Updated: Aug 26, 2019
Ji Liu is an assistant professor at the University of Rochester and the Director at Kuaishou Seattle AI Lab. Prior to Kuaishou, he was the Principal Scientist at Tencent. His research interests include Machine Learning, Optimization, Reinforcement Learning with emphasis on big data involved scenarios. Professor Liu is also interested in applying technology to solve real world problems in the area of recommendation, image analysis, game AI design, vision understanding, etc. The asynchronous parallel algorithm he proposed has been widely used in many machine learning platforms, such as Google’s TensorFlow. In 2018, he was also named “35 Innovators Under 35 in China” by MIT.
This episode is a live recording of our interview with Dr. Ji Liu at the 2019 CVPR conference. He shared his experience in deep learning research and how it is applied to optimize user experience at Kuaishou.
Highlight 1: AD-PSGD Algorithm for Deep Learning
Robin.ly is a content platform dedicated to helping engineers and researchers develop leadership, entrepreneurship, and AI insights to scale their impacts in the new tech era.
Subscribe to our newsletter to stay updated for more CVPR interviews and more inspiring talks:
Host: Good morning, Professor Liu, it's great to have you with us. Thank you for joining us.
Host: So you're a very experienced scientist in machine learning, optimization and reinforcement learning. Why did you choose the field of machine learning, especially deep learning optimization for research at the beginning?
Yeah, so because machine learning is a kind of methodology, studying how to solve a general data-driven problem, it can provide a generic tool to tons of application problems like computer vision, like NLP, like speech and many other things. I want to learn the general tool and methodology, so that's why I chose machine learning as my main research direction.
So back to your second question. Why [choose] deep learning optimization? As we know, deep learning is so popular and powerful nowadays, that makes deep learning optimization super important; Optimization efficiency is the key to accelerate the process from modeling to product. You may know, everyday people run the optimization or training processes for different deep learning models, the programs, for many, many times. If it can be accelerated, for example, 10 times, then everybody can save 90% of their time on running the stuff, so that can significantly accelerate the process to make the model, the algorithm to real product.
Host: We're going to talk about later around how we see that in industry then as well. So we'll be able to talk about AI commercialization, and then you can see how the research is brought into industry, so it would be a good area to explore further. As one of the MIT innovators under 35 in China, you created the asynchronous decentralized parallel stochastic gradient descent algorithm (AD-PSGD). What are the contributions of this algorithm?
Yeah, well asynchronization and decentralization are important and powerful tools to accelerate the training process of deep learning. These two new technologies significantly reduce the synchronization overhead. It breaks the traditional ways to do it. And these two technologies right now have been widely used in popular solvers like TensorFlow and PyTorch. And we believe this kind of things can bridge the gap between the traditional optimization people and the implementation people.
Actually, there is a big gap between these two groups of people. Optimization researchers understand the algorithms, but usually know little about the implementation, so when they design an algorithm, they consider very little from [an] implementation perspective. But the engineers taking care of implementation program the code in CPU or GPU may lack the deep understanding of the algorithm side, so this is the gap. And my work is trying to fill this gap and making the joint optimization, optimize both the algorithm and the implementation of hardware, to optimize performance, to optimize the efficiency of the training process. That is a main contribution and the purpose of my work.
Host: So you've been an assistant professor for five years. You're also working as a research scientist with corporate. Why did you choose to work in both academia and industry?
Very interesting question. In fact, many other people asked me the same question. I think that I am taking benefits from both areas. In academia, people usually study a deep point in a problem, which requires the skill of achieving the extreme at a single dimension. In comparison, people need to solve a complete problem in industrial, which requires the comprehensive skills. They are two very different experiences, and make me grow in different dimensions.
Another point worth to point out is that the problem considered in academia may be unrealistic. In particular, it might not be the real problem or important problem in practice. On the other hand, people in industrial may over focus on the specific problem, but ignore the fundamental rule and the general form of their problem. As a result, many people repeatedly do the same thing.
Therefore, the academia and the industrial represent two different groups of people. They can gap but can take benefits from each other. My goal is to bridge these two areas and make them take benefits from each other.
I think my experience in industry is also very helpful for my career. I know really what the real problem looks like. In academia, we sort of live in the ideal or virtual world and we only care about a few very general problems. But in industry, we care about the real problem, and that could be something very different from the problems we consider in academia. It doesn't mean the skills we learn from academia is useless. It's useful but depends on how to use it. And there is a gap, and my goal is to try to fill this gap.
Host: Yeah. Let’s talk a bit about Kuaishou. So you joined there at the end of 2018 as the director of the FeDA lab in Seattle. Can you tell us more about the mission and research focus of the FeDA lab?
Yes, it’s a good question. Commercialization is one big project recently in the company, and the research component in this project is to recommend the video advertisement to our users. The challenging part is in Kuaishou; actually we have two groups of users. One group of users share videos, and the other group of users just watch videos, but there’s a big overlap between the two groups of people. The principle of the company is to service these two groups of people simultaneously, but as we know, advertisement sometimes may hurt users’ experience. So that requires better technologies to do personalized recommendation. How to minimize the hurt to the users, to maximize their happiness; that is the challenging part. And from a technical point of view, how to formulate the happiness or satisfaction of users. That's the key and given that part, we need to design new algorithms, new technologies to do the personalized recommendation tasks.
And that's also very closely related to the computer vision staff, because in Kuaishou, all the advertisements are video-based. And in traditional recommendation problem, the context is mainly just text or image. Because of the new forms of the advertisement, it requires new technologies, new staff, and that's why I'm here to attend the computer vision conference.
Host: Right, to learn more and how the field is still progressing, you talked a lot about the user experience and user value, and we know that's very important in creating new products. So can you elaborate more from Kuaishou’s perspective on how you build this into your product development?
Yeah, that's a very interesting question. Yeah, Kuaishou does care about the user experience, and we put the user experience first. And the commercialization is second, to be honest. That is the principle of our CEO and the CTO and basically the whole company. We have the common sense: user experience is always first. So every time we do anything, we need to try to understand what is the gain from our users’ experience, and that requires a complicated data analytics task and also requires some new technologies, the artificial intelligence to understand that. Human is a very complicated system, right? Sometimes it’s very hard to use a pure or single theory to explain it. People say: Oh, psychology is a way to explain the human being, sometimes is true, but it’s not always right. We have to combine the expertise from different domains to understand our users. So that's why in Kuaishou, we recently built up a new team, stats analytics. One important goal for this team is try to understand human behaviors and the human satisfaction for us to build a better product.
Host: So you keep a very direct human-centered approach. You’re taking on this new role, how do you manage your team? What is your leadership style?
I don't want my leadership style to be static. I think my leadership style is dynamic. I think my leadership style needs to fit the needs of the team, depends on our members. In some stages, for example right now, the management style is more centralized because we are growing. We want to grow fast and it’s kind of centralized, but my ideal management scenario is decentralized. I'm like a coach and my members are players and they can communicate with each other. The cooperation is among them, not everybody needs to communicate with me to make their cooperation better.
Host: I’m hearing a lot of words around decentralization, collaboration, integration moving forward, that’s all great. What are your plans for yourself and Kuaishou in the next three to five years?
Ji Liu: I think Kuaishou has a very good user economy, and the product matrix right now is quite sparse, and we do have many things to do yet. For the next three to five years, my goal is to really understand the key problem in business. I am kind of a problem-driven person, a product-driven person, and I have lots of AI technology, but to utilize or to fertilize the technology in AI, my idea is to understand the product first. Understand our user, design the best product and then, we can design algorithms, we can design new technologies, we can define new things to serve that purpose. So that basically is my goal in the next three to five years.
Host: That’s wonderful. We look forward to tracking your progress and seeing how you guys grow, and I'm looking forward to celebrating with you as well. Professor Liu, thank you so much for joining us today.
Ji Liu: Thank you!