Updated: Jan 28
Robin.ly Exclusive Interview at NeurIPS 2019
Liwei Wang is a professor in the Department of Computer Science and Technology, School of EECS at Peking University. In this episode, Professor Wang highlighted one of his three research papers accepted at NeurIPS 2019, shared his perspectives on data privacy, the differences in the trends and challenges of AI research between China and the US, as well as the responsibilities of NeurIPS area chairs.
Prof. Wang’s main research interest is machine learning theory and has published more than 100 papers on top conferences including NeurIPS. He was the first Asian researcher who was named among “AI’s 10 to Watch”. He served as the Area Chair of NeurIPS and the Associate Editor of PAMI.
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 NeurIPS interviews and inspiring AI talks:
NeurIPS 2019 Paper Highlights
Margaret Laffan: We're here at NeurIPS 2019 with Liwei Wang from the Peking University. Welcome, Liwei.
Margaret Laffan: Three of the papers that you advised got accepted by NeurIPS this year. Is there any paper that you would like to highlight?
We have three papers and two of them are spotlight papers. The spotlight papers are selected as the top 10% of all accepted papers. But I'd like to mention the other one, its a poster. In that paper, we fully solved a 10-year open problem in the online learning area, and got very interesting results.
Margaret Laffan: What's the contribution of this paper to research?
In this paper, we provided a new algorithm that can learn if the data comes online. It can solve many practical application problems more efficiently than before.
Machine Learning on Differential Privacy
Margaret Laffan: As an expert in machine learning and pattern recognition, you've contributed to a book on machine learning and differential privacy. Data privacy, of course, is a hot topic for all of us. What are the limitations of the current technology?
I've been working on data privacy for more than 10 years. About 15 years ago, they were already a lot of work on data privacy, but there's a critical problem among all these work. There's no clear definition about what is “data privacy”, and what do you mean by “privacy is preserved”? There were many definitions, but none of them were about heuristic algorithms. For example, one day someone invented an algorithm to protect privacy, then the other day, another smart person invented an attack algorithm that can attack the privacy-preserving algorithm. This happens because there's no clear definition of privacy.
But in 2006, there's a breakthrough with a new definition called “differential privacy”. This is a very rigorous definition of privacy. If you design an algorithm and prove that it protects differential privacy, then whatever the computational power the machine has, no matter what prior knowledge it has, you can't break the privacy. The privacy is preserved and we can assure that. After 2006, most of the work in the data privacy area was built on differential privacy. So this is a breakthrough, a big step ahead.
But of course, there are also limitations in differential privacy. It’s very rigorous, but we need to inject a lot of noise to the data to protect privacy, and that noise hurts the utility of the data itself. This is a limitation in the current technology.
Margaret Laffan: How can you see us better preparing to solve this problem?
This is a very good question. I think one way is to find out if there are other good definitions of privacy. Differential privacy is one possible definition, but there may be other good definitions as well. So on one hand, we can protect privacy, on the other hand, we still have very good utility from the data that we want to use for machine learning or other statistical purposes.
China vs. U.S.
Margaret Laffan: You're an influential researcher for more than two decades and have attended many conferences globally. What would you say about the difference between the US and China regarding research direction, business application and the product commercialization, these three areas?
Let's compare China, the US and Canada. In China, just as I said, if we have a technology, we can make it better, apply it to business and make money very quickly. I think this is what China does better than the US, Canada and Europe.
But for foundational research in AI, I think the US, Canada and Europe are doing a better job than China. In China, currently there are only a few people working on foundational problems in AI. If China wants to become very successful AI, we need to do foundational research as well as applications, engineering and technology.
Margaret Laffan: When we think about the business application, we think about product commercialization. I know you've touched somewhat already about what you see in terms of the speed of adoption. How does it compare in China to the US?
In China, the commercial adoption comes very quickly.
Margaret Laffan: Do you see that people have any concerns around privacy, for example? How does that compare between the US and China?
In China, of course, there are privacy concerns, for example, for medical data. There are a lot of startups in China that are applying AI in medical analysis. There are privacy concerns, but the government of China also makes policies very quickly. For example, how to protect the privacy and encourage companies to do business in this area. There’re maybe dozens of startups working on the medical analysis using AI technologies. So I believe in this area, China goes very fast.
Trends in AI & NeurIPS
Margaret Laffan: What are some of the trends and challenges of machine learning in China that you've observed in the past 10 years?
There’re a lot of changes. 10 years ago, NeurIPS (formerly NIPS) was in Vancouver too. I was here at NIPS 2009. At that time, I clearly remembered that only 1% of the candidates were from China, but now, there’re a lot of Chinese people here at NeurIPS. And 10 years ago, there were maybe only four or five papers from China, but now this number is much larger.
In the past 10 years, AI and machine learning really developed very fast in China. Many students, young faculties, also researchers in the industry were doing AI related works. It was growing very fast. Also, I think machine learning and AI has found its applications very quickly - a new technology can turn in to a product even within months.
Margaret Laffan: So the AI adoption is faster.
Yes, I think it's even faster in China than in the US.
Margaret Laffan: Where do you expect the next major trends in AI and deep learning in China and globally as well?
One trend I can see clearly is to apply deep learning and general machine learning, AI to various aspects of industry. Currently, deep learning is mostly applied to computer vision and speech, natural language, but in many other disciplines, AI has power that is not being applied yet. For example, communication. I have talked to some communication companies, and I think there are opportunities to use AI technology to improve their current efficiency.
Margaret Laffan: Your are a mentor of NeurIPS as well as an area chair. Can you share some of what this responsibility entails?
Let me first talk about the area chair. In the past a few years, the submissions to the conference grew exponentially. Every year we have 50% more submissions than the previous year. Reviewing all these papers is really a heavy burden. The area chairs need to handle the quality of the review process to make sure good papers are accepted. That’s a lot of work. We need to read every paper and see if the reviewers make the right recommendations and give responsible comments. I think this is the responsibility of every chair.
Margaret Laffan: What type of criteria do you set then when you're assessing papers for submission?
I think contribution is the first - if there’s novel idea and good contribution to this field. The presentation and writing is also important for the audience to understand.
Margaret Laffan: Professor Wang, the final question for you: What are you most interested to learn at NeurIPS this year?
I’ve seen many talented young students and young faculties. I talked to them and learned a lot from them. This is amazing.
Margaret Laffan: Awesome. Thank you so much for your time.