning about building careers. And some of the people landed at Google Robotics or Columbia University and are building great careers right now. What do you think made you excel? And what do you think is the secret why you’re going forward so quickly in your career, as compared to let’s say, the general population? Other people who, let’s say, are also smart in terms of their capability to understand mathematics yet they cannot really work out the career like you do.
Krzysztof Choromański 12:42 I think there are lots of different factors here. I mean, one is simply. You know, the priorities that you have in your life and people just choose different paths. And I will give an example. One of my friends from the times of Polish Mathematical Olympiads. He was extremely talented mathematician. He qualified to International Mathematical Olympiad, like the Polish team.
Krzysztof Choromański 13:17 And I got the pleasure to spend time with him when we were all preparing for the international competition in Polish camp in [13:26 – unintelligibly. And I don’t want to mention the name, it’s not that relevant. But the point is that I remember, he won I think like a silver medal at IMO, if I remember well. And I know that at some point, he just kind of quitted mathematics.
Krzysztof Choromański 13:48 I think he was always also very good in languages. And I think he … I don’t know if he learned Chinese or Japanese. But he went like to Asia and you know, he kind of quitted mathematics like very abruptly. And you know, these are the decisions that people make. Because as I said, like mathematics like research is not everything. There are other priorities people have in life. Sometimes they just decided they want to do something very different.
Krzysztof Choromański 13:48 The other thing is that, even if you want to continue this research path. For instance, the PhD life is kind of like different than what you have in competitions. Competition is like you have a couple of problems that are, you know they are solvable. And because otherwise you wouldn’t have them again in the competition; and couple of hours to solve them. When you work on the PhD problem, I mean, the problem for your PhD thesis.
Krzysztof Choromański 14:47 Sometimes you even don’t know whether how hard or how feasible it is to solve it. And that’s additional risk and you need to work on it not for 4 or 5 hours. You need to work on it like for a couple of years usually. And some people simply don’t have patience. It’s just a different like style that they have. It’s more like focusing on something you can really solve quicker, more problems than a single problem that we need to wait for a long time.
Krzysztof Choromański 15:22 And the best example, I think I was watching interview with a great British-American physicist Freeman Dyson, who said that was his problem with the PhD system. He cannot really work for a long time on one problem that. You know, if something doesn’t work, he wants to jump to the other problem. But when you are in this PhD format, then you want to spend more time on one problem. If you can jump, but you shouldn’t jump every couple of days.
Krzysztof Choromański 15:56 You know, these are all these factors why some people don’t even try to go for a PhD. Just a different style. And, you know, as I said, like, I met so many amazing people in my life with amazing stories and doing like different things. I’m not the last person to say that, you know, there is one particular template that one should follow. At the very end, it’s what makes you happy. I mean, you go to Asia. You learn about Chinese or Japanese but you have fun learning the culture, go for it.
Krzysztof Choromański 16:36 If you want to do great research, go for it too. In my case, I think from a very early childhood, I had imprinted in my brain like all these stories of Nobel Prize winners. And you know, these amazing discoveries that they made. And I wanted to, you know, do something that others will remember me for. And that was driving me to do this research. Even though there are lots of obstacles, and it wasn’t the case that everything was smooth and easy.
Krzysztof Choromański 17:17 But as I said, people have just very different driving forces behind them and so many different factors, like we all hear that. Now, that’s how we have all these different stories at the end. But I know, like amazing mathematicians, amazing researchers, lots of my colleagues from competition times, that ended up in many different places. And I’m in touch with them. I know they enjoy their time and you know, that’s the most important thing. Make smart decisions in your life that make you happy. That’s what I can say about this. Okay.
Natalia Bielczyk 18:01 That’s a very humble way of thinking of yourself, I think, because it’s clearly that you developed a great career. I also am curious about your current job, of course, because for many PhD graduates and not only to a lot of people … Google, first of all, Google is a great brand. Google is at the frontier of IT industry in many different aspects. And I’m sure that Google Brain Robotics is at the frontier of robotics, and also partially neuroscience as well.
Natalia Bielczyk 18:41 I’m curious if you can disclose anything. Any of the projects that you’re working on or that Google Brain Robotics is working on at the moment. I’m sure that this is one area of interest in terms of potential future career paths for many graduates of computer science and also neuroscience today. I would like to know a little bit more about what you’re working on?
Krzysztof Choromański 19:10 Definitely there is. I mean, there’s obviously some confidential efforts. But there are lots of things that are publicly available and these stuffs I can talk about. You know, I think that the main theme is giving the machine learning revolution, how we can adapt those techniques into robotics. That’s the big theme the standard, I would say, pre-ML. Robotics was mostly like control theory, and this way somehow limit it.
Krzysztof Choromański 19:49 I mean, you can do lots of things. But now with ML, there are new opportunities. The very big theme is, you know, whatever we could do with ML and the standard, normal setting, like how we can translate the success into reinforcement learning/robotics. And given that, you know, we have expertise in ML. And the question is, you know, how we can leverage like those skills.
Krzysztof Choromański 20:20 We have three robots in the company. And that makes a big difference like when we work on stuff because it’s not reliant just on the simulator. At the very end, we want to deploy it on the real hardware. And at the very end, the real problem starts when you want to deploy stuff on the real hardware. Like stuff that works in the simulator are no longer works that well, like on the real robot.
Krzysztof Choromański 20:49 You know, in terms of with the projects and the research scope, it’s very broad. I mean, there is very exciting research on locomotion. And there is very exciting research on manipulation. And also, general robotics, General Reinforcement Learning; like making advances in ML algorithms behind them. I myself, am very interested in adopting very recent techniques leveraging attention, transformers pipelines, in reinforcement learning and robotics.
Krzysztof Choromański 21:33 This is in particular, the class of methods of algorithms that, we simply call performers. A new class of Transformers that me and my colleagues develop in order to make regular attention algorithm most scalable, much longer sequences to match high dimensional data. And applying it in particular, in reinforcement learning robotics, even though this is probably not the most standard. It’s not the most canonical application of those methods.
Krzysztof Choromański 22:09 This is kind of, in a nutshell, you know, what we are working on. I would say that, you know, the most important part is, robotics right now is a really very broad fields that can use techniques from many different branches of mathematics/applied mathematics. There’s control theory, there is machine learning. Now, when we talk about vision, there is an amazing revolution and using machine learning for vision.
Krzysztof Choromański 22:46 Now, how this can be adopted in robotics? Like, it’s a huge opportunity there. And of course, there are many companies that see that. I mean, Google is doing amazing stuff and other companies. And academic centers, they understand that the potential of those methods. And so, it’s a very exciting time for all of us working on that stuff. Because again, I think people understand that it’s just a matter of time to have this new robotics revolution with those techniques that turned out to be very useful, and more kind of like standard pre-robotic applications.
Natalia Bielczyk 23:33 I would like to ask just one more question about these transformers. I think this is a new term for me. Explain to us, very shortly, what this is?
Krzysztof Choromański 23:45 Transformers is a new class of machine learning architectures that was proposed a couple of years ago. It’s a pretty recent method, I think 2-3 years ago, by researchers from Google. Some of them I think are Google brain, so my colleagues from my team. And what the idea is transformers are designed to process sequential data. If you think about your data as a sequence; sequence of tokens.
Krzysztof Choromański 24:19 Now what those tokens are, well, they can be anything. You can think about the post- standard application text data. Different tokens will correspond to different words, words are ordered. And there are some particular order of words, and they have content. The relationship between words. The other application would be images, where the tokens can correspond to patches or even individual pixels. And again, there is some ordering that you have.
Krzysztof Choromański 24:53 The idea of transformers is to somehow encode these inductive priors, modeling the relations between different tokens in the sequence. Prior to transformers, the standard approach to sequential data were recurrent neural networks. The idea is that there is one hidden state that tries to encapsulate the entire sequence. But the problem with this approach is that well, this hidden status pretty compact and initially encapsulates the entire history.
Krzysztof Choromański 25:24 The sequence can be very long, think about the text data or think about images when you have your millions of pixels. And because the hidden state is very compact, it’s also limited in its expressivity. The idea of transformers is to use the attention mechanism. The attention mechanism is a structure for modeling, explicitly modeling, the relationship between different tokens in the sequence. And this attention mechanism is combined with more standard mechanisms that were used previously in standard network architectures like multilayer perceptron (MLP).
Krzysztof Choromański 26:03 But this is really the attention that is a game changer. That’s roughly speaking, what transformers are. They try to explicitly model these relationships and it helps in order to get better quality models. And transformers took over strong machine learning, becoming the new state of the art for processing text data. I mean, substantially outperforming RNN.
Krzysztof Choromański 26:33 LSTMs and RNNs are no longer use for this data. Instead, people use transformers. But surprisingly, they’ve become very efficient in other domains, like working with images. People even start applying Kit in reinforcement learning and robotics that I just mentioned. Then speech is another setting where transformers are used. And so, this is like a very new set of techniques that somehow apply a bunch of different methods.
Krzysztof Choromański 27:05 But really this attention module, attention block is a game changer. And attention itself was not in machine learning for years. I mean, many years before transformers were proposed. But transformers were really the first architectures that utilized very efficiently this concept of the attention, where you explicitly model the relationship between different tokens and turn out to be extremely successful in the entire machine learning field.
Krzysztof Choromański 27:37 People are very excited about it. And I would say that almost every day, you see a couple of new papers on transformers; different transformers variance. It’s nice to see that because with those new methods, we can do much more than with standard techniques.
Natalia Bielczyk 27:57 Are transformers also perhaps used in any way by Google Translate? Because I was just observing how it must be improved in the last 2-3-5 years, like translating between, let’s say polish and English. Now, it’s almost perfect text. Just a few years back, it was quite laughable, but now it’s almost like indistinguishable from human text. And I’m quite amazed by that. I was just curious if transformers have some input there as well.
Krzysztof Choromański 28:33 One thing is in terms of, some confidential projects. I must be a little enigmatic. But you know, what I can say is that it’s definitely … I mean, big companies such as Google, understand the potential of transformers and especially in the context of processing sequential data.
Natalia Bielczyk 29:03 I’ll look into that. Okay, super. My next question for you would be, you know, since you very effectively balance on the edge of research in academia, and also in industry. I have a bit like a provocative question. Since you know, how research in machine learning and reinforcement learning in particular is done. Who they think is currently the leader in reinforcement learning research? Is it Google, or is the rest of the world?
Krzysztof Choromański 29:39 I mean, I work at Google and definitely my opinion, maybe not the most objective one by definition. I think lots of companies do really great research on reinforcement learning and robotics. And I think, you know, we are definitely very humble at Google. I mean, we understand that. At the same time, you know, I think I can definitely say that the research that is done at Google on reinforcement learning and robotics is, definitely the top research. I think, we are on the frontiers of modern reinforcement learning and robotics.
Krzysztof Choromański 30:38 This type of research is very specific, because it requires a bunch of different things. Of course, I mean, researchers that are experts in the field and this type of this feature. I mean, a lot of great researchers; also look at academia. But the other important thing is resources. And I think here, in general, in industry has an advantage over academia, in terms of having those resources. You know, we can come up with great algorithms.
Krzysztof Choromański 31:20 But in the field, like robotics and reinforcement learning, you know, it’s not just about, you know, having the algorithm to prove the correctness but you need to show that this stuff works. It’s very practical in many different ways. It’s more about thinking of Applied Mathematics. Or for me, you know, I think about it a little a bit like, physics. I have a very warm feelings to physics.
Krzysztof Choromański 31:47 I myself, was in a physics International Olympiad, member of the Polish team. You know, our work in robotics/reinforcement learning, I think about it a little bit like a physicist. Like where you need to come up with a really nice theory, nice model.
Krzysztof Choromański 32:04 But then at the very end, you want to get the Nobel Prize and stuff, like needs to be verified experimentally. This is very important. And if the model, you know, is beautiful, mathematically elegant, but turns out not to work. Nobody really cares about it. I mean, you can still write the paper about the beauty like of mathematics and stuff., But really what matters, it’s at the very end, whether it reflects what nature is doing.
Krzysztof Choromański 32:33 And it’s somehow similar in machine learning is that you can come up with amazing algorithm for learning your policies of your robots. But at the very end, if it doesn’t work on the real hardware, then it’s not really that exciting. I think that, you know, industry has this advantage over academia, that there aren’t these resources, compute resources, that are needed in order to train those models and it just takes time.
Krzysztof Choromański 33:05 And of course, as I said, it’s the case of Google that we do have resources. And also, the case of other great companies that do have them and leverage them effectively. And I think academia understands that the time has changed a little bit. I mean, if you work on abstract mathematics, you don’t really care. All you need is a piece of paper, a pen and coffee.
Krzysztof Choromański 33:36 Paul Erdos had this nice quote that, mathematicians are machines that are transforming coffee into theorems and proofs. If you work in robotics and reinforcement learning, in general, in machine learning, you need a little bit more than a coffee. It’s still good to come up with stuff at the very end, this needs to be verified. And that’s why, you know, these top academic institutions, they understand this.
Krzysztof Choromański 34:00 And they are engaged in collaborations with industry in many different ways, by sending students for internships by taking part in joint projects. And I think Google again, has a very unique position here. Because for instance, on the East Coast, we have lots of great academic centers Columbia, Princeton, Harvard, MIT, and we collaborate with all of them.
Krzysztof Choromański 34:38 We have interests from these institutions. We have a project with those institutions. It’s just understanding. I think that working in symbiosis here, it’s really the best thing that can be done. And as I said, academia also understands that having that potential luck of really very smart students, what’s needed are resources. And those resources are obviously more limited, if you work in academia than if you work an industry.
Krzysztof Choromański 35:17 I think, getting back to your question. We, as a company, are doing top research in reinforcement language and robotics. But as I said, we also understand it’s a very hot area. Lots of other companies, institutions that do an amazing job and we admit it. We keep humble about it.
Natalia Bielczyk 35:41 This sounds like Oscar winners speaking from the stage. You know, that’s like humble bragging, I think.
Krzysztof Choromański 35:47 Maybe one day I will. I mentioned people like changing completely careers. Maybe acting is something that is waiting for me in the future.
Natalia Bielczyk 36:04 Maybe you can keep a poker face when you speak like this. Maybe you have a good luck in the industry. I would like to add something to this. I also feel that there is this integrity problem in academia, that is so individualistic. And so focused on individual success, that this makes it hard to building new software, new complex algorithms on top of what is already being done.
Natalia Bielczyk 36:34 Because these are complex projects and these cannot be done in isolation. From what I know about Google, and the way Google even hires employees. Even at the interview, you are encouraged to write a code that is understandable, rather than a code that is the most efficient and the shortest in terms of the number of signs. There is like a huge focus put on collaboration and on communication.
Natalia Bielczyk 37:03 And this is not true in academia, like every PhD student starts from their own world from their own project. And there is just so little focus put on collaboration, that even if you are a part of a project, you still have to think of your own CV at the very end. I think in mathematics, that might work because you only need one brain to prove a theorem.
Natalia Bielczyk 37:27 But once it comes to complex, groundbreaking projects like building new groundbreaking concepts in IT, then, of course, hundreds or 1000s of people have to collaborate. In this case, I think Google’s workflow and way of approaching project just is much more functional than what we have in academia. That’s one thing.
Krzysztof Choromański 37:51 I think that, definitely, in terms of this aspect of the teamwork. I think understanding that in fields like machine learning, it’s usually not the success of one person. Because the scope is huge. At the very end, in order to be successful, there needs to be a team of people working on stuff. It’s just infeasible. It’s not, you know, when I work on the erdeskainal conjecture, sure, maybe tomorrow. I will come up with the proof or disprove it.
Krzysztof Choromański 38:34 And that would be like, you know, a single person team and I will have a paper with the complete story. In machine learning, it’s almost never the case. I mean, as I said, the scope is just too large. There are so many things that needs to be done in order to verify the algorithm to make sure that the method really works. And obviously, in Google and other companies, they understand this. It’s important for us to understand that it is a team work.
Krzysztof Choromański 39:13 It’s important to kind of … I mean, people are. It’s fine to have more individual point of view, and people are just different. But at the end of the day, we are most efficient when we work together. And think in terms of academia, it’s different. Like one thing is that, you know, there are lots of projects that don’t require such a big scope. Lots of students working very individual, if you’re working on something like abstracts.
Krzysztof Choromański 39:44 Even if you work on machine learning, but you are a theoretician, like you’re proving like the properties of the algorithm. You don’t necessarily need such a big engagement. I still think that you know, working collaborative, even when you’re at academia is definitely important. Even if you believe that, you know, your style is kind of more individual style, because you can simply learn what others did and not repeat other’s mistakes.
Krzysztof Choromański 40:11 But yes, it’s just like a little bit different flavor. I don’t know, you know, whether it’s necessarily a drawback of academia. I would say, it’s just that academia is different in many ways. You know, you can say that maybe it should be more teamwork. On the other hand, this style, it’s just historically that’s how lots of people worked.
Krzysztof Choromański 40:46 We had still big projects, but there are definitely many kinds of individual projects. And you know, people like doing also amazing stuff sometimes very individually. Academia is not industry, will never be industry, industry will never be academia. But you know, the fact that those two entities are complementing each other is good in my opinion. Because as long as people from academia and industry can talk with each other. And understand that there’s different strengths of different groups, and they’re just picking the best from each other.
Krzysztof Choromański 41:28 Then this is when it works really well. And I would say that, definitely that’s the case with Google collaborations, Google engagements with academia. And that’s why I also enjoy it so much working at Google because I can still be engaged with academia efforts. But at the same time, I think that I can do much more than I would be able to do in academia, if I want to make like breakthroughs in machine learning.
Krzysztof Choromański 42:04 Simply because all these other items that academia doesn’t have, that we have like in industry. And just keeping this balance is important. But yes, there are definitely like two very different entities. That’s for sure.
Natalia Bielczyk 42:25 I guess it works well for you to keep this balance?
Krzysztof Choromański 42:30 Yes. I think, I really enjoy also having some time at Columbia University. I mean, just like you know, the fact that I can work with students from Colombia. As I said, I mean just like a very fresh point of view. Sometimes when you don’t work on the problem for years, you just like seeing other solutions, see other opportunities. You simply don’t know that something is impossible. That’s what I mentioned before and that’s good.
Krzysztof Choromański 43:02 One thing that you know I learned in my life is it’s very important to respect other’s work. But you shouldn’t be, …. What is the word I should use? It shouldn’t make you feel that … You simply cannot do study because of others; others didn’t do that. With respect to others work, admitting the importance of others’ work is very important. But never will you think that this is what limits you. Others didn’t do that then you won’t be able to do that.
Krzysztof Choromański 43:48 Because that’s the right mentality. I mean, just not being scared that these big names didn’t do certain things and have the right mentality to make progress. I mean, otherwise you will always be limited by others; others would be an upper bound for you. And I think that working with students, most of them they have this really good mentality, when they don’t really think that, … They just even don’t know that this problem wasn’t solved for that many years.
Krzysztof Choromański 44:20 They just go for it, they want to find a solution. And that’s what, in my opinion, is critical in order to have a successful career and as a researcher and really make a difference where others did not. Always respect others’ work, but never make others an upper bound for you, because that’s never good.
Natalia Bielczyk 44:47 I also feel that there are lots of blind spots, both in science and in IT industry. Like when I was building my website, and I was building an aptitude test I was working on for the last two years. I also realized that even at WordPress, which is basically 30% of the whole pool of worldwide websites, there are still so many missing plugins. We had to create a few custom plugins to be able to host this project just because there was no infrastructure.
Natalia Bielczyk 45:19 And I could see how many businesses; how many technological issues are not solved that are basic. The fact that Google didn’t do something that doesn’t mean that you shouldn’t do it. because there are so many opportunities that are just overlooked by others; by big players by big companies. I always feel that if you have an intuition for a project you should give it a go, because there is a fine chance that it was just overlooked.
Krzysztof Choromański 45:54 I think that’s really the most important thing. It’s kind of like self-confidence, what I think is like critical hard work. But also, as you mentioned, if you have an exciting idea, go for it and see how you can exercise it. As I said, especially in machine learning, there are so many people working on so many different things. And you would think at first glance that it’s really hard to make a progress.
Krzysztof Choromański 46:37 Because almost certainly someone else somewhere else is working on similar stuff. But it’s not that bad as I said. The nice thing about mathematics is there are so many different ways in order to solve the problem. It’s almost impossible for people that are already experts in their field to investigate all these paths. There are always unexplored paths and some of them lead to these new revolutions.
Krzysztof Choromański 47:15 Mathematics in this way is unbounded. There will always be opportunities just like a fresh mind and that’s what I say. It helped me a lot. The mathematical background that I got from times of me being a student in high school, taking part in the competitions than a student at the University of Warsaw.
Krzysztof Choromański 47:41 And learning mathematics and understanding the beauty of mathematics and the power of mathematics, that’s something that helps all the time. Because as I said, it’s impossible for an individual entity to go deep enough into mathematics and explore like all the feasible paths. As long as you have a fresh mind and you are excited about mathematics, there is a huge opportunity for progress, always.
Krzysztof Choromański 48:12 And so, that’s my advice to everyone. Don’t be overwhelmed by what others did and think about what you can do. My opinion is that a mathematical background helps a lot. It doesn’t really matter whether you end up being a particle physicist, or you work in IT or you work as a professor in academia. The mathematical background is the core building block of your success. Especially like in machine learning, where right now we just rediscover matter mathematics in so many different ways by just applying it different.
Krzysztof Choromański 49:00 Fields that you know, you wouldn’t think about applying like graph theory into robotics some number of years ago and we do it right now. That’s what I’m happy about very early on. I fell in love into mathematics and it stays with me for my entire life.
Natalia Bielczyk 49:24 Not without the reason it’s called the queen of sciences. I’m still tempted to ask you one question. I know that we are tight with time, but if I still have a minute. I am intrigued by this concept of lifelong learning since it’s among your numerous interests. You mentioned that what also interests you is applying robotics to lifelong learning. I’m intrigued what that means and how can this be achieved.
Krzysztof Choromański 49:53 One way of thinking about reinforcement learning is you have an agent or a robot that wants to achieve a particular task and you try to train a policy. A policy for the robot or an agent that would be a recipe and then a given state of what the action should be taken. And usually, those actions are very reactive. What you basically react to is current state that you are at, in the particular moment.
Krzysztof Choromański 50:27 But not so much, you know, what happened before; many steps like before when you ended up in that particular state. The idea of lifelong learning is to make robots think like a little bit more like humans, when we try to accomplish certain tasks. There are lots of tasks where we can be purely reactive, where whatever we will do is really determined by the current state.
Krzysztof Choromański 50:55 But the most challenging tasks will require in particular, memorization. Require us to exercise the experience that we got potentially many years ago. And that is still very challenging in robotics, because it’s a big question how you can leverage memory; what memory is, how our memory works. Like, how we can extract from our brain the skills that are most relevant or seem to be most relevant to solve a particular task.
Krzysztof Choromański 51:31 How we can get back to things that we learned like long time ago, to use it in a particular situation. The idea of lifelong learning is to make sure that robots can leverage those skills. They can learn from the experience, not just short-term but long-term and can use those learned skills not necessarily right now, but later on. And efficiently find it and use that memory and solve this task.
Krzysztof Choromański 52:08 In order to do that you also need to learn all the time. That’s kind of like roughly speaking, you know, what this term means and it’s very challenging. Because it really attaches these fundamental questions about how our brains work and we still don’t have answers to most of them. We say that the human brain is still mostly a mystery, why this is the case that we operate in such an efficient way.
Krzysztof Choromański 52:09 I have a one-year-old daughter and I see every day how she learns new stuff. And she can generalize from just a couple of examples. Machine learning algorithms needs millions of examples. The models might need to be trained on huge amount of data in order to get good quality. She can operate in this very heterogeneous environment, when most of the things she seeing for the first time and she can do it efficiently.
Krzysztof Choromański 53:15 And she’s stored those memories; she can leverage efficiently. And she trains her brain, pretty much non-stop from all these experiences, from negative examples from positive examples. If you think about it, it’s very efficient like an agent, reinforcement learning agent operating in a completely new environment. You know, the Holy Grail of people working in robotics is to somehow emulate the way how robots can be trained.
Krzysztof Choromański 53:49 But of course, it’s a big team and nobody really knows right now, how to do it efficiently. We are still in a very preliminary phase. It might be the case that it will require completely rethinking the way how we do machine learning. There are experts saying that the way how we train this big neural network models right now is simply wrong.
Krzysztof Choromański 54:10 Because we cannot generalize so easily, we need massive number of examples. There might be the case that it will require some revolution in machine learning. But that will also imply a big revolution in robotics. This lifelong learning, it’s really an idea that we should rethink the way how we train these agents. Because that’s how we as humans are operating in this complicated world.
Natalia Bielczyk 54:44 Perfect, perfect answer. Thank you so much for this. We have to finish right now. If you could in one sentence say something to all these young people who would like to become Googlers one day.
Krzysztof Choromański 54:58 What I can say is just do what you are passionate about, then definitely. You know, if you are interested in companies like Google, there are so many research teams working on so many different things that you will almost certainly find a topic that works for you. I would say, you know, my one advice. If you want to be successful, or even like you know, think about like you have a particular plan, don’t think too much about it. That’s kind of my experience.
Krzysztof Choromański 55:40 I didn’t really think about joining Google at some point, I got the call. And it all started like with the internship. Be a focused on your passion, focus on the stuff you are good at. And the other things will just naturally come. Don’t overthink, don’t think too much about your career. You have like a really good career at the end, if you are passionate about this subject; and that’s really critical.
Krzysztof Choromański 56:10 Because it can be in the great company, you can be working with great people. But if you don’t have fun on what you’re working on, it wouldn’t make sense. On the other hand, if you have fun, you know, despite all the difficulties that you will have. Because each of us will have in your life, that will be something that will be always a driving force.
Krzysztof Choromański 56:34 Maybe, you know, it’s not the most satisfactory answer. But I would say, you know, you want to think about a career at Google, like other companies. Don’t think about it too much. Focus on your passion and things will go naturally. That’s my advice.
Natalia Bielczyk 56:48 Perfect, perfect. Sounds convincing to me. I fully agree. I mean, ever since I focused on my passion, I feel much happier. And I feel like I’m going in the right direction. I couldn’t agree more. And thank you so much Krzysztof for being with us today. Thank you so much for your deep insights. And I wish you all the very best, I hope to hear more about your career in the future.
Natalia Bielczyk 57:14 I will be tracking your success on LinkedIn and on other media. And I hope to see more of your fantastic achievements. And yes, thank you so much once again, for being here today. And if you guys who are watching this episode would like to ask Krzysztof any questions. I think you can reach out to him through LinkedIn, that’s probably the best way. And of course, if you’d like to get more of this type of content, then please subscribe to the channel and see you next time.
Krzysztof Choromański 57:44 Thank you so much. It was a great pleasure and talk to you soon.