Keywords: Intelligent Computing System Artificial Intelligence Professional Course
Editor's Note: Five years ago, Chen, the author of this article, was invited to publish an article entitled "Dream of Artificial Intelligence for Architectural Researchers" 1, which publicly introduced his mental journey from general-purpose processor to deep learning processor for the first time (that article was later included in "Selected Articles"). After five years of hard work, Chen has made internationally recognized contributions to the direction of deep learning processors: his academic papers have been cited by more than a quarter of the papers in top conferences 18, and his technical achievements have been applied to nearly 1 100 million smartphones and servers such as Huawei, Shuguang and Ali. Therefore, he was appraised as the "pioneer" and "leader" of smart chips by Science magazine. Nowadays, when the deep learning processor has become a hot spot in academia and industry, Chen turns around again and moves towards a new stage of artificial intelligence system course teaching. This article will share with you his mental journey towards this new stage.
Setting up the course of intelligent computing system
I am a young researcher at the Institute of Computing Technology, Chinese Academy of Sciences, mainly engaged in basic research on the cross direction of computer system structure and artificial intelligence. You may feel a little strange when you see the title and author unit of this article: What is the relationship between the author and the artificial intelligence professional course? Because in the traditional sense, attending classes is the responsibility of teachers in colleges and universities, the staff of Chinese Academy of Sciences are not obliged to attend classes unless they voluntarily serve as post professors of HKUST.
The main reason that drives me to actively think about the curriculum design of artificial intelligence specialty is such a well-known phenomenon: the more the upper layer of artificial intelligence (algorithm layer and application layer, see figure 1) is studied, the greater the contribution of China researchers to the world; The lower the level (system level and chip level), the less the contribution of China researchers. In all kinds of ImageNet competitions, the algorithm models of many domestic institutions have shown the trend of "dominating the list", which can be said to represent the world's leading level. However, most of these algorithm models are developed based on CUDA 2 programming language, Tensorflow programming framework and GPU. Among these low-level "hard technologies", the contribution of China researchers to the world is relatively small. The lack of research ability at the bottom will not only hinder the basic research of artificial intelligence in China, but more importantly, it will make China's intelligent industry a castle in the air and embark on the old road that the information industry is subject to core chips and operating systems.
Figure 1 General level of artificial intelligence research
There are many reasons for the lack of basic research ability of artificial intelligence in China, and I think the most fundamental reason lies in the lack of education and training of talents in this field. Without fertile soil, no towering trees can grow. Without students majoring in artificial intelligence with systematic thinking, it would be difficult for China to produce intelligent computing system masters like Jeff Dean, and it would also be impossible to produce internationally competitive intelligent system products. Therefore, in the cultivation of artificial intelligence students, we should actively change the status quo.
20 18, 35 universities across the country set up artificial intelligence undergraduate majors, which is a great opportunity to rethink and sort out the artificial intelligence training system. As far as I know, most colleges and universities adopt pure algorithm and pure application teaching ideas when considering the artificial intelligence curriculum system. The students trained in this way are still biased towards upper-level application development and lack a comprehensive understanding of intelligent computing systems.
In fact, there are many people of insight in colleges and universities, and bypassing the system courses in curriculum design is often subject to three objective difficulties: first, there are not many artificial intelligence system courses available for reference in China; Second, there is a shortage of teachers in domestic artificial intelligence system courses; Third, there is a lack of teaching materials for artificial intelligence system courses in the world.
Based on my own research background, I have some superficial understanding of the algorithms and systems of artificial intelligence. Can you do something to solve the difficulties in artificial intelligence system courses, teachers and teaching materials? Is it possible to train some artificial intelligence students with systematic thinking and ability?
Therefore, in the 20 18 academic year, I applied to the University of Chinese Academy of Sciences to offer a systematic course of artificial intelligence, named "Intelligent Computing System" (once named "Intelligent Computer"), hoping to cultivate students' comprehensive understanding of the complete software and hardware technology stack of intelligent computing (including basic intelligent algorithms, intelligent computing programming framework, intelligent computing programming language, intelligent chip architecture, etc.). This course is very popular with students, and many students who choose not to take part in the course follow the whole semester. What moved me in particular was that students from other research institutes spontaneously came to Huairou to attend classes from Zhongguancun. It took three hours to drive back and forth from the last class, and it was already late at night when I returned to Zhongguancun. This may show that this course is attractive to students, and everyone has a real gain from listening to the class.
Cultivation and curriculum system of artificial intelligence students
The curriculum system design of artificial intelligence specialty should serve the training goal of students. So what kind of students should be trained for artificial intelligence major in colleges and universities?
There may not be a unified answer to this question. Compared with the computer major which is very close to the artificial intelligence major, it is obvious that the computer major in colleges and universities trains researchers, designers and manufacturers of computer machines or subsystems instead of computer users.
When the predecessors of computer specialty in China started to set up computer specialty more than 60 years ago, they had a far-sighted plan to train researchers, designers and manufacturers of computer machines or subsystems. After 60 years' evolution, this scheme has basically maintained the original intention of that year. Nowadays, computer majors in colleges and universities basically offer compulsory courses in computer composition principle, operating system, compilation principle, computer architecture and other systems (see Figure 2). In other words, although most computer majors are engaged in software development after graduation, they still have a basic understanding of computer hardware systems.
Figure 2 Netease Cloud Classroom 3 Computer Professional Training Program
The training goal of students majoring in artificial intelligence should be researchers, designers and manufacturers of artificial intelligence systems or subsystems. Only by achieving this goal can the talents trained in colleges and universities continuously and comprehensively support the industry and research of artificial intelligence in China. In order to achieve this goal, the curriculum design of artificial intelligence major should be both soft and hard (just like computer major). If the artificial intelligence major only offers courses such as machine learning algorithm and audio-visual application, it can only be regarded as "artificial intelligence application major" or "artificial intelligence algorithm major" at best. After all, the algorithm is just the tip of the iceberg, and 90% of the iceberg is still hardware and systems.
Take Google as an example. This company has the largest AI algorithm research team in the world. Google Chairman john hennessy is a computer architecture scientist and Turing Prize winner. Jeff Dean, general manager of Google AI, is a computer system researcher; The three most remarkable advances of Google AI (Tensorflow, AlphaGo, TPU) are all systems, not just a specific algorithm, and the algorithm is only a part of the system. Therefore, from the international academic mainstream of artificial intelligence, systems are as important as algorithms.
Students who have only studied algorithms may be good at adjusting model parameters, but they have no feeling about the time-consuming and labor-consuming of an algorithm. Such students do not have the basic skills to apply an algorithm to industrial systems (because if an algorithm is really to be used, it must meet the limitations of delay and energy consumption).
Only by joining the course of system line can students really understand how artificial intelligence works, including how an artificial intelligence algorithm calls the programming framework, how the programming framework handles the operating system, and how the operators in the programming framework run on the chip step by step. Such students can build complex systems or subsystems by themselves, which will have greater potential in scientific research and stronger competitiveness in the industry. As the insiders said, "Tensorflow will be used to earn 300,000 RMB per year, and Tensorflow will be designed to earn 300,000 USD per year."
Many teachers told me: "There really should be some systematic courses for artificial intelligence majors. But there has never been such a class in China, and there is no suitable teaching material. Our college also lacks corresponding teachers to teach such courses. " These are three practical objective difficulties, but they should not affect the curriculum design of our artificial intelligence major.
From the students' point of view, what courses the artificial intelligence major offers should be based on what the state and enterprises need from the students' union, rather than what the current teachers will do. Sixty years ago, few people in our country had seen computers, let alone started classes. However, for scientific and engineering computing tasks such as two bombs and one satellite, China still established the Institute of Computing Technology of Chinese Academy of Sciences, held training courses for computer teachers in the Institute of Computing Technology, and opened computer majors in universities such as Tsinghua and China University of Science and Technology (I can't list them here), which led to the huge computer industry in China today.
From the teacher's point of view, the systematic research of artificial intelligence has become an international academic hotspot. Teaching such courses is a process of mutual learning, which can help teachers walk to the forefront of international academics. This year, American computer researchers from Top4 University (Stanford University, Carnegie Mellon University, University of California at Berkeley and Massachusetts Institute of Technology) and other international institutions jointly published a white paper called "Machine Learning System". Training a group of young teachers in such an emerging hot direction will undoubtedly be of great help to enhance the influence of universities and even China in the academic field of artificial intelligence.
Therefore, it is necessary and feasible for students and teachers majoring in artificial intelligence to offer systematic courses.
What is an intelligent computing system?
Simply put, intelligent computing system is the material carrier of artificial intelligence. At present, intelligent computing system is usually a heterogeneous system integrating general CPU and intelligent chip (NVIDIA GPU or CAMBRIAN MLU, etc.). ), and provides an intelligent computing programming framework and programming language for developers. The reason why intelligent chips should be added to the general CPU is mainly because it is difficult for the general CPU to meet the increasing speed and energy consumption requirements of artificial intelligence computing. For example, 20 12, Google Brain spent16,000 CPU cores running for several days to train how to recognize cat faces, which is difficult to accept for industrial applications. Obviously, in order to really use artificial intelligence technology, heterogeneous intelligent computing systems must be used. In order to reduce the programming difficulty of heterogeneous intelligent computing system, it is necessary to have a programming framework and programming language for intelligent computing.
In fact, intelligent computing systems have been widely infiltrated into our lives in various forms. IBM's supercomputer Summit uses machine learning to predict the weather, BAT's data center runs a large number of advertising recommendation tasks, Huawei's mobile phone integrated CAMBRIAN deep learning processor handles image analysis and speech recognition, and Tesla's autopilot system … can all be regarded as intelligent computing systems. In the intelligent age, China and even the whole world need a large number of developers, designers and users of intelligent computing systems.
The development of intelligent computing system will not happen overnight. In 1980s, special computers (Prolog and LISP) oriented to symbolic intelligent processing can be regarded as the first generation of intelligent computing systems. But at that time, artificial intelligence lacked practical application and the algorithm was not mature. Moreover, Moore's Law was still in the stage of rapid development at that time, and special-purpose computers did not have much advantage over general-purpose CPU whose performance could double every 18 months. Therefore, the first generation of intelligent computing system gradually withdrew from the historical stage.
The course Intelligent Computing System focuses on the second generation intelligent computing system, which is mainly used for machine learning tasks such as deep learning. Compared with the first-generation intelligent computing system 30 years ago, today's second-generation deep learning intelligent computing system can be said to have met the weather, location and people. At present, artificial intelligence applications such as image recognition, speech recognition, natural language understanding, games, and advertising recommendation have begun to land, and the development speed of deep learning algorithms is overwhelming. More importantly, the performance development of general CPU has stopped. To support the continuous development of deep learning algorithms, we must rely on intelligent computing systems. Therefore, deep learning intelligent computing system will be the focus of academic and industrial circles for a long time. The course "Intelligent Computing System" will help students understand the intelligent computing system of deep learning.
The second generation intelligent computing system mainly supports machine learning tasks such as deep learning. In the future, if human beings really want to take a step forward on the road of general artificial intelligence, the algorithms that the third generation intelligent computing system needs to support in the future will go far beyond the scope of machine learning, and must include advanced cognitive intelligence algorithms such as association, reasoning and emergence. My personal guess is that the third generation intelligent computing system may be a virtual world environment that incubates general artificial intelligence. The course "Intelligent Computing System" may stimulate students' curiosity and attract them to devote themselves to the research of the third generation intelligent computing system in the future.
General situation of intelligent computing system course
The course "Intelligent Computing System" is mainly for senior undergraduates or graduate students majoring in artificial intelligence, computer and software engineering. The goal of this course is to train students to fully understand the complete software and hardware technology stack of intelligent computing (including basic intelligent algorithms, intelligent computing programming framework, intelligent computing programming language, intelligent chip architecture, etc.). ) and become the designer and developer of intelligent computing system (subsystem).
The preface of the course includes C/C++ programming language, introduction to computer composition principle and algorithm (or machine learning). The class hours of the course are flexible, which can be the last semester of the university (40 class hours, see table 1), the short semester can be concentrated for one week (20 class hours), and other machine learning courses can be embedded as a supplement. For the short-term study of 20 hours, the course hopes that students can "know what it is" about intelligent computing systems, mainly for practical operation; After 40 hours of long-term study, the course hopes that students can "know why" the intelligent computing system, so the mechanism should be thoroughly explained.
Table 1 Intelligent Computing System Course Outline (40 class hours)
In course teaching, we should adhere to two principles. The first is application-driven. A good engineering course should apply what it has learned, especially the course "Intelligent Computing System". It is basically ineffective to learn some theorems and formulas only after the course. Another principle is full stack penetration. In the past, there was a problem in the curriculum design of computer major, that is, there was obvious compartmentalization, such as the separation of operating system and computer architecture, what requirements the operating system put forward for the computer architecture, and what support the computer architecture had for the operating system. None of the courses linked these together. As a senior undergraduate (or graduate) course, "Intelligent Computing System" has the obligation to help students connect all the past knowledge of artificial intelligence software and hardware to form an overall understanding.
For the course "Intelligent Computing System", the driving paradigm is a grasping hand. In the process of National Science and Technology University, we chose video style transfer as a driving example. Simply put, style transfer can keep the basic content of each frame of a video, but change the painting style of the picture (for example, from ordinary photos to Picasso style or China ink painting style, as shown in Figure 3). For students, this is an interesting and driving example within their ability.
Fig. 3 Driving example of changing from ordinary photos to Picasso style
Focusing on how to realize the real-time style migration of video, we take students to write algorithms step by step and transplant them into the programming framework, write operators for the programming framework, design chips for operators, build a multi-chip system, evaluate the advantages and disadvantages of this system from the aspects of speed, energy efficiency and accuracy, and then optimize the system in closed loop iteration. Finally, we will give you an experimental environment of intelligent computing system, including camera and intelligent chip development board, so that students can realize a "semi-finished" application of real-time conversion of painting style from video shot by camera.
label
My mother is a middle school teacher. When I was growing up, my predecessors and Hu, who helped me a lot, were also famous teachers who were immersed in the front line of teaching all the year round and had great teaching enthusiasm. From childhood, the words and deeds of these elders made me deeply feel that education is a great cause, which can profoundly change students, industries, society, country and mankind. The knowledge of artificial intelligence that we teach students today may affect the competitiveness of our country in the intelligent age tomorrow. Therefore, although I have many basic research tasks at hand, I am still willing to take the cultivation of artificial intelligence system talents as my most important mission in the future, and spend most of my time and energy on the teaching and promotion of the course "Intelligent Computing System" in various universities.
It is very gratifying that the new course "Intelligent Computing System" has been supported and encouraged by many teachers and students, although there are still many defects. We have or will jointly offer this course in Chinese Academy of Sciences, Peking University, Beihang University, Tianjin University, China University of Science and Technology, Nankai University, Beijing Institute of Technology and Huazhong University of Science and Technology. This year, all PPT, handouts, teaching materials, videos, codes, cloud platforms and development boards of this course will be open for teachers to criticize and correct. You are very welcome to send me an email and put forward your valuable opinions.
In the future, I hope to cooperate with more colleges and universities that train students majoring in artificial intelligence, participate in the exchange and discussion of artificial intelligence system courses extensively, and improve the teaching level of artificial intelligence system courses. I believe that through the joint efforts of all of us, the practical difficulties in setting up the course of artificial intelligence system will be solved, so that the artificial intelligence talents trained in China in the future will have no technical shortcomings.
As a young teacher, I have a huge gap with my predecessors in many fields of education in teaching ability and experience. I need to learn more. The course "Intelligent Computing System" can at most play the role of cultivating the ability of artificial intelligence system in China. Just as Mr. Lu Xun wrote a passage for young people in Hot Wind Random Thoughts: "If there is a little heat and a little light, it will become a firefly, or it can emit a little light in the dark ... If there is a torch and the sun comes out, we will naturally ... praise the torch or the sun at will; Because he takes pictures of humans, even me. "
Footnote:
1 this article was published in the 5th issue of China Computer Federation Newsletter (CCCF), 20 14. /Institution/Institution Details? id=37388758630748 16。 _ack= 1 .
2 CUDA: Compute unified equipment architecture, compute unified equipment architecture.
See also.
Click "Read the original" to join CCF.