Lihi Zelnik-Manor: Industry picks up domains that are mature, that have short-term outcomes. Short term, academia has the option and also I think the duty to think long-term.
Gil Elbaz: Welcome to Unboxing AI, the podcast about the future of computer vision, where industry experts share their insights about bringing computer vision capabilities into production. I’m Gil Elbaz, Datagen co-founder and CTO. Yalla, let’s get started.
So welcome, Lihi, so it’s exciting to have you here. We have here with us today Lihi Zelnik-Manor, an associate professor in the Faculty of Electrical Engineering at the Technion and the General Manager of Alibaba Damo Israel Lab, Professor Lihi Zelnik-Manor holds a PhD and MSC with honors in computer science from the Weitzman Institute of science and a BSc. in mechanical engineering from the Technion. Her main area of expertise is computer vision. Professor Zelnik-Manor has done extensive community contribution serving as the General Chair of CVPR 2021, Program Chair of CVPR 2016, Associate Editor at TPAMI (IEEE Transactions on Pattern Analysis and Machine Intelligence) served multiple times as the area chair at CVPR, ECCV and was on the award committee of ACCV 18, CVPR 19. Looking forward, she will serve as the General Chair of ECCV 2022 happening in Israel. And as the program chair of ICCV 25 in China.
Wow. I didn’t know that they plan so far ahead.
Lihi Zelnik-Manor: Now it’s like six year planning.
Gil Elbaz: Okay, that’s exciting. I’d love to maybe know a little bit more about what it actually entails being the chair of these programs.
Lihi Zelnik-Manor: So it really depends on what chairs (you are talking about.) If you’re the program chair. then you’re responsible for the entire reviewing process, paper acceptance and the program. So it’s more academic. And being a general chair is more dealing with the operation. So we need to make sure that there is a venue and that people can attend. And there are hotels. So I think general chair is considered kind of, uh, distinguished, but it’s administrative work.
Gil Elbaz: So you’re the person that makes it happen. There’s actually been a lot of feedback on the actual review process in the past few years, with the enormous storm of people joining the field. And so many people trying to get accepted into these conferences as both a way into the industry and as a way to make an impact. But we see a lot of smaller iteration or increments on existing work. How do you think about that regarding like the way that the review process is done today? Is there anything that’s going to be changing in the near future?
Lihi Zelnik-Manor: So I think the computer vision and also machine learning, those fields have attracted a lot of attention from industry and we’re seeing many challenges. In my view, this is because the industry is not playing in the academic game. So the rules of the game have been set up for academia and they’ve been working very well in many fields for many years.
And all of a sudden there are new players and those new players have different goals, different agenda and different numbers. Things don’t add up. And also the field has grown exponentially very, very fast. So now then the process, it doesn’t work as well as it used to be, at least in my eyes and a level of professionalism, their review process is different. I don’t want to criticize it too much.
It’s hard, but it’s hard to find enough people that have the experience and the years of knowledge to do the review process. But when I was doing my PhD, PhD students did not do reviews. I remember when I was invited for the first time to do a review, it was like an honor, so I felt, wow, what an achievement. I got to do a review for a conference paper. And today, if you had one paper, you get to be a reviewer.
Gil Elbaz: Exactly. I did an oral presentation in CVPR 17, and then CVPR 18 they wanted me to review papers and I actually accepted initially. And then when I saw the papers, I felt that I wasn’t actually, I didn’t have the enough of the background in order to do a thorough review in a reasonable amount of time. And so I actually declined to do the review itself.
Lihi Zelnik-Manor: So this is very interesting, because I’ve been in the field for many years, and I still feel often that I don’t have the required background to do the review properly. And the area is just so big and so vast, and so many things, it’s very difficult to keep up.
So we are really in a problem that is not just yours, it’s everybody’s problem. And I do expect to see changes, but what changes, time will tell. So there, there are other areas where you look at I don’t know if similar, but somewhat similar events have happened. And those areas just shifted to industry and the industry lost its interest in the academic game. It’s a numbers game. There’s just not enough people to do the review properly, etc.
So industry went into the academia and then took this, area and got bored of playing the academic game and moved on. And then the problem was actually resolved by separation or the academia focusing on forward-looking their actions and industry focusing on the more near future. And this is how the problem was resolved. It was not the change in the review process, but rather the scale was changed back.
Gil Elbaz: Yeah. And a change in focus, which is interesting. The other big feedback that I hear a lot is that the industry players have substantial resources and teams and ML Ops teams you know, in Data Ops teams and all of that, that really gives them a very unfair advantage against the new master’s student that’s working with a professor on an interesting problem, but that might not be as mainstream or might not get as state-of-the-art results.
Lihi Zelnik-Manor: You’re getting into the philosophical point of view and I have my own point of view about the roles of academia and industry, and I don’t see this as a problem, what you’re mentioning. Indeed, it is true. So in industry, you have resources that you might not have in academia, but also in academia, resources that you don’t have in industry. For example, time. Time to think there’s actually even a movement called slow thinking. And some of my friends are members of these organizations. Please allow us to think slowly.
We’re moving a little bit beyond philosophy. So the way I see it, and I’ve worked in industry and academia, I just think academia and industry have different roles. Industry picks up domains that are mature, that have short term and short-term outcomes, short term impact.
So it doesn’t necessarily have to be two months. It could be one year. It could be three years. Rarely you will see in, in industry also things that are longer term. Some really rich companies can invest in that in, in really, really long-term. But most of the industry work, it adopts the short-term. Academia has the option, but also I think the duty, the responsibility to think long-term so deep learning that we’re all enjoying grew in academia. Quantum grew in academia, and there are many other domains that they grew in academia. And then it’s great that they shift on. And then, what I hope to see is that academia will find what it should focus on in those areas. Either stop working on machine learning, if everybody thinks it’s solved, I don’t think it is solved, or find those domains of academia should focus on. Those problems that need to be thought of. The new things. The next thing that will work in 10 years.
Gil Elbaz: Yeah. I agree. In many senses. I think that the one distinction is, it’s great to have you in the room you know, you have experience in both worlds and substantial experience in both worlds.
Lihi Zelnik-Manor: I agree with you, that in academia, you have substantial time and you can think about these things, but with the current pace of progress, which is exponential at this point, at least in the field of deep learning, what I see is that you actually don’t have time, even if you’re in the academia because you’re working on something and then there’s, a group working on something similar. They can actually publish before you reach…
Maybe you shouldn’t be working on that. I think the challenge comes. So for me as a professor, I can think in 10 or 15 years, and I’m not stressed and I’m not concerned about anyone publishing before me. That’s not part of my worries. Where does it become part of my worry?
I have students and I need to teach. So my other role in academia is not just long-term research. It is also teaching and students do come in and they want to do their bachelor’s master’s and PhD, and they want to become experts in a domain and they want to become experts now, and they want to control, learn, and digest and be ready for the industry life. And that is more immediate. And they want to have publications now and they want to have impact now. And this is where you need to kind of play the game. The fact that we are now kind of competing or playing in the same field game. This is where it becomes challenging.
So I think there was just the recently there was a paper that published statistics of the publication rates from industry versus academia in, I think it was, machine learning and not computer vision, but I’m not completely sure. And you see indeed that the industry is rising in the percentage, but you still see fair amount of papers coming up from academia.
You see the strong, universities still doing very well, still publishing papers at the top venues being accepted for oral presentations, etc. Probably a little bit more difficult for the weaker places, but the stronger places are still on the map and are succeeding to do that. Then you succeed to find those points where you do have your leverage. Okay. You have a startup. There are those dreams or the things that you want to do, and there’s the reality of what you can do with your resources and your capabilities, and you make a choice. So the same is true for a master or PhD student. You need to choose what you can do.
Gil Elbaz: So I definitely agree that it’s all about having a finite set of resources and trying to direct those resources a way. I also, I don’t want to put aside positive things that we see with industry being so involved in these kinds of forward-looking problems, even if they’re three years forward today, three years is a long time, much longer than what it was maybe 20 years ago. And so, you know, you can see that DeepMind with AlphaFold is now making, not only headlines, but real progress in real scientific domains that are beyond machine learning.
Lihi Zelnik-Manor: It’s amazing. For me, as someone who started working in this field in the early two thousands, this is super exciting. It’s so much fun. You know, I just talked to my son yesterday. I told him, I wish for you that the area you’ll go for will explode. Like what has happened to me. It is just amazing.
Gil Elbaz: It is amazing. I mean, on that note, let’s say in another 10 years, right. Do you have any industry in mind that you think has the potential to grow like computer vision grew?
Lihi Zelnik-Manor: Everything that has to do with machine learning I think will continue. But what I think might happen is that we’ll stop calling it machine learning. Machine learning will transform each and every field. So you’ll see revolution or a huge change in healthcare. You’ll see a huge change in chemistry. You will see huge change in manufacturing. You will see huge changes everywhere. And I really hope that it will see other, like, I think one of the spotlights right now, this is not me, everybody sees it, that quantum computing is getting a lot of question marks and attention. And you see the big corporates all have, quantum labs. This is often an indication that many people believe this would be the next boom, who knows. So we’ll see.
Gil Elbaz: So I think we both agree that AI is the future. Maybe to take you back into the past a bit, I’d love to ask a little bit about how the academic scene looked when you started out, who you guys were working with. And today, we think of AI as one thing, but I want to also provide a little bit of perspective to our listeners.
Lihi Zelnik-Manor: it’s a good question. I like this question because maybe because of ethical reasons. People are very concerned about AI consuming their lives. And I wouldn’t call it AI. I started calling it AI because everybody else does and then people understand what I’m talking about. For me this is just an algorithm. Deep learning. Okay. Transformers. Okay. It’s just algorithm. It’s math. It’s math. Exactly.
So if you go back, so there were people thinking about perceptions. Specifically about computer vision, people thinking about perception, human perception, machine perception. In the sixties, seventies, eighties, really before I was in the field.
So I can start talking more about let’s say around the year 2000, let’s start the proximity over there. The computer vision community was small. A few hundred people would attend the main conferences like 300, 400, I think ICCV 99 in Corfu. I would say 400 or 500 people. I’m not sure, I didn’t check. It was very, very much a community. You felt like you knew everybody. Very academic. There were already companies doing research and Microsoft Research was there.
Very academic, very university-oriented. And the focus in computer vision was mostly on geometry. So people who are trying to solve geometry. By the way, everybody knew that the holy grail is recognition. So if you look at the seminar papers, as I said, from the seventies or eighties, Wiedermann and Shimon Ullman from Israel.
You’ll see them talking about recognition, and thinking about it. But the actual publications, most of them were about geometry and people thought they need to solve geometry, and also were able to solve geometry. So it’s a mix again, of what you want to do and what you can do and the resources you have. People didn’t have digital images. Can you imagine working on computer vision without images to test on?
Gil Elbaz: That’s challenging,
Lihi Zelnik-Manor: That’s challenging, right? So, so in early 2000, we already had the images, right? But for example, videos, if you wanted to have videos on your computer, the video cameras were analog and you had to use a frame grabber, which was expensive. And you had know how to download the videos into your computer and opening a video file. I spent personally a couple of months on figuring out how to read a video file. So I had to code it in Linux using FFmpeg and tools. And so even the technical aspects were difficult at that time. But geometry was the name of the game.
I think he, you can look at the people like Richard Hartley and Andrew Zisserman that had wrote books and had a lot of influence. We’re working on that, but then very quickly around year 2003, first successes in recognition started to appear.
Gil Elbaz: Can I ask you to clarify a little bit about what you mean by geometry versus recognition?
Lihi Zelnik-Manor: Yes. So geometry was mostly about aligning images, figuring out where the camera is positioned, 3D reconstruction. If I see an object from two different viewpoints, was it possible to solve it? And it also had utilization in industry, like panoramas, and people were very happy to have panoramas on their cameras and 3D reconstruction people thought will be very, very useful.
Back then they thought it is essential for recognition, but even for recognition, if you look at David Lowe’s work. So David Lowe wrote one of the most seminal papers of computer vision, and maybe even computer science of all time. It was about SIFT (scale-invariant feature transform) features, an old school, computer vision paper. which is interesting when you asked me about the pace.
So the ICCV paper, which is the conference paper, was published in 1999. The journal paper- -back then people really cared about journals– came out in 2004. So five years between the conference paper and the journal publication. Wow. So the processes is to get the journal paper accepted were a little bit long. Five years is extreme, but, it would take at least a year or a year and a half to get your paper through the publication process in a journal like PAMI or IJC.
But if you look at David Lowe’s work before and after. So he was thinking about recognition and then he said, okay, what can I do? I can find the features if I defined them well enough, I will succeed to align the images. And then it won’t be able to align all the images. And through that, I will recognize that it is the same object because it has the same features and the features align. How can I recognize a face? I recognize features that are eyes and nose and mouth. And then if I aligned all the face images like, okay, these are all faces. So this was this type of thinking of using geometrical alignment and geometrical computation for eventually recognition as the goal.
Gil Elbaz: Interesting. Very interesting. Yeah. Just, it throws me back to the courses that I did on classic computer vision in the Technion.
Lihi Zelnik-Manor: Yeah. I think there’s a lot to learn from that then actually now you see, I see a lot of going back to classical computer vision, right? We reinvent the wheel. So you see methods like transformers that come out and it is much more correlated with old school computer vision. So correlation between image patches is a topic that the computer vision community has dealt with for a long time, somewhere between 2000 to 2010, yet that decade a lot of works on non-local means and patch based methods, example based method. And now transformers are leveraging that knowledge.
Gil Elbaz: Yeah. So you see it as looking to the past, grabbing some useful ideas, concepts, and then bringing them to life with today’s deep learning capabilities. Yes, exactly. And was there any hype back then? Was it a lucrative kind of field was it crazy on the other hand? What did they think about the field?
Lihi Zelnik-Manor: Lucrative. No, I wouldn’t say that. A very applied field. So we were working in applied field, but the success in the real world was very limited. So this puts you in a question, mark, why are you working on this field then? Several times when I was pursuing my career, I got the advice: “switch fields.” “That’s not the right field. You need to change to computational biology.” “Biology is the future.” I actually had some very wise people talk to me when I went out on my postdoc and try to convince me to switch to computational biology.
Gil Elbaz: Wow. We’re all happy that you didn’t switch. And was there really any focus on theory back then? Or was it mainly engineering focused, like solving problems?
Lihi Zelnik-Manor: There was always the mix. And today, it’s also true. , even if, I look at my own papers, we’re doing a lot of work on architecture search, so I have one paper which is very, very theoretical and it tries to look at the bounds of how over-parameterized networks need to be in order to converge during the training.
At the other extreme, I have very practical papers. the key ideas are just, this is a method that works very, very fast and very, very efficiently; use it. So I have the full spectrum because I like, I like to have the full spectrum and you see, I think it’s a reflection of what you see in this field.
There was always everything there, a good mix. And I think I would cite Vladlen Koltun on that. He gave an excellent tutorial a few years back at CVPR about what is research and what we should think about when we think about research. And he said, it’s a collaborative effort. So there are like many, many pieces, many small boats, and the boats should point in the same direction. And then the big one will win the championship and you don’t have to win the championship yourself, but you make a little impact and you go in that direction then together, it flows.
Interesting. Very, very interesting. Yeah. It’s been a real challenge to add theory into this industry. At least from what I’ve seen, there’s always been an attempt try to connect the engineering efforts with the theory. And, I’m definitely going to read up on your paper afterwards, but I think that in practice, we’re pretty far from understanding some very basic things, right? In a theoretical level, like how much data do I need, where is my network weak.
So first of all, the good news, is if you go back between 2012 and 2014, the deep learning revolution kind of started to boom. there were the first results in 2012 and people started talking about it and there were many skeptical people, many skeptical people that said, why does it work? We don’t believe in this approach. It has no substantiation and We have no idea why it works, but it works amazing. It works too good not to explore. And there was really no understanding of anything.
Now we have a lot of understanding. And how do you know that we have a lot of understanding. You see, so many companies have working products. This means that the engineers there, even if they can’t express it easily yet they know what they’re doing.
They know which architectures are choosing. They know the learning rate they know how to train it. They have this know-how, they know something, and it’s reproducible and it repeats in many different places. So we do know a lot now going to Siri, can you prove how much data you need? I can’t prove it, but if you ask any one of my engineers, how many images do you need? To solve a certain problem contribution. They will give you a ballpark number.
Gil Elbaz: It’s a very, very interesting and also counter intuitive kind of concept. And I agree with you like my engineers as well, right? They can say, oh, I need to, you can do this with like 10,000 images, get a decent result. Or on different problems, we need a lot of data. We need half a million, a million images just to, to make this work.
Lihi Zelnik-Manor: You’re a mechanical engineer like me, right? Yeah. So I don’t know if you ever heard this, but when I was a student, I remember in that you have to, this was one of the influential sentences I heard when I was, in my, at the Technion doing my studies for the bachelor degree. And this was in the course on, um, fluid dynamics dynamics. Yeah. So we were learning about that Navier-Stokes equations, and it always is, don’t know how to solve them. And we can’t really compute the flow. And we were all kind of perplexed.
And then the lecturer told us, don’t worry. Eventually you will go out there. And the blue collar worker, he will tell you that you need 10 inch hose. It’s gonna work. And trust him, he knows you need a tiny champ at 10 inch hose. So this gap between the theory or the research and the practice has always been there and mechanical engineers have always been the ones that have not waited for the physicist and the mathematicians to prove what needs to be done. They just acted upon it. And they just built systems and they build trains and they build airplanes and they did not fully understand the physics. They did not fully understand the aerodynamics. They said, okay, this works. And they just said they were doing experiments, rough estimations. So I feel very comfortable coming from that perspective. I feel very comfortable with building systems that work without fully understanding them. And for computer scientists, I see this is often difficult. My son is a physicist and he tried to take a class in deep learning and, and he hated it. So no logic at that, the things don’t add up. And I said, okay, yeah, there, there are no proofs, but it works.
Gil Elbaz: Yeah. There is logic. There’s just no solid proofs. It’s interesting. It’s like the information is hidden in the deep features in our brains in a way. And we act upon it, even though we can’t actually formalize it at least today.
Lihi Zelnik-Manor: True. But there’s more and more that is coming out. If you look, or it’s just progressing tremendously and people are succeeding and you see their paper comes out, for example, a new idea, like transformers very, very quickly people get it. They understand and they do better. They improve it very quickly. How can they improve so quickly? This is because they already have this knowledge on what works. So they have an understanding.
Gil Elbaz: Yeah. I think that it’s both, that they have an understanding and they’re also like various vectors to improve it. Right. For instance, with transformers, you want to move from like a quadratic, um, time to, to more of a linear kind of timescale and then, or a computational complexity. And then, you know, there are various ways to do it. And then you see people trying to make preformers and reformers and all of these other formers and it actually does work and it does make progress. I do think that there is kind of a surprise moment, many times though in deep learning, which is fun. But, but yeah, it goes back to the same point, if we can dive into the main topic for today, I think that this is something that’s not talked about enough. But it’s probably one of the more important subjects in our field. The challenges hiring and finding people and making people happy. we can connect this to the academia in a way, regarding your students How did you find, you know, the right student to join you? Is there like a specific set of characteristics that you always look for?
Lihi Zelnik-Manor: So I think in academia, okay. The way I chose people or in academia was a little bit different from what they’re doing industry, or maybe a lot different. It was very important for me to choose people that I just enjoy working with.
It is also important in industry, but there, this was the priority on that was very high and I wanted to enjoy it. I want to have fun. I wanted to, because we do a lot of one-on-ones and a lot of brainstorming and a lot of thinking. There needs to be a personal fit between me and everyone in my close team.
Usually when students would come in, I would tell them what my beliefs are, how I work and then send them to my lab and tell them, go gossip, go ask question, go ask others. How was it to work with me? And you choose if you want, this is the package. Choose. If you want to take this template. Those that said yes, we would launch a test project and we just see if it flows well.
So thing in academia, this was very kind of simple. And the reason I didn’t have like high bar, actually, most of them. We would just hit it off and get going. Very few. It never happened. And the reason is that I saw my duty two reasons. So one reason is that most people are, I enjoy working with apparently.
And just most of the students that came in through my door probably saw already both I was working on before they wanted to work on the same things and it kind of clicked and was fun. And the second reason is that I saw my duty to teach people. So this is my duty to teach people, whoever comes in. If, they have shortcomings that they don’t have enough background in math or enough background in programming or enough confidence. It is my duty to teach them. So I take them if they want to. And they believe in it, I take them. So this was my academic perspective.
Gil Elbaz: That’s a great approach, I think. And I know that there are some professors that have a very different approach that are very, very tough to talk to purposely create a lot of walls that you need to pass. And then they see if you want it enough, if you’re hungry enough. I appreciate that. And I think that in many cases, measuring people by the grades that they get in their classes is not necessarily the best way to measure them and the best way to filter them out.
Lihi Zelnik-Manor: Okay. So I’ll tell one last thing then maybe we talk a little bit about the industry. So when I started my career as a professor, I went to ask for advice, how should I choose my students? And people gave me different advice. So people said things like, look at what you said, like choose the ones that are best in math. And one of the recent physics, all kinds of stereotype things.
And then the advice I really liked a lot was choose someone that is very excellent in something unrelated to work. So it, can be excellent in music or excellent in sports or excellent in whatever. Choose those people, they will fit you better. So maybe this is a good guideline.
So how would we connected the industry? Different people have different preferences, different goals. I truly honestly appreciate other professors have different ways to choose because they have different goals than mine. I wanted to teach certain people, and this was my goal. And this is how I chose people that want to learn from me. Other people have different objectives in their lives.
And when I go to industry, again, I define the objectives. So I need to know what type of team I want to build. What are, of course, in terms of skillset, but also in terms of culture. And then these are my objectives. And then I select people who are working to those objectives. So it’s not the same thing.
Gil Elbaz: So what are the main objectives in industry, or maybe we can connect that to how you moved to Alibaba and maybe we can talk about the objectives and in industry.
Lihi Zelnik-Manor: Okay. So I’m going to say, I’ll give a brief introduction to Alibaba Israel and who we are, and then maybe there will be some context.
So I moved to Alibaba three years ago, late 2018. So the story of Alibaba and Israel is they acquired a company called Visualead. And this was the seed for Alibaba Israel. They team of six people joined with the CTO, Mr. Friedman joining. And he was the one who actually founded Alibaba, Israel. And then I joined a few months after that and we’ve been growing through recruiting and acquisition of Infinty AR. A big team joined Alibaba, Israel throught that. The CTO Matan Protter has also joined us. And, they’re still my people they’re in Alibaba. And then we started recruiting and we have now people like, Asaf Noy who leads a research team. And Nadav Zamir who leads the algorithm team. And we’ve grown into an R&D center that develops product.
So we are working on a product called Aliyun or Alibaba Cloud Drive, which is a hybrid version of you’d like something between Google Drive and Google Photos adapted to the Chinese market. So Aliyun plan is the Alibaba’s drive product. And our team is not doing it by itself. We’re doing in collaboration with two other teams.
One is more the front-end for the drive. And one that is doing the storage on the backend. And we actually also do end to end for anything that is related to photos. So we have actually, we have all the way, a broad spectrum of people from product manager, mobile developers, backend, front-end, infrastructure, and algorithm people, data team, and researchers.
Gil Elbaz: I actually didn’t realize the full scope that it’s not only the research side of things. It’s actually the full product being developed here.
Lihi Zelnik-Manor: Research is just part of the team. We’re doing research with more than two goals. So one goal is branding, which is great, but also knowledge actually our research team outcomes, going to the product eventually, but a bit longer term than the algorithm team that are working directly on the features, but we’re really an R&D center developing features and like image, search, or creating automatic albums from your personal photos.
And you can imagine there’s more need to design the application. You need to define the product and you need to program it and do the, again, the front end of the mobile and the there’s also web version. So also the front-end on the web backend grab algorithms into microservices and deploy them in Alibaba Clouds, multimedia platform, et cetera. So we do it the full spectrum. And if you asked me, how do you recruit, then you need to, so it’s about skills, but also need to find people feeding for, for each of those roles in terms of, what it takes .
Gil Elbaz: it seems like a big shift from the academia where you’re very, very focused on algorithms to a much, much broader spectrum, more closely related to let’s say a startup or some kind of org focusing on a product How was the kind of, how was the move for you? was it what you expected or was it, how was it different?
Lihi Zelnik-Manor: In every possible way. It’s a different life. So I think before you asked me some simple questions and one of them was, how is academia different from industry? You know, What is similar, come on. It’s different world. It’s like asking me why the academic career is one wayIndustry career is another way. Military career is another way. There are so many options for career. You can choose whatever you want. I’m lucky enough to have a chance to experience both actually today answering more seriously people in academia and field like mine, computer vision, or machine learning, have a chance to enjoy both worlds.
it is great that it is possible. And, and I really hope universities in Israel will enable this more and more as we see it also in other places in the world. it’s actually been quite standard in some fields. So for example, architecture economy law, people have always been with two feet, like one foot at the academia and one foot in the real world.
Not real world, but industry. it was because they made is just as real. And now it’s happening also for in engineering and you can leverage both worlds, so really want it for the long-term or for the research and to explore what are your thoughts and the other one to take it into practice and actually create products at work and people can hold in their hand and use.
Gil Elbaz: So speaking of people holding the product in their hand, how many people are using this product, is it, I’m assuming that there’s a very, very wide reach if it’s the Alibaba version of Google cloud for our audience.
Lihi Zelnik-Manor: You’re right. So I can’t give you specific numbers. These numbers are confidential, but you can imagine, you know, China is big 1.5 billion people, but we’ve just launched. So the first version came out in March and we are really in the process of exploring what is the right product. And maybe I’ll leave a little bit of background about specifically Drive products. So in the west, they’re very common, just looking at Google Drive. They have 2 billion users. And in China, there are competing products, but the market is really not taking it. the biggest one in the market has 70 million users. Oh, wow. So just a fraction is using drive product, which tells you that not hinders potential. On the other hand, it tells you nobody figured out what is the right drive product for the Chinese user. And the Chinese users are different from the American user or the European user.
They have different preferences. For example, if you just open any app in China, it is the appearance you will see immediately, very cluttered, very colorful. It’s really the opposite of the Google search, which I remembered the date launched. It was just a white page with the search bar saying I’m feeling lucky and that’s it.
And in China, it’s really the opposite. It’s red and orange and blue and green. And like a hundred mini apps inside your app and you look at, and as a Westerner and you look at it and say, are they crazy? I know they’re, they’re just have different preferences and this is what they like and enjoy.
Gil Elbaz: It’s so interesting to see how, the cultural differences can be visualized in this way.
Lihi Zelnik-Manor: There’s also a very small anecdote, which I like, I think like red and green have reversed meanings in the East and West. If you look at the stock market, what marks, the stock went up and went down. it’s reversed Yeah. So even the colors are different, so it’s, uh, it’s the small things and the get people hooked and want to use your products.
Gil Elbaz: Can you tell us a little bit about the teams themselves? a little bit about the structure communication between the teams, how you go about scaling that.
Lihi Zelnik-Manor: So I think there are two aspects that could be interesting.
So we building. Alibaba, Israel is very similar to building a startup within the a huge organization. So our ocean is not exactly the blue ocean that you can choose whatever you want. You need to choose your sea. Your sea is what makes sense for Alibaba and how you can leverage Alibaba. But apart from that, we built something very new.
So we are the first international R&D center of Alibaba. First foreigner side, we would say, so how do we work? Uh, with other teams was unclear, how do we do things, operational things was unclear. So we are very much like a startup, trying to choose our direction and building a team for that.
And our team today, as I said before, it spans from product, we have a product managers, we have mobile developers, a mobile team. We have a data team. We have engineering teams that does the back and front-end and, MLOps and an algorithm team that are responsible for the features that go into the product. And we have a research team that is more long-term and I’m bringing them making sure we’re doing state of the art technologies.
And they have a chance to contribute into the product at a slower pace, but keep making sure that we are activated and everything we do is really, really excellent and the best. So this is kind of for team structure, but what is interesting is to ask me, how do you scale up? And one of the things that we’ve done just recently is thinking about our development processes. And I think there’s a lot of know how now in the industry, people know that squads seem to work better for, deployment and deliveries. So we have done this recently.
Gil Elbaz: Can you mention what squads are?
Lihi Zelnik-Manor: So the traditional organizational structure, if you go back to the 1990s, teams were organized by profession.
people expertise, experts in graphics. There was a graphics team and people who were expert in front-end or the front-end team and mobile developers, developer and the team for apple, a lot of developers, et cetera. teams were organized by your expertise. And then if you want, for example, to create a product or a feature in a product that requires a mobile developer and a backend developer and an algorithm and a product manager, the multiple teams will have to kind of communicate and align their.plans. And then at some point, I think the first that made the buzz about it as Spotify people call it the Spotify model, they started talking about squads these are like teams that have people with multiple different expertise, diverse expertise. And they’re like an independent unit that can deliver the feature or the product.
So you will have a squad, could have one person from each of those teams. It depends on what, for example. So we have one squad that it is more about core technologies. And for example, does not have a mobile developer because there was no need. They just deployed to cloud. And maybe even a squad could have only algorithm experts, because this is all you need for that task.
So squads could be formed for a certain task and they could die. You could disassemble that when the task is complete, or if the task continues they could remain as a squad. And then it makes it easier to manage and tie all the loose ends together in order to deploy in a timely manner, because everybody is aligned, their OKRs, are aligned and they work together for the same goal.
Gil Elbaz: Yeah. I can tell you a little bit, from our perspective, we don’t officially call it squads, but we have a very similar method of taking on specific projects and getting multi-disciplinary people from a bunch of different teams to come together, we have a temporary kind of lead for this project. We call it a project lead And then he, or she manages the entire scope of that project, all of the people from all of the teams. And so we still have teams that are divided into the professions that they have, and they’re constantly learning, constantly challenging each other. So we do feel it is important to have that kind of guidance. And we have very senior people on each team. It could be 3d artists in our case or algorithm developers or backend developers. But in addition, there’s a lot of, these temporary multidisciplinary teams that are formed for a specific project. they’re usually like, it’s known that they’re temporary, that we usually don’t keep them around for too long.
Lihi Zelnik-Manor: So thank you for that. Bring up a very good point. So I call them squads. You call them project teams. The name is not really important. But what is important is that, or what is interesting is I know I talked to many different companies and you see that every company actually has different needs and different preferences and there’s no one way to do it.
So in my case, we chose similar to you. We have, the organic teams are based on profession because indeed they need to know how and learn from each other and, and, uh, learn technologies and make sure they’re up to date, et cetera, et cetera. So knowledge needs to be transformed and problems need to be solved together.
But then we also formed those quads that are, it could be temporary, it could be long-term and other, other companies might have just squads as their organization structure, like with the reporting lines and managers. the multidisciplinary teams will be the team. It might make more sense for them.
Some companies will have a mix. So it really depends on what you’re doing and what your needs are in terms in terms of development. And also who your people are. So people need to be ready for that. It needs to fit them. You need, someone wants to lead this project. Definitely someone needs to be willing to do that, and people need to want to work in that format. So it’s about maybe the mindset of the team.
Gil Elbaz: Definitely. We sometimes also see that nontrivial people are needed can actually accelerate projects substantially. we recently added operations person to a very technical project. But that required some, external resources outside of Datagen as well. We saw a substantial speed up in the entire project and she took on not only the external communication, but then also managed a lot of internal things that we saw as bottlenecks. And suddenly we got like a very nice boost in performance, something that was surprising and very positive recently.
Lihi Zelnik-Manor: That’s amazing. I think it’s beautiful. the main learning I have is that, there is no one solution fits all. I really love your story because personally, I strongly believe in diversity. And I think when you bringing someone from. Which is very different and you change things. And all of a sudden you have new ideas and things start to shift and say, oh, wow. I didn’t think of that. that’s just cool.
Gil Elbaz: Yes. She came with a mindset that was very different to our standard engineering focused mindset, which is very refreshing helped a lot.
Lihi Zelnik-Manor: Was she challenged at the beginning?
Gil Elbaz: Honestly, she came in, uh, running. She was like, ready to go. She wasn’t, I think, surprised at all or had a tough time.
Lihi Zelnik-Manor: That’s great. Often I see that those even that come with the different background or perspective, for example, let’s say you recruit the product manager that has never worked with AI. The questions that they would bring, the ways of work they would want to instill .The team might say, oh, that can’t work ,and dismiss what they say.
challenge this person who comes out with this different mindset. And then it takes time to see that actually there’s something to it. And maybe we can learn from someone that came from a different perspective. So often I found people they are used to something they want to keep that way and changing how they think and how they work.
Gil Elbaz: Yeah. Yeah. I completely agree. I think it takes time, but it also takes good communication. even with very different perspectives or new concepts ideas, a lot of times what makes or breaks the collaborations are the single individuals that communicate in a good way or not.
Lihi Zelnik-Manor: I think there’s a quote from Steve Jobs. I don’t remember the exact wording, but it is something like this, that it’s not about the technology. It’s about the people Communication is very challenging because uh, often people think it’s really has taken place and maybe it did not. So at the end of the day, I don’t think the main barriers for the success of our products and technologies are the technologies and the product are the people that the,
Gil Elbaz: I agree. And especially when scaling up, right? Like when moving from a small organization to a large organization, where, we’re at right now, and I’m sure that you are as well hiring a lot of people, we have a ton of open roles. This isn’t supposed to be an advertisement, but it is, but in practice, that’s one of the main things that we’re looking for, even in the most technical roles and also in the, obviously in the operations roles, having good communication is super important on our side.
Lihi Zelnik-Manor: What I find is often difficult for companies in scaling up is maintaining your organizational culture. Okay. So you know what the country you want to instill, and you’ve been doing that you recruited people according to that, but now you’re, maybe you were as the, as a startup founder, you’re no longer doing the recruiting of everybody.
Actually, you will find people that tell you that they interviewed each and every person that joins their team. Some people continue to do it, but most don’t and you start trusting your, your team leaders to do the interviews. And often companies suffer from pain points at that point of scaling up. Because those people that recruit may not fully be aligned with you on what to look for. And all of a sudden, you see a different culture that you are not hoping for. And people that don’t fit the culture you want to instill, and then you need to change. And, um, hard.
Gil Elbaz: That’s hard. Yeah. I agree. We’re not yet at that stage. So we’re around 50 people now. We’re not yet at the stage where we don’t have the capacity for either me or Ofir, my co-founder to interview, but eventually, I definitely see that it will happen. And then hopefully we can still instill some kind of culture, but it’s a lot about the top of the funnel and the people that actually come in.
Lihi Zelnik-Manor: First, it is possible. I really hope that you succeed in the, in doing that, as long as you’re thinking about it already, how do you scale up with your team and maintain the team that you want, then you’re already, you’ve done part of the way, right? Because you’re thinking about, so you’ll act upon it.
Gil Elbaz: A lot of our involvement today as, founders is around HR and around around people, sometimes also algorithms, which is the fun part. And just out of curiosity, maybe to take us back to the areas around methodologies of working, are there any unique methodologies that you’ve found. From either Chinese culture or Alibaba culture or internal Alibaba things, or are we already in like a very global situation where Spotify is, let’s say came up with something interesting and then it has kind of a global effect.
Lihi Zelnik-Manor: So first of all, also about Spotify, the account is something, and then you can find blog posts saying that Spotify doesn’t use the Spotify model. So I don’t know how much credit we should give them, but I do see changes and I see different teams were operating in different ways, also in Alibaba and propulsive, globally being part of Alibaba is I think maybe similar two other corporates in the sense that you have this framework to, for example, you talk about the team and how do you make sure people are working well and they’re happy and are motivated.
And how do you get the right team culture that they enjoy working there, right? You’re a founder, but you want to enjoy it, right. If it’s not fun, then. And working for a big company, then those companies, they come with their own culture. They come with the values. I think this was actually one of the things that for me personally, was very, very interesting and enjoyable.
So Alibaba has six values. Work, really follows those values, Twice a year, we have the performance process at Alibaba. You’re even evaluated on how well you align with the values. So it’s not just your outcomes. It’s also, the values are evaluated.
Gil Elbaz: Can you give us a sample of some of these?
Lihi Zelnik-Manor: Sure. Well, the values are actually posted on the main web page of Alibaba Group and you go to the first page, Alibaba talks about its culture. A strongly believe in the culture. Maybe even stronger than, than Western companies. It’s always there. And you always go back to the values when you’re in doubt. So they’re kind of like your compass.
So it starts from customer first, employee second, shareholder third, which is very, very different. So Alibaba is famous for when they went to NASDAQ in the IPO, the people that were there standing on the bench were not the founders. They were the first customers of Alibaba wow, which is interesting. Other companies. So customer first. And then trust makes things simple. So you’re talking about the people, communication, you need to trust people. When communication doesn’t work very well and you start having doubts and you’re not sure, but you trust people and it’s okay. He must be doing the things right. Let’s see what he’s doing. Let’s talk, let’s revisit it, et cetera, et cetera. I think one thing that is unique to Alibaba for example, is the third value, which is: change is the only constant. It’s always changing. It’s always rapid. Maybe this is something that is, kind of special to the Chinese culture or specifically for Alibaba.
The speed in which these these organizations start projects and shut them down is very high. So big corporations often move slowly, but not Alibaba. In Alibaba is OKRs, is to remain agile. So 250,000 people work for Alibaba and other wants to be agile organization.
Things change a lot, even for us Alibaba Israel we worked on the multiple different projects. And now we’re in a project working on this Alibaba Drive and we’re pivoting the product continuously every three months, every quarter or six months, the product changes direction because we’re in exploration mode and this exploration will remain for awhile.
So how do I recruit how do I keep people happy from point one, I look for those people that will feel okay with those changes that will enjoy them, and will see them as opportunity for growth. So if you find those people and then they join your team, then they’re happy with what’s happening. Then you have a win. This is how you close the cycle.
Gil Elbaz: And you have six values, but let’s say change is one of the big ones are having a dynamic environment that’s constantly changing. And then you try to look for that in the interview process as well.
Lihi Zelnik-Manor: Yes, another value is the, is about growth. So performance always is always improving. So I recruit people that really care a lot, about their personal growth, that they’re ambitious and they want to grow things. In a corporate.
That’s not always their distinct you would look for. Some corporate to look for people that are just happy doing their job, and they want more stability. And that’s not the type of people I’m looking for. I’m looking for people that want to grow and get a chance to do things they have not done before.
Gil Elbaz: Yeah. On our side, I can say that kind of in the startup environment you know, we don’t need to tell them that it’s dynamic. We do look at them. Yeah. We do look to see that they’re willing to adapt to the startup culture, but, but yeah, it’s, it’s a very, very dynamic environment. So it’s not something that we need to explicitly look for And I think it’s very interesting to see the contrast in Alibaba, taking a similar idea and expanding it to a giant company. In addition, on our side, we look for ownership. So taking things end to end, we try to understand if they’ve done this in the past. If they believe in it, something that’s very, very important communication of course is also super important on our side. And we look for this as well. And yeah, I think that there are a lot of smart people everywhere. Right. And so the differentiator on a team level is really the soft skills in many ways.
Lihi Zelnik-Manor: That’s an interesting point because I’m sure that you look for people that believe in your vision and mission, right? Definitely for sure. And the soft skills. This is a point that is very interesting. So people ask me young people ask me, should we go to university? Should we study? Why go to the Technion and a bachelor’s degree or a master’s degree? And I say, well, one thing is the knowledge that you accumulate, but I think it’s a soft skills. I see big difference in people and people soft skills that has correlation to the studies they went through.
Gil Elbaz: Definitely. I think that both the military service has a lot to do with these communication or soft skills and the same for university.
It’s like a life-changing experience and it’s a substantial amount of time and effort that goes into it. I think the only way to get through the Tiffany on saying is with friends and doing things to. You’d be surprised. You know, I thought that I thought that too, but they’re actually statistics. Some people do get by prefer to work on their own.
Really? Yeah. They exist. They’re not, that’s more than number, so different people go through it in different ways. But at least my perspective, let’s say you were really one of those people that like to work alone, you prefer to work alone and you can succeed doing your studies all by yourself. So let’s say you have the full thing.
You don’t actually don’t need anyone in your studies. You are forced to do some projects in pairs. You are forced to do some lab projects, or you’re not allowed to do it by yourself. And it forces you to go out of your comfort zone as someone who wants to do things on their own, it forces you to work with somebody else.
And then if someone comes to me and they say, well, I’ve done all those projects. These are my two projects that I did for graduation. And, uh, I was forced to work with someone, but I didn’t like it. So I did that or they were weak or whatever reason they will have and say, actually, I’ve done all the work by myself and I was okay.
Maybe that’s not someone that would work with me. Yeah. It’s yeah. It’s a hard sign for me for maybe for, for some rules, maybe it fits. But, um, I think some people need to learn as with communication need to learn, to work with each other. If you don’t know how to work with others, that’s a minus. So it’s awesome that you’re super smart and capable of doing end to end everything.
But if eventually at the end of the day, if you don’t do it with others, then that’s a minus and studying gives you a chance to do that, to work with others. Yeah, I agree. And maybe on the technical side of things, beyond the communication, I think that for us, at least one of the more interesting points or.
Points that we kind of focus on is to see that people are genuinely interested in what they’re doing. And so it’s not just about, you know, knowing, you know, the new architecture that just came out. Right. It’s, it’s just seeing that they’re, they’re really passionate about what they’ve done in the past, seeing that they’re passionate about the fields that they read blog posts and they, they do things that enrich their, their understanding because it’s interesting for them and not just because they need to.
That’s great. Yeah. So maybe to touch on, um, one of the big questions of today is kind of around how do we motivate our people? So as you know, as the job market’s been even more intense and as we’ve had so many people kind of moving between places at a, an increasing speed, I think the average time is less than two years now for an engineer to be in a single work environment.
How do you keep your team mode? And we can focus on the computer vision team because you have many different teams. Of course. Yeah. But actually I think it’s kind of the same for, for all. So I think you answered this question in what you’re doing. If you recruit the right people that like the vision and mission, they believe in it, they’re passionate about their profession.
Doesn’t matter if they’re a mobile engineer or an algorithm or a researcher. If they’re passionate about what they’re doing and they join a company that they laughter it’s mission and. And they fit for the culture. Then you have everything in place. Now, different people have different preferences and it’s all fine.
Some people really put high weight on the income. Others put really high weight on being able to continuously progress in their career management. Others, people just want to keep learning all the time. Other people appreciate, really want to work in maybe the big company, because they want to change every two years what they’re doing.
So it’d be company that have maybe some options. Some people really have the dynamic passion for startups and they like to chase the dream and others prefer stability. So there are people that switch jobs every two years, but the people that stay at the same job for 20 years, probably you, you like using the term a product market fit, right?
So it’s a product market fit. If your company is a good fit for. Forgive me for treating people as that, that engineer is a product. And if you’re a good fit for them, that higher, uh, more likely that they will stay longer time. Yeah. I can see the, kind of the concept of product market fit, going to engineer company fit.
We can call it or we just have a, we have a new, uh, ECF. Okay. So that’s going to be famous now, but yeah, I definitely agree that once, you know, once there is an ECF, I’m just getting a, once there is a, an engineer, uh, company fit. It makes things much, much, much simpler. And it’s interesting because like you said, everyone’s very, very different.
And of course there are good people that have very different preferences. And so. I think that one aspect comes to the filtering and the beginning of the process. Do you have like a process of identifying what’s actually really important for these engineers or kind of understanding? It’s a good, it’s a good tip.
Of course, of course. I think that in our interview process, uh, we give a lot of emphasis to HR. So HR interviews each and every candidate, and also we have an interview process and we aligned between us. What are the things we want to check? So if someone focuses more to professional aspects, someone else will focus on the HR aspects.
We cover all the bases. Definitely. And I’d want to just finish off with two last questions. What are the kind of next big trends that you see in computer vision? Maybe even beyond the timeframe of two, three years. So, I don’t know about the trend or milestones, for instance, I can say what I hope we will see.
So I think that these are interesting times and just yesterday I had a conversation with Tamara Friedman and, uh, he told me, no, I see all those companies that tell me that they’re done with the algorithm development. There there’s no need for further improvement. It’s just about a product now. And I said, wow, okay, interesting.
So many technologies are mature enough. They w so what’s that. Now there are all those things that don’t really work or not. They don’t work well enough. So what is the next thing beyond neural networks or today’s formal networks? I think this one, this is the interesting question. What is the next stages will take us one step further.
We’ve touched upon this at the beginning of our conversation and said, okay, call it AI, because everybody calls it the habit. It’s not AI. It’s not artificial intelligence in any way. We’re just doing deeply. We have not even solved machine learning. So we just teach very, very specific things. So what is the next step?
Can we improve on, on really mimicking some of the human perception capabilities? Well, that happened, and this is an interesting question. And, um, what’s your perspective on, uh, the question of consciousness building a system that we would consider at least conscious? I would say that this is very unclear to me, what that would be like, and are we ready for it?
So I think taking into something much, much simpler, for example, autonomous driving so many years just to get through the, the people factor, right? It’s not about the technology as my, not just the technology, people accepting it, the baking, the legal issues are . All that and accepting it and saying, okay, it’s okay to drive in such a car, let them drive.
And this is simple. It doesn’t touch upon what is in our heart. We really, we really are concerned about, right, that the rise of the machines and the AI taking over the world and these things that are scary for us, none of us driving is, is practical and technical useful and would get this rate of driving.
So conscious machines like any, like, like any technology could be totally awesome. Totally amazing. Change our life for the better, but could also be very bad. I agree. I think, I think it’s a double-edged sword. It’s, it’s an interesting concept, but between the lines, I’m going to assume that you can correct me of course, that you think.
Anything’s possible. Right? I think so, but I don’t see it in the near future, if this is what you’re asking. I don’t see in the near future, the city three years or 10 years, I think it will take much longer. Interesting. Okay.
Gil Elbaz: And last question that we ask all of our guests of course, is, uh, for new people in the field that are joining now either after their first degree, after their second degree, uh, maybe they’re just exiting some kind of military service.
What are your kind of recommendations or what are kind of your thoughts on how to start on the right foot and how to create kind of a, an initial career path?
Lihi Zelnik-Manor: So I’m still a little bit old fashioned because of the data I’ve seen until today. And it depends a little bit, I’ve been asked this question actually for many of ’em from, by many people, my, my kids’ friends and I tell them first, you need to define what is the career path you’re looking for?
But if you want to go into an area like computer vision, that you really still need to accumulate a lot of knowledge, I would recommend going through the traditional path, studying, getting a strong bachelor degree and then PhD. It really, if you want to be an expert, I would recommend that. So PhD gives you a chance to become a world expert in something.
It gives you a chance. Not everybody leverages this chance. Well, some people go and do a PhD, just kind of to have a checkmark, oh, I have a master’s of your avid PhD, but it is a chance that you could leverage. And if some people do leverage it and then you, you see them go out there and do amazing things like to start companies technology-based we get past technologies and they have this unique knowledge base that they take with them for life.
And from my perspective, being a, having a career for many years, I see myself using different aspects of the things that I’ve learned. So I used to be geometry that is statistics and probability and optimization and programming in different languages. And they changed, it used to be C and C plus, plus I coded in, oh, we don’t know what office to do that.
And now it’s a, and then MATLAB and Python. So this is just to give an example, it changes all the time. So how can you prepare yourself for a long career? And now we’re talking about long careers that you can work until you’re 80. So we’re talking about like 56 years of work. How do you prepare yourself for that?
Do you really think that doing a bunch of projects, even if they’re amazing in the military, give you the right infrastructure? I would recommend building a very strong infrastructure in. Computer science and physics build the infrastructure, have the tools, and then you can always learn and evolve and continue to, to remain relevant.
Very interesting. Yeah. I have less of a perspective of the PhD side of things, but from my experience in the academic studies, I can definitely agree that it gives you a very unique perspective. It gives you time to think time to really learn a lot. And I definitely think that the time there is time well spent, especially in the beginning of a career to build that infrastructure.
What I do today as a startup CTO is very, very much based on a lot of the knowledge that I gained through the academia. It’s good that you mentioned being CTO, different people have different goals in their career. So if you wanted to go for the tea, she wanted to go for the attack. If you want to be a tech expert for many years and remain.
For many years then yes, built this infrastructure and maintain it. And then you can, you can do that and enjoy it. It’s fun. That’s fun. It’s definitely fun. Thank you very, very much, Lee. It was a pleasure to have you here. Thank you. I enjoyed it.