Another industry AI leader has joined the academic circle, and it is becoming more and more common for industrial AI talents to transfer to academia. The special feature of Dr. Zhou Bowen, a fellow of the Institute of Electrical and Electronics Engineers (IEEE Fellow) and former senior vice president of JD.com, is that it emphasizes the integration and innovation of production, education and research, and pays attention to the industrialization of results from the source.
According to public reports, Zhou Bowen previously founded Xianyuan Technology. Check out the website of Xianyuan Technology, a technology company focusing on creating leading artificial intelligence technology platforms and products for the booming trillion-level industrial digital intelligence track.
Zhou Bowen recently told CIC (www.thepaper.cn) the original intention of joining Tsinghua University, discussed the current three limitations of AI talent influx in industry and conversational AI, and proposed artificial intelligence for young students.
For more than 20 years, he has been engaged in the research on the basic theory and cutting-edge technology of artificial intelligence. Zhou Bowen has made outstanding achievements in the fields of speech and natural language processing, speech machine translation, and deep semantic understanding. Many papers have been cited by international scholars for over 1,000 years.
In an exclusive interview with CIC, he said that in the past few decades, he has been doing the same thing, which is to expand existing knowledge, existing capabilities, and existing models at the forefront of technology and apply these innovations to solve practical problems. The accumulation and precipitation in the past also made him determined to further explore and consolidate the university scientific research team integrating industry and academia in the next stage, and work hard to truly make cutting-edge achievements of academic significance, economic benefit and social value.
“On the one hand, we should be encouraged and motivated by the progress of artificial intelligence in the past few years. On the other hand, it has indeed raised many potential questions. Why am I willing to return to colleges and universities to do cutting-edge research at this point in time? I think cutting-edge research There are still many problems that have not been solved. Especially in recent years, many ‘hard-bone’ problems encountered in industrial application scenarios need to be returned to colleges and universities through the establishment of a closer collaborative working mechanism of production, research, learning, trial, and application to Zhou Bowen said, “Tsinghua University is the most active and far-reaching university in China in terms of collaborative innovation. The scientific research team I established in Tsinghua is not only committed to bringing the core problems of the industry back to the laboratory to overcome, but also to The most essential thinking, experimental paradigms and cutting-edge achievements in the academic world can be better productized and planned. The biggest difference between this team and the existing model is that what we most hope to achieve is the circular promotion of basic research and innovative products. It is more closely and organically integrated with industrial R&D.”
Zhou Bowen’s existing academic achievements reflect the close integration of actual scenarios in the industry. In 2017, he served as vice president of JD.com Group, president of JD.com Artificial Intelligence Division, and founding president of JD.com Artificial Intelligence Research Institute. Business), Chairman of the Group Technical Committee, and Senior Vice President of the Group.
Before joining JD.com, he held positions such as Dean of IBM Research Artificial Intelligence Basic Research Institute, Chief Scientist of IBM Watson Group, IBM Distinguished Engineer, etc. He was responsible for IBM’s strategic planning and research in artificial intelligence and deep learning basic research. He is also deeply involved in the productization and commercialization of artificial intelligence technology.
As early as 2003, he developed the world’s first embedded large-vocabulary two-way speech translation system, and successfully promoted its subsequent productization and successful market application. He has led the team to develop the IBM Watson Platform, define and launch the JD NeuHub artificial intelligence open platform, and its core technologies are widely used in large-scale Internet applications such as cross-modal search, voice translation, intelligent customer service, shopping assistants, content generation, and digital virtual humans. , and industrial digital-intelligence applications such as intelligent supply chain, AI quality inspection, intelligent product design, and organizational digital-intelligence collaboration. Among them, in 2019, NeuHub was awarded the National Artificial Intelligence Open Platform for Intelligent Supply Chain by the Ministry of Science and Technology. He and his team also built a digital and intelligent exhibition platform to support the development of many major events including the China International Service Trade Fair in 2020 and 2021. Online and offline integration.
With the in-depth implementation of the artificial intelligence strategy, China currently has many artificial intelligence enterprises that can solve a single problem well, and also has world-class single artificial intelligence technology. “Strengthening the artificial intelligence industry chain is to put artificial intelligence technology into the application of all aspects of the real economy. In fact, it depends on the integration and innovation of AI and the real industry.” Zhou Bowen said, “Today, the large-scale application of AI in the real industry is only It’s just the beginning. Academia and industry are entering a ‘golden age’.”
The return of AI talent in industry
In China, AI talents in industry are “returning”, and it is more and more common for AI leaders of technology companies to transfer or return to academia.
In November last year, Fudan University officially announced that Qi Yuan, the former chief AI scientist of Ant Group, joined Fudan as a “Fudan-Haoqing” Distinguished Professor and Dean of the Fudan Artificial Intelligence Innovation and Industry Research Institute. Before joining Ali, Qi Yuan was a tenured associate professor in the Department of Computer Science and Statistics at Purdue University.
In August last year, Li Lei, the former director of ByteDance AI Lab, announced that he had joined the University of California, Santa Barbara. Earlier, Zhang Yaqin, the former president of Baidu, established the Tsinghua University Intelligent Industry Research Institute in 2020, and served as the dean of the research institute and a chair professor of intelligent science at Tsinghua University.
Ma Weiying, former vice president of ByteDance and director of the Artificial Intelligence Laboratory and IEEE Fellow, will join Tsinghua University Intelligent Industry Research Institute as chief scientist in 2020. In March 2019, the Hong Kong University of Science and Technology and the Innovation Workshop announced the establishment of a joint laboratory for computer perception and intelligent control. Zhang Tong, who left Tencent AI Lab, is the director of the joint laboratory. Zhang Tong is currently teaching in the Department of Mathematics of the Hong Kong University of Science and Technology.
In fact, in the past few decades, the interaction between talents in academia and industry has always been a very obvious trend, and there are many successful cases.
When Zhou Bowen worked overseas, there were many cases around him. Before returning to China, he worked at the IBM TJ Watson Research Center for 15 years. “I’ve always had many colleagues in the center who joined CMU, Yale, JHU, Columbia and other universities after doing research in industry for a long time.”
“Academic research is not only limited to papers, but there are also many researchers who hope to solve practical problems in industry. They also have great influence. Whether it is from academia to industry, or from industry to academia, It depends on each scholar’s current focus and research question.”
Zhou Bowen said that from a global perspective, this kind of industry-university-research integration is not something new in recent years, and it is often the results of successful industry-university-research integration that have profoundly changed our lives. Turning cutting-edge theories into epoch-making applications requires deep insights into technologies, products, and markets, especially for disciplines like artificial intelligence that pay attention to implementation. It is necessary to conduct more exploratory academic research in combination with practical problems.
In the process of industry-university-research integration, academic research and industrialization have different characteristics. The purpose of academic research is to explore new knowledge. On the basis of having a broad vision, it is necessary to focus on in-depth knowledge. It takes a lot of knowledge construction to understand the work of predecessors from the literature and review, and then researchers need to propose a very specific development. The research direction is a work from simple to complex.
Industrialization, on the other hand, is more inclined to solve problems, especially to solve problems with many practical constraints through the innovative application of technological achievements. A very important work here is the productization of technology, that is, it needs to be oriented to clear users and to be embodied. Do a good job in product design that meets the boundaries of technical achievements and capabilities, and do enough testing, testing and verification in engineering to determine whether it can be replicated and applied on a large scale at a reasonable cost. The final goal is the marketization and large-scale products of the product. deliver.
Any product that can be successfully scaled and commercialized has a very lean design, development, integration, and production process. All the complexity is hidden behind the product, leaving users with a reliable, simple, and easy-to-use experience, from complex to simple. “So we see that the implementation of most far-reaching industry-university-research achievements requires a process from simple to complex and from complex to simple.”
In actual work, the R&D in the industry is more inclined to the reverse market problem orientation, and the business problem to be solved is forced to think about which product to use, so as to solve the practical problem under the condition of controllable cost and guaranteed user experience, and then Look for technologies around your product that can both scale and solve problems efficiently.
More often, technical awareness of market issues is a process. R&D in the industry needs to “abstract problems layer by layer”, “the front is the insight of market demand and opportunities, and the latter is concretely represented as technology and product issues, and the required breakthroughs in technological boundaries are finally summarized into the basic issues of academic research. “
Zhou Bowen believes that the new integration of production, education and research is no longer just the common single-chain transformation of scientific research achievements from schools, companies to the market, but also needs to start from the market. At the same time, it can improve the market efficiency, improve the competitiveness of enterprises, and improve the standard of living. A good ecology in which the world and academia draw each other, and the double helix rises synchronously.
make real questions
In 2007, the iPhone was born. Earlier than this, in 2003, a large-vocabulary two-way speech translation system on handheld devices had appeared, and two people in different languages could speak to each other in real time through a handheld device. The developer behind this technology was Zhou Bowen, who was still working at IBM at the time.
“The best commercial handheld PDA at that time, I remember only 32 megabytes of storage, 206 megahertz fixed-point CPU, and no floating-point computing power. Its computing power and storage are now as primitive as dinosaurs. And speech recognition is also good , machine translation, speech synthesis, all have high requirements on computing power and model size, and in order to meet the needs of users for free communication, the larger the vocabulary, the larger the required model, and the more complicated the calculation. So I was a When people started doing this project, everyone felt that the challenge was very big, almost impossible.”
Before the emergence of large-scale two-way speech translation systems for handheld devices, the previous system developed by Zhou Bowen and the work of other colleagues have proved that larger computing devices can achieve real-time translation.
But a practical problem is that users of the speech translation system cannot always carry a bulky computer with them, and it is impossible for the system to be connected to the server at any time in the era of communication conditions and cloud computing at the beginning of this century when cloud computing was far from popular. “So it is clear that only by putting this technology on handheld devices can it really bring value to users and solve the problem of information exchange in the case of language barriers.”
From the perspective of academic research, the technology of speech translation system has been developed. “The previous project has been successfully completed, I have also made my contribution, and the written paper has also been published. I could choose to do the next one. Hot academic research.” If precious scientific research time is spent on the research of a handheld speech translation system, it is certain that a lot of optimization and engineering work must be done. The workload is huge, but it is difficult to base on these work. Publish CIC, and what is uncertain is the viability of the idea, because no one has done it before. So, a lot of people would find this a less glamorous project.
But Zhou Bowen was firm at the time, “This thing is something that can really bring value.”
“Thank you very much. At that time, my leaders at IBM also supported me to try. So I spent almost a year. Basically all weekends and nights I wrote code and debugged hardware equipment in the laboratory.” Android also does not have an iOS system. At that time, almost all handheld devices used Windows CE as a development environment and operating system for artificial intelligence systems. It was very difficult.
Zhou Bowen had to design his speech translation system from the bottom operating system, adapt the driver according to the underlying hardware, compile the embedded Linux system from scratch, write his own development tool chain, and redesign the architecture and redesign of the speech translation system. Write code, and more importantly, find new algorithms. “It turns out that the general architecture that everyone does is waterfall, that is to do speech recognition first, then machine translation, and then speech synthesis. But such a ‘daochang’ architecture is placed in a ‘snail shell’ like a handheld device, the first The speed is too slow, the second memory is not enough, and the accuracy of the third speech recognition will affect the accuracy of machine translation.”
To this end, he proposed a new model and a new architecture to reconstruct the speech recognition system. One of the innovations is to combine speech recognition and machine translation for joint search. A number of innovations integrate speech translation in the case of constant accuracy. The speed has been increased by more than 100 times. “So after we launched this system at the end of 2003, it really changed the way of thinking of many people. Later, in the research and achievement competition in this field, many companies and universities such as CMU began to move to handheld devices.” Based on its launch The portable speech translation product of the company has also been successfully commercialized in several application scenarios, and part of this work was later published in a review article on the progress in the field of speech translation in the Proceeding of IEEE journal.
Looking back, if it was purely for the purpose of maximizing the influence of CIC, Zhou Bowen’s doing this was not a “smart” move, and the energy spent on productization can actually be used to write more new papers. . But his choice is to insist that technological innovation should serve practical problems, and choose to let solving practical problems guide the research direction. This is the value of soaking in the laboratory back then. Zhou Bowen said, “To solve the problems that users actually need.”
Three Limitations of Conversational AI
“I have been doing the same thing for decades, which is how to expand our existing knowledge, existing capabilities, existing models, systems, and algorithms at the forefront of technology. I am very fortunate that academically there are Some of their own unique achievements.”
Zhou Bowen’s research interests involve multimodality (language, speech and others) and knowledge representation, understanding, interaction and reasoning, and new ways of trustworthy artificial intelligence. He and his collaborators are the first researchers to propose the self-attention plus multi-hop mechanism. This new mechanism allows the deep neural network to learn words, words and sentences by using the internal structure of the language and through multiple self-attention. The dependencies between sentences have greatly improved the computer’s ability to understand and express natural language, and related work has been cited more than 1,800 times in papers including Transformer. His two new model structures for natural language generation have been cited more than 1,700 times and nearly 1,000 times respectively in the field of AI generation.
Whether it is in IBM Watson, leading intelligent customer service in JD.com, or presiding over major special scientific research projects in 2030, he is researching conversational AI. Zhou Bowen believes that conversational AI is not only an application, but also a means of AI learning.
Just like conversations between humans, a good AI conversation system not only helps users, but also enables AI to learn faster. “The dialogue process itself is a very good means of learning. Don’t just study conversational AI as an application, but also study it as a learning mechanism. This is where conversational AI needs to be expanded next.”
Prompt AI, a new branch of natural language processing (NLP), is prompt AI, which is an input form or template designed by researchers for downstream tasks, which can help pre-trained language models “recall” their own pre-training Something “learned” while training. The prompt prompts the pretrained model, and as soon as the pretrained model “sees” it, it knows what task it is going to accomplish.
“Prompt AI is a very simple method of using conversational AI as a learning method, which has begun to attract a lot of people’s attention. But in the long run, how to make dialogue a learning method for AI is a very important research. subject.”
The most impressive thing in the field of language understanding and human-machine dialogue is the emergence of very large-scale pre-trained language models. Conversational AI is making great progress. An important reason is that the large training model allows AI to “see” most of the context in pre-training, so it can better predict what to say next based on the context.
But there are still three major limitations of conversational AI. Zhou Bowen believes that, first, the underlying mechanism has not changed, and the AI system lacks common sense, a cognitive model of the world and an understanding of the deductive mechanism. “Conversations between people, the explicit content is only part of the whole dialogue, and the other part implicit in the dialogue is about your and my common understanding of the world.”
“For example, in the dialogue, we all know the concepts of ‘Tsinghua University’, ‘artificial intelligence’, and ‘IBM’ and the meaning behind them… But in the AI dialogue, we do not have a good model to integrate these concepts that are not presented in the dialogue content. The role of common sense in understanding and predicting conversations. That’s a huge question.”
Zhou Bowen proposed to build a scene-driven knowledge representation mechanism. In each round of dialogue, the content of the round of dialogue is the core, and the corresponding extension of the dialogue is constructed in real time. “Entities involved or not involved in the dialogue, construct their relationship and logic, and use the structured and differentiable derivation representation of knowledge as part of the dialogue model.” This is also one of his current work directions.
Secondly, conversational AI lacks real-time reasoning and induction capabilities, because the current large model training only makes model extraction based on the context in which a large amount of data appears, and lacks interpretable logical relationship reasoning.
A third limitation is that conversational AI lacks the ability to navigate conversations. The dialogue between people is dynamic. During the dialogue, the effectiveness of the dialogue, the feelings of the participants in the dialogue, and the distance between the goal and the goal will be evaluated, so as to adjust the dialogue. Conversational AI “either is very easy to converge and does not expand the extension of the dialogue; or the dialogue is pulled by the other party”. That is to say, it usually shows that it does not have the ability to open dialogue, the scope of the dialogue is narrow, and it is impossible to actively and efficiently guide the topic. If in complex task-oriented dialogue, dialogue AI lacks self-learning of dialogue strategies and more game theory research.
Although the past few years have made great progress, the shortcomings of current conversational AI are also obvious, and this also contains huge academic research opportunities. “I’m personally very excited about this field.”
A good AI must be grounded
Talking about turning to academia, Zhou Bowen said that another important driving force and future focus of his work is to cultivate compound artificial intelligence talents.
In recent years, domestic colleges and universities have set off a wave of setting up artificial intelligence disciplines. Since March 2019, 35 colleges and universities across the country have obtained the first batch of new AI professional construction qualifications, and artificial intelligence has become a “popular” for three consecutive years.
The Ministry of Education recently announced the 2021 undergraduate majors filing and approval results in ordinary colleges and universities, showing that 95 colleges and universities have newly registered artificial intelligence undergraduate majors, including Peking University, Tongji University, Zhongnan University of Economics and Law, Southwest University, etc.
“I personally believe that with the increasing number of high-level young people in our country, whether it is learning ability, academic ability or research ability will be strong, the shortage of artificial intelligence talents will be well alleviated, I also believe that we will definitely There will be a lot of original scientific research results.”
Artificial intelligence is a discipline that solves practical problems. It requires not only cutting-edge technology research, but also practical industrial problems to create value. After DeepMind used the artificial intelligence program AlphaFold2 to predict protein structure, it also used artificial intelligence to control nuclear fusion, and nuclear fusion energy is also the frontier direction of energy development in the world.
Make research solidly solve problems in real industry scenarios. “As advocated at the 70th anniversary meeting of the Department of Electronic Engineering of Tsinghua University in April this year, ‘let the research results be put on the academic bookshelf and the industry on the shelf’.” Zhou Bowen said that the proposal of these concepts is very exciting, “it really generates creativity. That’s what AI for impact needs to be.”
Zhou Bowen said that artificial intelligence needs compound talents. At present, artificial intelligence education should pay special attention to cultivating talents with real problem orientation, industrial perspective, and good combination of artificial intelligence and industry.
Such talents need to have the ability to abstract and abstract the actual scene into cutting-edge issues with academic taste. In the process of solving problems, the new cognition of the academic frontier can be expanded, and the transformation of the industry after the expansion can be achieved naturally.
“We need to encourage more two-way linkage between academia and industry, encourage more teachers and students in academia to help the industry implement better technical methods, and encourage more colleagues in the industry to bring more real problems to the school. .”
For the cultivation of compound artificial intelligence talents, Zhou Bowen suggested that students should have academic pursuits and pay attention to practical applications at the same time. “Good AI must be down-to-earth, so I hope that students will be down-to-earth as soon as possible, not purely thesis-oriented.”
Secondly, for students to learn artificial intelligence well, they must lay a solid foundation, and master probability and statistics, stochastic process, linear algebra, calculus, graph theory, programming ability, computer architecture and other engineering capabilities from the undergraduate study stage.
“I suggest that outstanding undergraduates can enter the laboratory after laying a solid foundation in mathematics, physics and majors, learn the methodology of scientific research and try to solve some problems.” Zhou Bowen proposed to carry out research and applications under the guidance of tutors, and cultivate interest and curiosity. , especially learning how to ask good questions.
In addition, as AI research deepens, the barriers between AI subfields such as machine learning, data mining, natural language processing, and computer vision will be lower. “20 years ago, those who were doing artificial intelligence basically didn’t understand what natural language processing was doing, and those who were doing natural language processing didn’t understand what speech recognition was doing.” Therefore, at the postgraduate level, it is necessary to integrate as soon as possible to form an understanding of different disciplines. Insights, cross-integrated innovation.
Charity In China report
Paypal account: firstname.lastname@example.org
Follow facebook page：Charity In China