Interview with Tencent AI Lab- Four Core Research Fields:Computer Vision, Speech Recognition, Natural Language Processing and Machine Learning

On March 23rd, 2017, the Tencent AI Lab had its first academic conference in Shenzhen, China. The conference officially announced that Tong Zhang, an expert in machine learning and former professor from the State University of New Jersey, would be the Director of AI Lab, and Dong Yu, former researcher from Microsoft and an expert in speech recognition would be its Vice Director. When the lab was first founded, Tencent was silent about it. However, it started to develop in speed.How does the Tencent AI Lab lead artificial intelligence (AI) research in China? As one of the earliest media reporting them, Synced recently conducted a special interview with the the Lab Director Tong Zhang. This article will combine the interview with our understanding of Tencent AI Lab from previous report, providing readers with a more complete overview about the Tencent AI Lab.

On March 22nd, 2017, the news “Andrew Ng Resigned from Baidu” blowed the entire IT industry. The turnover of talents in artificial intelligence has become striking in recent years, while these talents are the key for many technology firms to develop their artificial intelligence skills.

So far, many technology giants in the States or in China spend huge amount of fund to recruit experts in various fields related machine learning. Tencent is one of them as well. Tencent started to prepare for setting up this lab in 2015. In April 2016, Tencent AI Lab was finally founded. During this period of time, this low-key AI research organization absorbed more than 50 AI scientists and 200 AI application engineers. Most recently, Tencent AI Lab claimed Tong Zhang and Dong Yu’s join, implying the fast-paced development of the lab.

AI Lab’s Stratigic Position in Tencent

The success of Internet companies in China first depends on demographic dividend. The company can obtain certain market share by having a variety of products. Now the second core element contributes to this success is the technology.

Xing Yao, the Vice President of Tencent

“The success of Internet companies in China first depends on demographic dividend. The company can obtain certain market share by having a variety of products. Now the second core element contributes to this success is the technology. This is what Pony, a nickname Huateng Ma made for himself during his college time, said in one of his speeches,” said Xing Yao, the Vice President of Tencent, “This is also why we pay so much attention to AI.”

No doubt. AI technology is one of the core elements for technology giants to maintain their competitiveness, and allow them to explore more new opportunities. Nowadays, almost all the technology giants have their own AI research labs or related sub-organization. In the United States, Google founded Google Brain to further develop their AI technologies in 2011. In Google I/O 2016 conference, the company stated their mission of AI First. Around the same time, companies like Microsoft and Facebook kept recruiting AI talents and acquiring AI startups. E-commerce giant like Amazon.com also launched their Amazon AI Blog to introduce their AI research outcomes in the past February. In Mainland, China, Baidu has already took over fields like self-driving cars, and interactive voice response. As one of the three biggest Chinese Internet companies, Tencent was rather low-key compared to Baidu.

However, Tencent AI Lab is not the earliest product Tencent has for its AI technologies. Many years ago, Tencent had already set up various labs or product teams by combining its core services and fields like WeChat, QQ and Finance. For instance, at the end of 2015, Tencent worked with Hong Kong University of Science and Technology to set up their WHAT Lab, which specialized in image processing, pattern matching, machine learning and data mining. WHAT Lab is also one of the earliest AI teams Synced had kept in touch.

The mission of Tencent AI Lab is to enhance the creative AI research in Tencent. “In our Tencent community, we not only have AI Lab to develop our AI technologies, but also apply AI technologies in various products. Our AI Lab focuses more on technical research now. This is also why we invite global-level expert like Tong Zhang to join us. We emphasize on some technical research and research about creativity, hoping that we can use our research outcomes to support other Tencent’s products. Our products and the lab try to complement each other.”

“On the one hand, we do basic research. On the other hand, we develop systems and tools for Tencent’s product as technical support,” Tong Zhang added, “AI Lab support some major business products to accumulate and develop its technologies, and then popularize the mature ones within or outside of the company through cloud platforms, making these technologies available to a wider community.”

About the conversion from developing technologies to applying technologies to real products, Xing Yao mentioned that:”Our AI Lab is very similar to the combination of Facebook’s FAIR and AML two departments. Many people inside also develop applications. In terms of the organization of the AI Lab, our research and applications share a complete and closed circle that both elements complement one another’s growth. I believe that our process is better than other companies’ processes since we put both departments into one. This is the difference. ”
Even though Tencent started its AI research early and tried to apply the technologies to WeChat and QQ, Tencent is still behind in terms of the future direction of their AI technologies.

“In terms of technologies, we are more technical, which is similar to most of the Internet mainstream companies. Nonetheless, we still have our own features. We pay more attention to social AI, content AI, gaming AI, and AI tools. We have clear direction about these sub-fields since they are more related to our company’s featured products.”

However, these technology companies all share similar advantages in terms of developing AI technologies: huge user population, wide product usages, big data and talents. Xing Yao points out that these advantages were not exclusive to Tencent; all Chinese technology companies have it. In addition, he thinks that: “There are many more AI talents in the world”, and he believes that Chineses have a lot of potential in AI field.
This claim matched with Tong Zhang’s opinion in the first academic conference this morning. Why does the AI technologies in China can be at equal stage as those in the United States? Tong Zhang thinks that it is because the tech companies in Mainland educate plenty of AI talents; they have very strong experience in handling data and using machine learning technique to solve real-world problems.

Four Core Research Directions

Focus on fundamental research in four fields around the core services and products.

While Andrew Ng was still at Baidu, Baidu Brain is the center of Baidu’s AI research, covering speech, image and natural language processing. Based on Tencent’s core services and products, Tencent AI Lab will dig deeply into these four fields: computer vision, speech recognition, natural language processing and machine learning.

Tencent plans to develop its fundamental research in these four fields. At the beginning of the year, Xing Yao introduced the fundamental research Tencent AI Lab will conduct to support its core business services and products at the Tencent research annual meeting.

In terms of technical research, Xing Yao said, “Tong Zhang is leading a team focusing more on reinforcement learning research. Furthermore, Zhang directs many multi-layer models and generative models. These research are already at the top of the world.”

Reinforcement learning is one of the most active research fields in AI. It investigates how the agent takes actions in an environment to maximum the reward. The famous AlphaGo uses tons of reinforcement learning techniques.

Tencent AI Lab also develops their Go AI named Jueyi based on Tencent’s gaming AI services. Jueyi won the champion in the tenth UEC Go Computer Cup.

After the entire competition, Jueyi’s team director Yongsheng Liu told Synced that “What supports Jueyi from the behind is deep learning and reinforcement learning, the two hottest research fields. Its main structure is based on an AlphaGo journal paper published in Nature in January 2016. It is a pure machine learning system but it also produces some breakthrough beyond the paper while the program actually runs in real world.”

Tong Zhang and Xing Yao both agree that this kind of breakthrough shows the advantages Tencent gaming services have.

“Since we have some game context, as long as our technology gets better than they are before, we will be satisfied. We don’t need a perfect plan, so we can have many iterations. Similar to Jueyi, it is not an international-level player at the beginning, but it learns and grows slowly from all the games it plays. During such process, even just a little bit progress can be beneficial.”
Moreover, Xing Yao mentions that they have switched to other games now, and should have some research results coming up. Last year, DeepMind and Blizzard teamed up and worked on AI challenge like StarCraft II. Based on the gaming services Tencent owns, will AI Lab develop AI engines that work with complicated games like LOL? We are hoping to see Tencent AI Lab’s research result.

Tong Zhang, an sophisticated, professional researcher

“Theory basics are the most important.” — — Tong Zhang

Doctor. Tong Zhang from Tencent AI Lab

Dr. Zhang is an expert in the the Recruitment Program of Global Experts program, also known as The Thousand Talents Plan, under the Organization Department of the Communist Party of China. He has a Bachelor of Science in mathematics and computer science from Cornell University, a Master’s degree and a Doctor’s degree in computer science from Stanford University. Before he joined Tencent, Dr. Zhang was a professor in the State University of New Jersey, a researcher in IBM T.J Watson, an office researcher from Yahoo! New York lab, a Vice Director in Baidu Lab and the director of Big Data Lab. He participated and led the development of various machine learning algorithms and application systems.

In this interview with Synced, Dr. Zhang expressed his opinions about the research development in machine learning by combining his research and career experience. He thinks that at the shallow learning stage, it mainly is the theories guiding the practices; even though the current deep learning is lack of theoretical foundations, theories will push the practices forward again as related research in the field keeps developing and growing.

When Dr. Zhang first started related research, deep learning was not as popular as it is now. He mentioned that his research direction was data computation during his time at Stanford. After he graduated, he joined IBM T.J Watson to do research about machine learning and natural language. That was the time that data-driven machine learning techniques developed really fast in the industry. Later, he joined Yahoo! Lab in New York City, getting to know more about text and big data analysis in various scales.

After that, Dr. Zhang returned to university to do research; the research he did at that time focused more on fundamental research in machine learning, such as statistical machine learning. He explained that his industry experience made him realize the importance of integrating theory and practice. For example, he worked with Kai Yu from NEC Lab in 2010, building a series light-layer models based on statistical machine learning theory. Their image identification technique in this project was at the top level of the world, and won many important international competition, like the first ImageNet. He also claims that he is interested in deep-layer models, and is working on some deep learning applications and research.

Dr. Zhang is one of the Chineses who published the most journals in international conferences like NIPS and ICML. Speaking of the development or changes in recent NIPS conferences, Dr. Zhang comments that NIPS is one of the popular conferences in machine learning since its research direction is interdisciplinary and broad, allowing professionals from different fields to communicate, exchange innovative ideas and broaden research horizon. At its earliest stage, the conference was smaller but its coverage was much wider, including neuroscience, statistics, speech and image application research; it emphasized more on new statistical models, and how to combine theories with practices. The NIPS we have now is much larger, and has similar trend to other conferences, like the differences between NIPS and ICML are not as obvious as they were before. The machine learning research now focuses more on practice: researcher develop a model or algorithm super fast, adjust the parameters a little bit, and publish their paper after they obtain good result from collecting data. However, NIPS still tries to keep its own original features and will publish some paper that contain real innovative ideas annually.

Research cannot just focuses on practical applications; theoretical foundation is also very important. From Xing Yao’s perspective, “Theory basics are the most important.”

However, sometimes theory is something we would like avoid. Due to the complexity of deep neural network, some relevant research is basically working in the Blackbox, giving people impression that neural network is too complicated and the theory is too difficult to catch up with the practices. In this case, Dr. Zhang think that the gap between theory and practice should not be a reason that prevent practitioners to move forward. Deep learning develops fast in practice is a good thing, which makes theorists reflect what they can contribute to the more complex models we have now. He claims that there will be more and more theoretical research in deep learning, which might be helpful for research that focuses on practices, just like the situation happened at the early stage of shallow learning.

Xing Yao agrees with Dr.Zhang’s opinion. Yao states, “Most of the deep learning now works with heuristic and experienced; it lacks some theory support. If deep learning wants to go even further, people have to fill this theory gap. One of the missions at our AI Lab is to make deep learning go further. This requires us to be down-to-earth, work on fundamental research, and then develop some research on fundamental algorithm theories. Methods driven by applications, experience and heuristics are limited. This is also an important reason why Dr. Zhang joined the AI Lab here.”

Original Article from Synced China http://www.jiqizhixin.com/article/2550|Author: Pan wu | Localized by Synced Global Team: Jiaxin Su, Meghan Han, Rita Chen