Updated: Nov 12, 2019
Mei Han is the director of the US research lab for Ping An Technology, the core technology arm of Ping An Group focusing on finance, healthcare, automotive services, real estate and smart cities. She received her PhDs at Tsinghua University and Carnegie Mellon University and worked at NEC Labs America and Google prior to joining Ping An. At Ping An, she focuses primarily on machine learning and researching its financial and medical applications.
Mei Han served as the Corporate Relations Chair at CVPR this year. This episode is a live recording of our interview with Mei Han at the CVPR 2019 conference. During the interview, she shared her thoughts on the growth of CVPR and computer vision over the years, her career path and her vision for the AI research in Ping An Technology.
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 inspiring talks:
Wenli: We have Mei Han here with us. She's the director of Ping An Technology here in the US, and she is also the Corporate Relations Chair of this year’s CVPR. Thank you so much for joining us here today.
Wenli: So I heard that you are the Corporate Relations Chair of this year’s CVPR. So what is the responsibility for this role?
Actually, I did this two years ago. That was the starting point to get more sponsorships for CVPR [and] to get more corporations involved in this conference. So that year we got a record-breaking 100 sponsors, but this year the number is much larger. My responsibility, as a research scientist from industry, is to contact different companies. My role is to build a bridge between the corporate sponsorship with the conference and academic research.
Wenli: The CVPR conference has been growing rapidly through the years. And there must be a huge difference between two years ago and today. So, what do you think of this collaboration, the bridge you built between the academic world and the industry?
I think the essential reason for these two parts, the cooperation part and the academic part, to get closer is not because of the sponsorship of the conference. It is because of the reality of the research. This big data-driven world takes more effort from both sides to work together. One side has practical problems, true application scenarios, and big data from industries like Google and Ping An. And the other side is more focusing on [long term] research, like professors and students. So naturally, they should work together on the problems and solutions to make true impact.
Wenli: What do you think about the growth in both areas? Is there a gap? Which one is doing better, the academic or industrial world?
There’s always a gap. When I started my PhD in 1995 at Carnegie Mellon, [theoretical and long term] research was mainly done from the university side, and the companies were doing more business-driven research. I was an intern at Microsoft Research in 1997, then I joined NEC Labs America in 2001. So at that time, we basically focused more on corporate related research or business-driven tasks. But now the line is fuzzier, so everyone mingles together to try to solve a particular problem.
Wenli: On the autonomous driving side of AI, I know that industries definitely have more data and almost unlimited GPU use. So what do you think the academic people are going to do? Do they need help on this?
We do see more and more professors and students get involved in the corporate research. There are many professors taking sabbaticals, filling in positions in the industry. They send their students for internships. Many big companies sponsor the research on the university side. Google Cloud offers free membership [and computation power] for students to try their work on Google Cloud.
Wenli: We know that you have very solid research background, both in Tsinghua and CMU. What was the PhD research back then?
At that time, there was no deep learning. We did more research on traditional and typical computer vision problems. For example, by taking a sequence of images, we tried to estimate the 3D scene, the camera locations and internal parameters, which is called camera calibration. So that's a very traditional or typical computer vision problem, called structure from motion. This is [basically] a math problem which we solved by matrix factorization. We put all the tracking of feature points from the images together because these images are generated by the true scene, which is 3D, by camera projection. So we factorized this huge matrix into two [parts]. One part [includes] camera parameters, the other covers true 3D points. So it’s a math problem. That's the framework at that time.
Wenli: Amazing, so many things have changed. What do you think of almost two decades of the revolution in this field?
I think the big jump is the big data, the deep learning thing and Feifei’s ImageNet to make it possible with huge datasets and deep learning approach. So basically it’s a data-driven time.
Wenli: Those are the breakthroughs in these two decades?
Those are indeed big breakthroughs, but I still think most of the training for computer vision PhDs to understand features, understand geometry, understand the [physical] world [and image formation] at that time is still helpful. Even now, not everything is like jump-in and jump-out, or just big data. For medical and computer vision research, you still need to understand the data and the feature, the essential thing computer can’t grasp. So I think it is a big jump, but [the training] is still valid and connected.
Wenli: Back to the days in CMU, your advisor was Takeo Kanade. He's such a big name, a pioneer in computer vision and the director of the Robotics Institute. Anything that you learned from him still inspires you today and influences you?
Yeah, I think he's very famous for being a hard worker. Even at the age of 60 and 70, he still works overnight for a lot of project deadlines.
Wenli: How was it like working with him?
He always had late night or early morning meetings, like 2 am, 3 am. After overnight meetings, he still goes to golf or play tennis. He's really energetic. That's one thing that impressed us all. His energy beats all of his young students. Another impressive thing is that he's really paying attention to the details. He checked our code, confirmed every detail of the demo. He even modified the presentation slides for me including the font, the size, and the color. He taught us how to present and write diligently. He really cares about all the details, very hands-on. He is a great advisor.
Wenli: Do you still keep in touch?
Yeah, when he is in the Bay Area, all his students will get together; we’ll have dinner together. He always asks to try the spicy Chinese food, so he really loves to compete with us, he is really a fighter, even for who can try the spiciest food.
Wenli: I think that personality also has gotten him far in the career. We've talked about the revolution in [these] two decades. And we have gone so far in this area, so what would be the current challenges, like the bottleneck in the industry?
In terms of research, I feel like we are moving from the recognition perception phase to cognition, [which is] to really understand the world. Another thing, [especially] because I'm working at Ping An right now, there is an urgent need to do explainable AI because we are doing financial and medical applications. For these traditional areas, the business people [in the finance field] and doctors, really care about the true principles behind the big data. Deep learning cannot work as a “black box”. Lots of researchers have put their heart in this problem to solve the explainability of AI. [It is] a principle problem. There is math [and theory] behind AI, or explainability behind deep learning. We should try to solve this in a principle way to understand how big data and deep learning work for the businesses.
Wenli: I guess some businesses will see it as a very disruptive technology that might change the game forever. Are there certain parts that are threatening?
I think on the surface it’s the big data thing which makes the bar much lower. Everything can work to some extent with big data and deep learning. But somehow, to make it truly work, to make [it] robust, reliable, and to convince the business people to use this, we still need some principles to understand the whole process, to understand the math and theory behind the scene.
Wenli: What's your strategy on this? What’s Ping An’s strategy?
Ping An has been doing theoretical research for machine learning method[s] and principles. There are collaboration programs and intern programs to work with students and professors to do principal theoretical analysis of deep learning. And we just hired a postdoc to do expandable AI, basically for more research in the areas of finance and medical.
Wenli: Are there any challenges or obstacles in medical AI in the future?
The hardest part is to communicate with doctors to understand their requests, and what is most helpful for doctors and eventually for the whole community and for the patients.
Wenli: Can you explain to our community a bit about how those AI technologies will benefit us on the business side?
For Ping An’s businesses, we cover a lot of different areas like medical, finance, smart city. For smart city, we do some video content understanding and smart education to help kids to learn better. In terms of this, we try to use data to analyze the behavior of students, of teachers, of the class. So we built up the structure of the classroom video per student, per class, per teacher, per knowledge point. So we can basically structure the whole video stream to index the student behavior, to understand how they behave, and to understand how the knowledge point is presented.
Wenli: That’s an amazing use case.
In this case, we can help the students to learn better, to help the teachers to teach better.
Wenli: That also lowers the corporate cost in the future.
And to benefit the whole community. For some schools and for some areas without great teachers, we can broadcast their videos to the rural areas, so the kids or the teachers there can learn how to learn and teach the knowledge points better.
Wenli: That's amazing. You joined Ping An Technology Lab in 2018 to lead AI research including the areas that we talked about, finance, healthcare, smart city and video understanding. What is your current mission at Ping An?
Yeah, that's a great question. I have been thinking and re-evaluating our missions. From the beginning, we try to pursue two missions. One goal is to advance the technology in AI research. As one Ping An office located in Silicon Valley, we try to absorb the experiences of the talents from Silicon Valley to advance the technology. The other mission is to benefit the business - not only Ping An's business - to really understand the AI scenario. Because Ping An is a huge company with a lot of AI application scenarios. Through the platform of Ping An, we understand the scenario, understand the application and understand our data better. In terms of this, we try to push forward the whole research community.
Wenli: What are some of the achievements that you have done? I know you just joined for a year.
Yeah, a bit over a year. We are very excited here to do some impactful research. For computer vision, we have four projects. The major one is the agricultural monitoring tool. Basically, we use remote sensing and satellite images to monitor the crop and the yield to try to estimate the possible damage, to understand which crops can be planted here, the yield after the flood or any disaster, how the yield will change later in the year. In this way, we can help the farmers to plan better.
Wenli: I know we already have so much satellite data out there, but how do you label them?
Yeah, that's a very good question. We try to use a systematic and smart way to label the huge amount of data. For crops, we use NDVI images which are multi-spectrum images. Then, we try to cluster the spectra of the samples first. Ping An takes it seriously. We even go to the farms for every sample location. We do it smartly and try to sample the critical points, then go to the sample locations to do some measurements. That’s the true ground truth.
Wenli: What about the other research areas besides this one?
Like I mentioned, we also do smart classroom to study student and teacher behaviors. And in terms of smart education, we also do English speaking education. We try to estimate how to help Chinese kids improve their oral English, their speaking abilities. We try to talk with the kids, encourage them to talk for a few minutes. Then we estimate their vocabulary, speaking speed and if they are pronouncing the words correctly. We do put attention to these details.
Wenli: That’s really meaningful work that you’ve done.
Yeah, we are also working on English speaking conversational AI, like a chatbot in English to serve the English speaking financial business partners.
Wenli: There are a lot of plans going simultaneously, almost at the same time. So you have a team for each research area?
We have six to seven projects going on. Some are focusing on computer vision problems, NLP problems, and we do a little bit of FinTech research.
Wenli: This is really exciting. There're so many things going on. Each project will take a few years, right?
Yeah, we really have some long term goals, and also midterm projects with milestone deliverables. The good thing about working at Ping An is that we have some true tasks, we have some real requests from business partners within Ping An. These tasks are driving the whole process to move faster. For example, to reach each milestone, we try to deliver something to our smart city partners, to agriculture insurance partners. Then we know we are on the right track. Gradually, we will build up our long term goal of accomplishing the top of the world agriculture monitoring process.
Wenli: What are the milestones that you wanted to deliver for next year? What's your next step?
For next year, we first try to expand the lab to be a good size. Now we have 30 plus people, we will go with maybe 40 to 50 researchers. So together we try to build a lab that will be focusing on computer vision, NLP and ASR tasks and try to benefit Pang An’s businesses and beyond.
Wenli: Amazing. Personally, I’m curious [because] you have an amazing solid academic background. I’m sure when you graduated from CMU, there must be choices if you want to stay in academia or you want to go to industry. Why did you decide to go to industry?
One thing is, I wanted to try out what I have learned over the years. I got two PhDs, one from Tsinghua University, the other from CMU. I felt that I stayed on campus for too long, so I really wanted to try what I have learned, if I can apply my experiences to businesses. I think the industrial lab is a good combination of research and application. So I joined NEC Labs America, then Google, now Ping An Lab. All these three places brought me a lot of benefits to learn the business, to understand the application scenarios and to apply what I have learned. Now to lead the lab and share my experiences with young talents. That's really inspiring.
Wenli: Amazing. Thank you so much for coming here to join us.