Recently, UCLA professor Zhu Songchun returned to China and joined the Department of Automation of Tsinghua University. At the same time, it will also work with Tsinghua University and Peking University to establish the "Beijing General Artificial Intelligence Research Institute" and serve as the dean. Zhu Songchun is a top scholar in the field of computer vision. His return to China will bring strong impetus to the development of domestic artificial intelligence.
Recently, according to a public list of new recruits that Tsinghua University plans to hire from the Internet, Zhu Songchun, professor of statistics and computer science at the University of California, Los Angeles (UCLA) and director of the UCLA Center for Computer Vision, Cognition, Learning and Autonomous Robotics, intends to join Professor of Automation at Tsinghua University.
In addition, Professor Zhu Songchun will return to China as a "National Strategic Scientist" and will also be invited to establish a private non-profit organization with Tsinghua University and Peking University, Beijing General Artificial Intelligence Research Institute, and serve as the dean.
The institute will focus on cutting-edge artificial intelligence technologies, and strive to cultivate interdisciplinary and original artificial intelligence talents in the field of artificial intelligence, while building a new generation of general artificial intelligence platforms.
To build a general-purpose strong artificial intelligence is what Zhu Songchun has been pursuing, and the current artificial intelligence starts with the breakthrough progress of computer vision.
Songchun Zhu received his Ph.D. in computer science from Harvard University in 1996. He studied under Professor David Manford, an international mathematics master. He has published more than 300 papers in top international journals and conferences, and won three times the highest international award in the field of computer vision-Marl prize.
Zhu Songchun has his own unique view on computer vision, and has made important contributions in the field of cognitive science, such as visual common sense reasoning and scene understanding.
He believes in a sentence, "If a nation forgets its history, she is destined to lose its future." This sentence is equally thought-provoking for computer vision.
He mentioned that there are many newly published visual papers, and very few of them can cite documents from 5 years ago. They all cite articles on arxiv in the past two years to compare some benchmarks.
Few people seriously read the papers 10 years ago, 20 years ago, or even 30 years ago. Some ideas and frameworks at that time still have important meaning for current research. Almost everyone uses the same method Than the precision after the decimal point.
Everyone is quite short-sighted, and only pays attention to the history and popular methods of the past few years, and it is impossible to inherit this discipline. Especially after the current wave of methods ebbs, these people will gradually lose their foundation and originality.
Speaking of his academic career, he believes that David Marr had the most profound influence on him.
Since the beginning of the 1960s, many people have studied optic neurophysiology and psychology, and some have done some edge detection work. But what problems does computer vision solve? How to achieve? Everyone is not in agreement, it is not clear.
David Marr has divided three levels to solve this problem, namely calculation (in fact, it should be said to be expression), algorithm, and implementation.
First, at the level of expression, how to write it as a mathematical problem. What is the mission? What is the output? This is independent of the method of solving the problem.
Secondly, when solving this mathematical problem, you can choose different algorithms, either in parallel or in series.
Thirdly, how an algorithm is implemented on hardware can be implemented with CPU, DSP, or neural network.
In addition, David Marr also clarified what vision is to calculate.
Marr has proposed a series of expressions, from primal sketch (primary minimalist graph), to 2 ½ D sketch (deep minimalistic graph), to 3D sketch.
It also includes texture, stereo vision, motion analysis, surface shape, and so on. Marr believes that visual computing is not simply seeking a solution, but a continuous process of calculation. The more you look and think about it, the more understanding you may get.
It is worth mentioning that Marr was diagnosed with acute leukemia in the winter of 1978. After learning that there is not much to come, Marr hurriedly compiled a book "Vision: Studying Human Visual Information Expression and Processing from the Perspective of Computing". He was only 35 years old when he died.
Zhu Songchun and his colleagues spent 8 years on this book, converting the early visual concepts proposed by Marr, including textures, image primitives, and primitive minimalistic images, into a unified mathematical model.
From then on, vision can be studied from a purely theoretical and computational perspective.
In addition to visual statistical modeling and computational theories, Zhu Songchun also implemented a parsing calculation framework for images and scenes, and expanded the syntactic pattern recognition theory of the founder of pattern recognition, Mr. Fu Jingsun.
Since 2010, Zhu Songchun has combined computer vision with cognitive science, natural language understanding, robotics and other disciplines to explore what he calls the "dark matter of artificial intelligence"-95% of intelligence that cannot be observed through perceptual input.
Now, Zhu Songchun's team has built a large-scale, physically realistic VR/AR environment for training and testing autonomous AI agents responsible for performing a large number of daily tasks.
These agents can integrate the capabilities of vision, language, cognition, machine learning and robotics, develop physical and social common sense in the process, and use cognitive architecture to communicate with humans.
People who are familiar with Professor Zhu Songchun never hesitate to praise him for his rigorous academic spirit.
Dai Jifeng, a researcher in the Vision Group of Microsoft Research Asia, visited Professor Songchun Zhu’s VCLA laboratory for more than a year, and shared some of the academic life of Professor Songchun Zhu.
Have first-class intuition for the general direction of the visual field
Professor Zhu Songchun's experiment has a large number of students, so it is natural to use strong funding to support it. In recent years, Professor Zhu is probably the most funded professor in the visual field of American universities (I don't know if it is necessary to add "one").
Since 2011, Mr. Zhu’s laboratory has received more than US$40 million in funding as a PI. The main reason for this is his "advanced research thinking."
Being able to get these big funding means that Professor Zhu "has first-class intuition and a leading and accurate grasp of the general direction of this field."
And Professor Zhu's "sixth sense" has been manifested many years ago.
In 2012, a large MURI project hosted by Professor Zhu Songchun was held at UCLA. He took the stage to talk about "vision meets language", saying that the combination of vision and language would be an important issue. For example, when you see an entire picture, the system should output Describe it in a paragraph, for example, when you see a bounding box area, you want to describe what happened inside, and how to implement this with a hierarchical And-Or graph.
At that time, many big shots in the field of vision felt a little fanciful. Unexpectedly, after a year or two, this will be the smashing VQA task, but it is realized with a neural network.
"Able to perceive the general direction of the future in advance." This is the top research feeling, and this is what Prof. Zhu Songchun is best at.
Professor Zhu Songchun’s general direction is wrong, but the probability of being correct is already high.
Meticulous in mathematics (especially statistics)
When discussing with Professor Zhu, the most often challenged is "this algorithm is wrong, mathematically wrong, and the latest technology in the CV field is statistically wrong"
For most researchers, probabilistic models are popular for probabilistic models, SVM is popular for SVM, and neural networks are popular for neural networks.
And Professor Zhu Songchun has faith, that is, his "probabilistic model", which once led the trend in the visual field before SVM. It's their own thing, so it's not as easy to abandon like others.
Professor Zhu once said, "Doing research is like playing Go. You can't play one game from the east and the other, because the territory is all occupied by others."
Tofu heart to student knife mouth
This is the most controversial place for Professor Zhu.
When I first went to his laboratory, he would be very uncomfortable with his criticism, but you can get to know him slowly.
He is very good for the long-term development and important interests of the students; although he is uncomfortable when criticizing, he will not hold grudges afterwards; there is also a balance and reconciliation with Professor Wu in the laboratory.
In fact, bosses in the academic circle are more temperamental and push a lot of students. This is a common problem among researchers.
But at the critical moment when looking for a job, he and Professor Wu from the same laboratory are very supportive and human.
Professor Zhu Songchun's daughter gave up her American citizenship and became a Chinese citizen when she reached the age of 18. Perhaps since then, Professor Zhu Songchun's plan to return to China has been put on the agenda.
Zhu Songchun's return to China this time will bring strong impetus to the development of domestic artificial intelligence, especially general artificial intelligence. He is also one step closer to the dream of "a unified theory of artificial intelligence."
Reference link:
http://www.stat.ucla.edu/~sczhu/research_blog.html#VisionHistory (Zhu Songchun: True Sources | A Probe into the Three Sources of Computer Vision and Artificial Intelligence, 2016)
https://www.zhihu.com/question/59182074 (The evaluation in the text comes from the Zhihu answer of Microsoft researcher Dai Jifeng)