Zhu Songchun: 30 years of research on general AI (Science and Education Figures)

Biography of a scientist

Zhu Songchun: internationally renowned computer vision expert, statistical and applied mathematician, artificial intelligence expert, dean of Beijing Institute of General Artificial Intelligence, dean of Institute of Artificial Intelligence and School of Intelligence of Peking University, and chair professor of basic science at Tsinghua University. He has won the highest international award in the field of computer vision - Marr Prize for 3 times.

At 9 o'clock in the morning, Zhu Songchun set off from Langrun Garden of Peking University to Beijing Institute of General Artificial Intelligence (hereinafter referred to as "Tongyuan") across the road from Peking University to deal with various affairs, and returned to Peking University in the afternoon to teach students of Yuanpei College, and then went to the Department of Automation of Tsinghua University to guide students to carry out scientific research. In one day, his identity switched between "Dean of Tongyuan", "Professor of Peking University" and "Professor of Tsinghua University", and he was very busy.

"In the more than three years since I returned to China, I have been doing this almost every day, and the pace of work is very intense, but I feel very fulfilled." Zhu Songchun said this in a recent interview with this reporter.

"At that time, I secretly decided that I would study well, go to university, and become a scientist."

At the end of the 60s of the 20th century, Zhu Songchun was born in Ezhou, Hubei. Although his family was poor since childhood, he has always been diligent and hardworking, and he has excellent character and learning.

In 1977, the gaokao system was restored, which was a major change that affected the life path of many young people and even the direction of Chinese society. In this regard, Zhu Songchun, who had just entered the third grade of primary school at that time, still had a very limited understanding, but he was also touched. One day, he recalled, he saw the playground of Changling Middle School near his home full of candidates from all directions, quietly listening to the school's director of education about the mobilization speech for the college entrance examination. "They are full of spirit and fighting spirit, as if they were going to the battlefield! At that time, I secretly decided to study well, go to university, and become a scientist! Zhu Songchun said.

"Also going to university" not only inspired Zhu Songchun to overcome many difficulties, complete junior high school, and become the only student in the town in the same grade to enter high school, but also excellent academically. In 1986, he was admitted to the Department of Computer Science of the University of Science and Technology of China with excellent grades, and his college dream came true.

"Technological innovation is not waiting, I just want to make some original results"

During the university, Zhu Songchun was eager to learn, and gradually developed a strong interest in AI, not only took the initiative to audit the introduction to AI course in the Department of Automation, but also took the neural network course for graduate students. In his junior year, he wrote to Zhang Daren, a teacher in the cognitive science laboratory of the Department of Biology, to participate in the study of the mechanism of early visual attention, during which he read the book "Vision" by British scholar David Marr. It was this pioneering book in the field of computer vision that led him to 30 years of research and fruitful results. In December 2022, Professor Zhu Songchun's academic monograph "Computer Vision: Statistical Models Interpreting Marl's Paradigm" was officially published, providing a complete mathematical model for Marl's framework.

"Is it possible to build a 'grand unification' theory to explain various complex intelligent phenomena and achieve universal AI?" This is a question that Zhu Songchun asked in his third year of university, and finding the answer to this question has become his main goal in scientific research since then. At that time, domestic artificial intelligence research was just beginning, and reference materials were very scarce. In order to collect information, he and a classmate took a train to Beijing, copied hundreds of foreign papers at the National Library, and returned to school to study with his classmates.

When he graduated from university, Zhu Songchun was sent to study for a master's degree with the second place in the whole department, but at that time, he could not find a supervisor in the intersection of vision and cognition in China, so he had to do research while looking for opportunities. Finally, in 1992, Zhu Songchun was given the opportunity to study for a doctorate in computer science at Harvard University in the United States, under the supervision of David Munford, president of the International Association of Mathematicians and winner of the Fields Medal.

"Breaking the waves and riding the wind, subtle science is passed on. The Chinese nation is not alone in its aspirations. Before leaving, Zhu Songchun received a poem from a middle school teacher, and he deeply felt the ardent anticipation between the lines.

Zhu Songchun lived up to his expectations and advanced in his studies. In 1996, he made his name for himself in academia when he graduated with his Ph.D. At that time, the mainstream research of computer vision was rules and geometry, and statistical methods were not favorable. The foundation of contemporary artificial intelligence is statistical model, Zhu Songchun challenges the traditional paradigm with advanced thinking, and for the first time uses statistical methods to accurately depict "what is texture" with mathematical models, solving the basic problems that have plagued the field of computer vision for more than 20 years.

Compared with the rapidly developing Internet field at that time, artificial intelligence research appeared somewhat deserted. Zhu Songchun found that many researchers around him went to Internet companies, however, he chose to stay in universities to continue to engage in AI research, immersed himself in exploring the "grand unification" theory of artificial intelligence, and successively taught at Brown University, Stanford University, and the University of California.

"Teacher Zhu is a very pure scientist." Zhu Yixin, an assistant professor at Peking University's Institute of Artificial Intelligence who graduated from Zhu Songchun's lab, said, "I often see Zhu sitting and meditating, and the scratch paper in front of him is densely packed. He was quick to give us unique insights on the questions we asked him. ”

"The time for technological innovation is not waiting, I just want to make some original results." Zhu Songchun said that scientists enjoy the wonderful experience of exploring the frontier of science and technology. He was tireless and enjoyable, often working late in the lab, and often returning home in the dead of night and starry.

"For China to win in the AI competition, it must cultivate a large number of top talents with international vision and national feelings"

At the end of August 2020, Zhu Songchun returned to China to organize the General Institute and led the scientific research team to explore the forefront of AI. "If China wants to win the AI competition, it must take an original technical route and cultivate a large number of top AI talents with international vision and national feelings." To this end, Zhu Songchun has been intensively leading the creation of a new artificial intelligence curriculum system at Peking University and Tsinghua University, and actively promoted the establishment of Peking University Intelligence College.

Zhu Songchun predicts that there are three major trends in the development of artificial intelligence: first, in the core research fields of artificial intelligence, such as computer vision, natural language understanding, cognitive reasoning, robotics, machine learning, and multi-agent, which are increasingly highly integrated and unified, realizing the transformation from weak artificial intelligence to general artificial intelligence; Second, the intersection development of intelligence and other disciplines, such as psychology, neuroscience, biomedicine, etc., expands its extension; Third, the theory and methods of general artificial intelligence have gradually become the foundation of humanities and social sciences, realizing the development of "new liberal arts" and "new social sciences" from qualitative to quantitative, and exploring the intelligent society of human-machine symbiosis. In line with this, Zhu Songchun positioned the training goal of artificial intelligence discipline as "general knowledge" in the context of the intersection of "new engineering", "new liberal arts", "new social science" and "new medical science", and integrated the "general intelligence" of the six major fields of artificial intelligence and the "generalization" of all walks of life.

According to Li Wenxin, deputy dean of the Institute of Artificial Intelligence of Peking University, in order to improve the teaching effect of physics courses, Zhu Songchun carefully studied the physics course syllabus, delineated the knowledge points, held a special seminar with the teachers of the School of Physics, set up a special course research group, and finally designed two physics courses tailored for the students of the general artificial intelligence experimental class (hereinafter referred to as "the general class"), covering the basic knowledge of physics from the beginning of high school physics to modern physics, highlighting the modeling ideas of physics.

In this way, through nearly 3 months of hard work day and night, Zhu Songchun and more than 20 interdisciplinary experts have repeatedly polished, and finally established a new set of artificial intelligence course system, covering six core areas of artificial intelligence horizontally, including computer vision, natural language, cognitive reasoning, robotics, machine learning and multi-agent, vertically running through the course system from mathematical foundation, artificial intelligence theory to cutting-edge practice, etc., while opening AI and art, philosophy, sinology, law, economy, society, Multiple cross-cutting courses in government and public administration.

Under the promotion of Zhu Songchun, Peking University established a general class in Yuanpei College, and Tsinghua University established a general AI direction for teaching and training programs according to aptitude in the Department of Automation. Up to now, more than 150 students from 3 grades have joined. Tongyuan has become their scientific research practice platform, gaining a growth channel from Tongban to Tongyuan.

"The new training program led by Teacher Zhu changed the trajectory of my life." Liu Yu is a 2020 undergraduate student in the Department of Automation of Tsinghua University, and originally thought that as an undergraduate student he did not have the opportunity to participate in such a high level of scientific research, but he passed the interview in his freshman year and became the first batch of students in the direction of general artificial intelligence at Tsinghua University.

Special lectures, summer scientific research practice, scientific research training camp... In addition to systematically learning the interdisciplinary disciplines involved in general artificial intelligence, Liu Yu also participates in various forms of scientific research and humanities activities. Under the guidance of Professor Jia Baoxiong, the supervisor of the General Institute, he began his research in computer vision by reading papers and writing abstracts. After just two years of enrollment, you can independently take charge of a research project under the guidance of a supervisor.

"To do scientific research is not to go to Beijing Tianqiao to watch tricks, where it is lively, but to do original innovation"

In the sci-fi movie "Super Marines", there is a "big white" intelligent companion robot that can accompany the male protagonist to study and play together, and when the male protagonist is emotionally lost, the "big white" will comfort him and give him a big hug.

Unlike an agent that can only complete a single task, "Dabai" is a general-purpose agent. This is exactly what Zhu Songchun is striving for: to create a universal agent with the ability to self-perceive, cognitive, decision-making, learn, execute, and cooperate socially, while conforming to human emotions, ethics, and morals.

"To achieve this goal, we must change the 'parrot paradigm' based on 'big data, small tasks' to the 'crow paradigm' based on 'small data, big tasks'." Zhu Songchun explained that parroting is a simple imitation achieved through repeated training, knowing it without knowing why; Crows drinking water seems simple, but it belongs to the behavior of independent reasoning, which is advanced intelligence and the future development trend of artificial intelligence.

With the support of Beijing Municipality and the Ministry of Science and Technology, the establishment of the General Institute is progressing smoothly. Zhu Songchun proposed that with Chinese thought, the science and technology of creating the world, find the foundation of the development of artificial intelligence discipline from China's excellent philosophical thought, and carry out organized scientific research with original scientific research paradigm and original technical route. As soon as the news of Zhu Songchun's return to China to establish the Tongyuan came out, it attracted the attention of peers at home and abroad.

"It's time to do a big job!" In a WeChat message to Yang Yaodong, assistant professor at King's University London's School of Information, Zhu Songchun not only talked about his scientific research ideals and the significance of general artificial intelligence to the country, but also carefully arranged for the young scholar to return to China to teach. "I also want to contribute to the country's general artificial intelligence research like Mr. Zhu." Yang Yaodong immediately resigned from his teaching position and returned to China to become a researcher in the General Academy, and at the same time served as the class teacher of the 2022 Peking University Tong class, taking on the dual burden of scientific research and talent training.

Like a huge magnet, Tongyuan has attracted more than 200 scholars to join, including more than 30 doctoral students graduated from famous overseas universities, and established a number of research teams such as general vision, natural language and human-computer interaction, machine learning, cognitive computing and common sense reasoning, multi-agent, and robotics.

"Tongyuan has deployed forces in each research direction, and we can freely combine with colleagues according to the needs of cross-research to jointly tackle cutting-edge topics." Yang Yaodong said. It can not only continue its own research in the field of AI segmentation, but also exchange and collide with outstanding talents in other fields, and the role of talent aggregation on the platform of Tongyuan is obvious.

"Doing scientific research is not about going to Beijing Tianqiao to watch tricks, and go wherever it is lively." Zhu Songchun often uses his own experience to inspire everyone, "To do original innovation, we must have a global view and forward-looking thinking, predict the frontier trend of science in advance, and set the ambush on the necessary road in order to achieve a leading edge in one fell swoop." ”

"Don't follow the crowd, dare to be the first in the world." This is what many people say about Zhu Songchun. When the United States participated in the major scientific research project of "explainable artificial intelligence", most of the other international experts used the popular visual deep learning scheme in the academic community, and Zhu Songchun led the team to create a new paradigm of "real-time two-way value alignment between man and machine". "At first everyone thought it was impractical and unattainable." Zhu Yixin said that Zhu Songchun led everyone to overcome many difficulties and make breakthroughs, and the relevant research results were finally published on the homepage by the international academic journals "Science" and "Science Robotics" at the same time, which was evaluated as "a big step forward in AI research in the field of human-machine teamwork".

"Thirty meritorious names are unified one by one, and 80,000 miles of Huai family country." Zhu Songchun's office hangs such a calligraphy work, which accurately interprets his scientific research and pursuit of dreams in the past 30 years, and also expresses his deep feelings of serving the country. Looking to the future, Zhu Songchun is proud: "Our goal is to build Beijing into a source of AI innovation with global influence. (Reporter Shi Fang)

Source: People's Daily - Overseas Edition