On December 21, the "2019 New Generation Artificial Intelligence Academician Summit Forum" hosted by Pengcheng Laboratory and the New Generation Artificial Intelligence Industry Technology Innovation Strategic Alliance entered the second day, with 12 keynote speeches and 2 roundtable forums. Panel pushed the forum. To the climax. In the morning, experts from Alibaba, University of Science and Technology of China, Tsinghua University, Institute of Automation, Chinese Academy of Sciences, and Pengcheng Laboratory shared the latest developments in their respective fields around AI chips and brain-like brains; in the afternoon, from China Ping An, Health Commission, Traffic Police, Ant Experts from Jinfu, the Institute of Computing Technology of the Chinese Academy of Sciences, and Yuntian Lifei focused on artificial intelligence and artificial intelligence empowerment, and deeply analyzed the current development of the artificial intelligence field and the direction worth exploring in the future. Two wonderful round-table forums also discussed the convenience of artificial intelligence to our lives from all aspects of daily life, transportation, social management, medical services, and finance.
This conference continues the strong lineup of academicians last year. This year, as many as 10 academicians and many top experts in the field of artificial intelligence and the top experts in the field of artificial intelligence have been invited to analyze the development status of artificial intelligence field and the direction worth exploring in the future. .
1. Academician Pu Muming: Research on Brain Science and Brain-like Intelligence
As the first guest speaker at the conference, Pu Muming, a member of the Chinese Academy of Sciences and a foreign member of the American Academy of Sciences, brought a report on the topic of "Brain Science and Brain-like Intelligence Research".
Pu Muming, academician of the Chinese Academy of Sciences and foreign academician of the American Academy of Sciences
Academician Pu Muming said: "I think a very important development direction of the next generation of artificial intelligence is the new type of artificial intelligence inspired by the brain. The major project that our country will launch in the next 10 years is called brain science and brain-like research. The overall framework is "One body and two wings", the main body is the neural basis of brain cognitive function, which makes the whole brain nerve connection map, and the two wings means that the research content is divided into the diagnosis and intervention of major brain diseases, and the research and development of brain-computer intelligent technology."
He pointed out that understanding the brain requires the analysis of neural connections at three levels:
The first is the macro map. Through MRI technology, millimeter-level nerve bundles can be seen. Each nerve bundle has thousands of cell fibers, and the conduction direction is bidirectional, but only understanding the direction of the nerve bundle does not contribute much to the understanding of brain functions.
The second is a mesoscopic map, with a spatial resolution of micron level. Use special methods to label different neuron types and understand the functions of different neurons. The Mesoscopic Neural Connection Atlas is the main direction of neuroscience at present.
The third is the microscopic atlas, the spatial resolution must reach the nanometer level. A lot of useful information can be obtained by studying the distribution of neuronal axons and dendrites and the rules of synapse generation at the microscopic level.
Furthermore, he believes that three levels of cognition should be considered in the study of brain-like intelligence:
The first level is the cognition of the outside world, including perception, multi-sensory integration and attention, classification, etc. It is a cognitive ability that many animals have. For this, you can find some information and inspiration from animal models;
The second level is the cognition of self and non-self, including self-awareness, empathy, and the ability to understand others, but it may only be available in primates, so research can only be based on primates, and this It is the most advantageous field in my country's future basic neuroscience research;
The third level is the cognition of language. This is the syntax, grammar and infinite open language structure that only humans have. It is for studying the neural mechanism and evolutionary origin of human language, and constructing non-human primate transgenic animal models. necessary.
For these brain network characteristics, what can be learned from the future research of artificial intelligence, but is there no place to learn from it? Academician Pu Muming believes that there are five important points:
First, neurons are very simple, with different properties and types of neurons, such as inhibitory neurons (information reversal), excitatory and inhibitory neuron subtypes (different discharge characteristics);
Second, neural networks have forward, reverse, and lateral connections. Now they are mainly forward, and adding reverse connections will bring benefits, and lateral connections are especially important and orderly.
Third, the plasticity of nerve synapses. In addition to the enhancement of transmission efficiency and the weakening of functional plasticity (LTP and LTD), the structural plasticity of synapse regeneration and trimming is also very useful.
Fourth, the storage, retrieval, and fading of memory are also very critical. The efficiency and structural modification of a specific synaptic group in the network is the storage of memory. Memory theory refers to the reproduction of electrical activity in the synaptic group that stores memory, and memory fading It means that synaptic function and structural modification fade over time. In addition, according to the output, the synapse group related to learning can be modified to perform reinforcement learning.
Fifth, the application of the concept of Heber neuron clusters, including the establishment of cell populations, the integration of multi-modal information, the binding of different information, synchronized information oscillations, time phase differences, and the map structure of input information can all be used for reference.
Finally, he put forward four suggestions for new artificial network architecture and machine learning algorithms:
One is to get rid of the temptation of deep learning network (DNN);
The second is to establish a new artificial intelligence model and algorithm with the goal of high efficiency, energy saving, semi-supervised and unsupervised learning;
The third is based on the Spike Neural Network (SNN), adding transmission delay to emphasize the importance of timing information;
Fourth, based on a simple (few layers) network, each time the characteristics of a natural neural network are added to generate a new architecture, and then new machine learning and algorithms are used to detect the effect.
In addition, he believes that the AI community should propose a new robot or intelligence standard to try to establish a "New Turing Test" from more dimensions such as the integration of language and perceptual abilities, and teamwork.
2. Academician Gao Wen: Digital Retina and Cloud Vision System Evolution
Another academician of this forum-Gao Wen, Academician of the Chinese Academy of Engineering and Director of Pengcheng Laboratory, followed Academician Pu Muming's introduction to neural networks, focusing on how to use the understanding of neural networks to improve the urban brain or smart city system For some problems existing in the existing cloud vision system, his report topic is "Digital Retina and Cloud Vision System Evolution".
Gao Wen, Academician of Chinese Academy of Engineering and Director of Pengcheng Laboratory
First, he pointed out the two main challenges that currently exist in cloud vision systems:
First, although there is a lot of video data, there are not many big data that can standardize it and dig out rules from it.
Second, most of these video data are normal videos, and there are few sensitive videos, so the value generated is not great.
In the end, these two issues are actually direct consequences of the entire visual perception system architecture. In this regard, Academician Gao Wen and others sought inspiration from the evolutionary history of the visual system of advanced living organisms. When designing a new second-generation urban brain or cloud visual system, they worked on the visual nerve channel in the middle. They call this idea the digital retina.
He further introduced that the definition of digital retina includes eight basic elements, which can be divided into the following three groups according to characteristics or functions:
The first group of features is a globally unified time-space ID, including two basic elements, a unified time on the entire network and precise geographic location;
The second group of features is a multi-level retinal representation, including three basic elements: video coding, feature coding, and joint optimization;
The third group is that the model is updatable, adjustable, and definable, that is, the three basic elements of model updatable, attention adjustable, and software definable are combined.
Unlike traditional cameras that have only one stream, that is, a video compression stream or a recognition result stream, this digital retina has three streams, namely video encoding stream, feature encoding stream, and model encoding stream. The three have their own division of labor, and some are in The front end can control and adjust in real time, and some are adjusted and controlled through cloud feedback.
In the past two or three years, Academician Gao Wen and others have also carried out a lot of work on this, focusing on the following four important developments in enabling technologies:
The first enabling technology is high-efficiency video coding technology, in which the scene coding technology they proposed has become a major contribution to the world made by Chinese scholars, raising the coding efficiency to a new level;
The second enabling technology is feature coding technology, which includes two international standards CDVS and CDVA that they completed with international experts, and CDVA is a standard established to support deep networks.
The third enabling technology is video and feature coding technology. This is because video coding and feature coding use different optimization models. Video coding uses a 2D optimization model, while feature coding uses the RA model. The curve direction of the model is completely different. For this, they proposed a joint optimization model, which turns RA and 2D into an objective function, and the joint optimization can be achieved by seeking the optimal solution.
The fourth enabling technology is CNN model coding technology, which can achieve advantages such as multiple model reuse, model compression (differential coding), and model update.
Finally, Academician Gao Wen concluded that the current cloud vision system is not very effective. For this, we can make it more effective through new concepts and technologies like digital retina, including reducing bit rate, reducing latency, and improving Accuracy, reduce cloud computing costs, and convert low-value video data into big data.
As Digital Retina has achieved version 1.0, how should we move towards version 2.0 next? The answer is: use the idea of spiking neural network to make Digital Retina 2.0.
3. Tang Xiaoou: Artificial Intelligence, Innovative Application
After the wonderful reports of the two heavyweight academicians, Tang Xiaoou, the founder of SenseTime Technology and a professor at the Chinese University of Hong Kong, brought a report with the theme of "Artificial Intelligence, Innovative Application" in a humorous speech style as always.
Tang Xiaoou, founder of Shangtang Technology and professor at the Chinese University of Hong Kong
"I don't know how to give a speech anymore. People often only remember the jokes I told. In fact, I also mentioned some profound philosophy of life."
There was a burst of laughter from the audience very "according to the situation."
In response to the theme, Professor Tang Xiaoou first shared his understanding of "innovative applications":
Innovation is to do things that others have not done before, and the result is to spend money, so in this sense, the main purpose of innovation is to spend money, which seems to be the function of universities;
Application is to implement innovation into the product and then sell the product. The core is to make money. This is what the company should do.
But how do the two seemingly contradictory points of innovation and application coexist in the company? He pointed out that for most companies in the world, especially startups, it is impossible for large companies such as Google to earn a large amount of income through mature business applications, thus investing huge R&D expenses for innovation every year. Pay the bill. How to find your own survival mode?
SenseTime uses AI+ empowerment to find a methodology that balances innovation and application by cooperating with all walks of life. He furthered the daily needs of the above-mentioned workers "from 8 in the morning to 10 in the evening", such as driving to work, building office, lunch, going out for meetings, games and entertainment, medical consultation, leisure shopping and other life scenes. He introduced Shangtang and Cooperation and development results of related industries in various application fields such as autonomous driving, smart parks, indoor navigation, and smart retail.
Based on his own practical experience, Professor Tang Xiaoou also gave his own experience and thoughts on the dilemma of "innovation or application": the innovation we initially understood is generally to publish papers, do research, and do research in the laboratory. Things that have been done, but such innovations are often unacceptable once they are put on the market.
Therefore, he believes that the final innovation is not only technical innovation, but also engineering innovation in the company, product innovation and even business model innovation, sales model innovation, partner cooperation method innovation, and the entire company management model. Comprehensive innovation, including innovation.
"Only in this way can we survive in the cruel market environment and find our own living space in front of extremely powerful competitors. In front, many companies, including SenseTime, have a long way to go. go."
4. Wang Haifeng: Frontier of Natural Language Processing
Dr. Haifeng Wang, Chief Technology Officer of Baidu and Director of the National Engineering Laboratory of Deep Learning Technology and Application, brought a report on "Frontiers in Natural Language Processing" and shared his experience and views of nearly 30 years of research.
Wang Haifeng, Chief Technology Officer of Baidu, Director of the National Engineering Laboratory of Deep Learning Technology and Application
"Natural language processing is a very important and very popular direction for artificial intelligence." Dr. Haifeng Wang said that natural language processing is the theory, technology and method of using computers to simulate, extend and expand human language capabilities. "Development Plan" is also listed as a common key technology.
Then Dr. Wang Haifeng reviewed the research and development history of natural language processing from artificial rules to deep learning models:
Natural language processing based on artificial rules requires domain experts and domain knowledge, and this knowledge is modeled into a computer system, and the development and migration costs are very high;
In the period of statistical-based natural language processing, automatic training and model selection can be realized to a certain extent. At that time, there were also many feature engineers dedicated to constructing various features; on the other hand, a large number of statistical machine learning models were used in different applications Different effects will be achieved, so the choice of the model itself has certain limitations.
In the era of deep learning, natural language processing has become simpler, more standardized and automated. A set of models can get better results for different data. Therefore, compared with previous machine learning models, deep learning is a very important The characteristic is that there is a set of methods that can be applied to different fields and different applications, which is very similar to the human brain.
He further pointed out that computing power, algorithms, and data are the three major elements that drive breakthroughs in natural language processing technology. With the rapid development of natural language processing technology, this research field has also shown many new changes, such as the level from traditional NLP. Formal structure analysis has evolved to direct end-to-end semantic representation; from the past limited to sentence understanding, it has developed to multi-text and cross-modal content understanding; machine translation has achieved a leap in quality and has achieved real-world applications from science fiction ideals.
Subsequently, Wang Haifeng systematically sorted out Baidu's achievements in pre-training language models, machine reading comprehension, multi-modal deep semantic comprehension, machine translation and other natural language processing technologies, as well as open source and opening up, and the fruitful results of promoting industrial applications.
5. Zhang Zhengyou: Intelligent Evolution of Robots
Director of Tencent AI Lab & Robotics X, ACM, IEEE Fellow, Dr. Zhengyou Zhang, as the first guest speaker in the afternoon session, brought a report with the theme of "Intelligent Evolution of Robots".
Director of Tencent AI Lab & Robotics X, ACM, IEEE Fellow Zhang Zhengyou
Before entering the report formally, Zhang Zhengyou recalled that last year, Academician Gao Wen also invited him to the Artificial Intelligence Academician Summit Forum to give a report. When Academician Gao Wen invited him again this year, he also emphasized the sentence "We must do the latest research report." He laughed and joked: "I guess he wants to test me, and want to see if I have made some research progress over the past year or so."
"Artificial intelligence is still in the early spring, not very intelligent, and there are many problems." Then, he took the camera's inability to recognize fake images that block the lens as an example and pointed out that the current artificial intelligence is only learning from a large amount of annotation data, and its generalization ability is poor. .
He believes that with the development and full application of sensor technology, the era of coexistence of humans and intelligent robots will inevitably come. This is also an important reason why he chose to return to China to join Tencent to create Robotics X Robotics Lab.
Dr. Zhang Zhengyou then introduced the six components of the robot, including the body, perception, actuator, dynamic system, interaction system, and decision-making. The future trend of robots is automation, intelligence, and autonomous decision-making in an uncertain environment. For the autonomous decision-making of robots, he proposed the SLAP paradigm, that is, sensors and actuators should be closely integrated, and with the help of learning and planning modules, they can improve their capabilities and make decisions.
Regarding the future breakthrough of intelligent robot technology, he once again mentioned the "A2G theory" shared last year, where ABC represents the basic capabilities of robots, A refers to the robot's ability to see, speak, listen and understand, and B refers to the robot body. C is automatic control; and DEF refers to a higher level of robotic capabilities, D is evolutionary learning, E is emotional understanding, and F is flexible manipulation; the last layer-G is to protect humans. This puts forward requirements for more advanced and smarter robots, and the ultimate goal of robots is to serve people.
Finally, Zhang Zhengyou expressed his vision for the development of robots, that is, human-machine coexistence, co-creation, and win-win. For this, it is necessary to start from "using robots to enhance human intelligence, care for human emotions, develop human physical potential, and realize human-machine collaboration. "Four aspects to create this kind of future.
6. Yan Shuicheng: Transform AI into Affordable Intelligence
Also presenting the second time at the Artificial Intelligence Academician Summit Forum is the Chief Technology Officer of Yitu Technology, IEEE and IAPR Fellow Dr. Yan Shuicheng. In his speech on "Transform AI into Affordable Intelligence", he pointed out the challenges of turning artificial intelligence into "Affordable Intelligence" and shared some work progress from the perspective of chips and models.
Yan Shuicheng, Chief Technology Officer of Yitu Technology, IEEE, IAPR Fellow
Yan Shuicheng mentioned that the core mission of an AI startup and Dachang AI Lab is to achieve the true landing of AI, which requires solving two problems:
- The first is the algorithm. On the one hand, it is necessary to ensure that the algorithm is "usable", that is, the accuracy is high enough to truly unlock a scene; on the other hand, the algorithm must be "enough to use", because now many scenes are already based on monomodal algorithms Unable to provide a satisfactory solution for users.
- The second is computing power. On the one hand, it must be “affordable” for users. For example, AI chips used to support computing must have high enough concurrency performance; on the other hand, it must be “affordable” for users. This kind of computing power must ensure that the power consumption is low enough, otherwise even if the user buys it home, it may be unavailable due to the high electricity bill of the data center.
As artificial intelligence is used in more and more scenarios, and as the technology has reached the stage where it can be used, now in addition to higher and higher requirements for computing power and algorithms, it is also more from Think about it from an angle.
Dr. Yan Shuicheng further pointed out that research shows that the computing power required to train and test artificial intelligence models doubles every three and a half months, which is much faster than Moore's Law. Just as at the NeurIPS conference that just ended this year, one thing that everyone is more concerned about is: As AI is used more and more, the power consumption it brings is also higher and higher. Will this have an impact on the environment?
Therefore, to truly let AI land in a scene, the two most critical engines are:
- The first engine is a high-performance AI model, which is the dimension of the algorithm. There are two ways to obtain a high-performance AI model, one is based on different Motivation models, and the other is based on NAS (Neural Network Architecture Search). In response to this, the main hope is to solve the "Affordable" problem in research and application.
- The second engine is a high-performance AI chip, which is the dimension of computing power. In response to this, chip manufacturers must first follow the algorithm and chip principles to ensure that the chip achieves high performance in a large enough use scenario; secondly, it must predict the most cutting-edge algorithm development trend in the field to ensure that this chip will be in the next few years. Be able to "play your strengths"; in the end, you must make the user's construction cost low enough and affordable.
Finally, he concluded that AI is being applied in more and more scenarios, and the accuracy and goals pursued are getting higher and higher, which puts higher and higher requirements on algorithms and computing power. At this time, AI's "Affordable" problem will become more and more important.
Moreover, if AI is to be converted into "Affordable Intelligence", high-performance AI models and high-performance AI chips are the dual engines that promote this conversion. Only in this way can our end users be able to "afford" and "affordable intelligence". Affordable".
7. Sun Jian: The frontier development of visual computing
Dr. Sun Jian, Chief Scientist of Megvii and the winner of the He Liang He Li Fund Award, gave the title "Frontier Development of Visual Computing", focusing on the introduction of the research process and progress of computer vision from the convolutional neural network and computer vision technology itself.
Sun Jian, Chief Scientist of Defiance, He Liang He Li Fund Award Winner
Convolutional neural networks started relatively early. In the 1980s, a professor in Japan proposed such a concept and developed it. The subsequent research work on convolutional neural networks mainly focuses on four issues:
The first is the convolution problem of neural networks. Now everyone uses 3×3 or 5×5 convolutions. Convolutions have previously experienced the AlexNet network, GoogleNet network, faster R-CNN proposed by Facebook, ShuffleNet V1/V2 proposed by Megvii Technology, etc. At present, the latest research progress is dynamic convolution/conditional convolution.
The second is the depth problem of neural networks. This is a problem that has troubled neural networks for many years. When the depth of the network is not large enough, it is difficult to achieve network training. The initial depth of the neural network was 8 layers. After two years, it increased to 20 layers. Then the depth of the deep residual network proposed by Microsoft increased to 152 layers. The idea of using the residual network can get good training results.
The third is the width of the neural network. When the complexity of deep learning exceeds a point, the larger the model, the error rate of training and testing will decrease at the same time, which is different from our traditional machine learning cognition, which is actually related to the width of the network. There are two relatively new directions: one is to start from the perspective of Kernel, and the other is to try pruning methods, such as MetaPruning.
The fourth is the size of the neural network. Generally speaking, the size of the neural network is constant during the training process. However, research has found that when the size of the neural network is changed during training, better network performance can be achieved.
Regarding the computer vision technology itself, Dr. Sun Jian focused on the direction of target detection and shared some problems and progress in the current research:
First, when the objects in the image are very close, the detection technology cannot accurately detect a single object;
Second, the design problem of computing architecture. In this regard, Megvii proposed a lightweight two-stage target detector-ThunderNet, which designed the integration of multi-scale architecture, and it runs very fast on ARM devices.
Finally, Sun Jian also pointed out some of the most important and most invested key issues in computer vision applications:
- First, it is very difficult to collect data for special scenes such as fire, and it is difficult to obtain data through data enhancement;
- Second, the need for new research methods such as self-supervised methods;
- Third, the problem of occlusion. Although there is some progress on this, deep learning cannot completely solve this problem;
- Fourth, deep learning and computer vision technologies are not yet able to track multiple objects that are in motion at the same time continuously;
- Fifth, the problem of visual control, for example, it is not possible to continuously control the robot or manipulator through visual feedback;
- Sixth, there are still great challenges to realize low cost, easy deployment and security in real applications;
- Seventh, the existing methods cannot achieve high-precision prediction problems.
8. Chen Xilin: Towards an understandable computer vision
Chen Xilin, a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences, ACM, IEEE, and IAPR Fellow, also served as a guest of the report and brought a report with the theme "Towards Comprehensible Computer Vision". In the report, he also shared some of the problems in the field of computer vision from his perspective and some of the exploration work he has done to address these problems, and gave his own thoughts on the future development of computer vision.
Researcher of Institute of Computing Technology, Chinese Academy of Sciences, ACM, IEEE, IAPR Fellow Chen Xilin
He pointed out that it has been almost half a century since the concept of computer vision was put forward, and it has mainly gone through the four stages of Marl's computational vision, active and objective vision, multi-view geometry and hierarchical 3D reconstruction, and learning-based vision. Although the progress made in this field is obvious, it also brings some problems, such as the emergence of evaluation benchmarks.
"Before, everyone did not compare each other in research. Even if you publish a paper, the results may stay in place. Then there is a benchmark for evaluation. However, one disadvantage is that now researchers, especially students, just take care of it. "Sweeping the list", this is actually not doing real research. So this is a big problem. "
He believes that when doing computer vision research, you need to know not only What and Where, but also How, Why, When, etc. In addition to the problems of research methods, the current computer vision research is also facing two serious problems.
The first is that the research is in a "closed world", which reflects that new data cannot be updated in time, knowledge cannot be borrowed from other fields, and the real connection between objects cannot be truly understood;
The second is the inability to handle open world issues well, such as the inability to distinguish between the language and semantics of the real world.
In response to these problems, Chen Xilin has carried out a series of explorations and works on interpretable decision-making models, the similarity between conceptual space, semantic space, and visual space, transferable comparative learning, and the use of context.
Finally, he concluded that in the past 50 years, computer vision has achieved a lot of success in the application, so what will happen in the future? ——In the future, computer vision research will develop in an understandable direction, that is, the knowledge behind the technology will play a more important role.
9. Li Shipeng: Internet of Everything
Li Shipeng, academician of the International Eurasian Academy of Sciences, deputy dean of Shenzhen Institute of Artificial Intelligence and Robotics, and IEEE Fellow, in his speech titled The development process of Intelligent Internet of Things) to IIoT (Intelligent Internet of Things).
Academician of International Eurasian Academy of Sciences, Deputy Dean of Shenzhen Institute of Artificial Intelligence and Robotics, IEEE Fellow Li Shipeng
Li Shipeng believes that apart from other factors, the current era of artificial intelligence mainly includes four basic factors: AI, humans, robots and IoT. Among them, human is the central factor. The interaction between human and intelligence is called human-machine coupling or human-machine collaboration. The combination of human and IoT is physical intelligence, and the combination of human and AI is a virtual scene.
The entire development process of IoT can be divided into three stages:
The first stage is the most basic stage of IoT. All devices that can connect to the Internet and transmit data are called IoT devices, which are mainly concerned with the connection problems between devices, data collection and communication problems. People mainly interact with IoT devices through commands or remote control. The level of intelligence at this stage is very low, and basically it can only do conditional control like IFDtt.
The second stage is called AIoT. This term is not proposed internationally, but a concept with Chinese characteristics. The IoT in the previous stage was basically without intelligence, and the application of data was simple or only on the surface. In this stage, the data generated by the IoT is processed intelligently. On the one hand, the user's interaction with IoT devices is becoming more and more intelligent; on the other hand, the collected data does not only stop at the interpretation of the original data, but combines the data to form some new knowledge. At this stage, AIoT always has a centralized controller to control all IoT devices, because it needs such a brain for overall control.
The third stage is IIoT (Intelligent Internet of Things). In the last stage, the independent intelligent object itself has a certain degree of intelligence, and can operate independently in many cases. At this stage, we need to explore how to combine the intelligence between intelligent and independent agents and what kind of intelligence can be formed by combining them together? The relationship between man and machine has become a more equal cooperative relationship.
Li Shipeng believes that the evolution of aggregate intelligence brought about by IIoT is taking place. This trend may break some of the existing obstacles in the artificial intelligence industry, and finally may pave the way for a future AI framework based on causality.
10. Xia Qin: Cloud is beautiful-the present and future of high-performance computing chips
Xia Qin, Chief Chip Planner of Huawei HiSilicon and Head of Product Management Department of HiSilicon Turing, shared the present and future of high-performance computing chips with the title of "Cloud Great Beauty-The Present and Future of High-Performance Computing Chips" Thinking, and taking Huawei's Shengteng chip and Kunpeng chip as examples, discussed the problems of high-performance chip design.
Xia Qin, Chief Chip Planner of Huawei HiSilicon, Director of Product Management Department of HiSilicon Turing
Xia Qin said that high-performance computing chips have become a very popular direction in the chip field. Many new theories and new technologies have been produced in this field. Now engineers in this field are nothing more than solving a ternary equation, including three dimensions: performance , Cost, ease of use. It is easy to solve from a single dimension, but if you want to achieve real mass production, you must solve the following three problems at the same time:
First, the nuclear design problem can be explored in the three directions of improving the main frequency, multi-core and micro-architecture (instruction set). For example, in the multi-core direction, Huawei has already launched chips ranging from 8 cores to 16 cores to 64 cores; in the micro-architecture, Huawei's Kunpeng 920 chip has implemented some measures such as disorder and instruction prefetching on the instruction set. To improve performance, better results were obtained.
Second, on-chip design issues. In this regard, Huawei's optimization work on the chip includes the use of the latest technology to reduce the chip area, increase the yield, and reduce costs.
Third, the peripheral interface issue. In this regard, Huawei has made a lot of innovations in the memory channel interface of the chip, including interface speed, multi-IP interconnection, and accelerator.
In addition, the design of AI chips also faces unprecedented challenges in software: First, AI-based computing requires greater parallel processing capabilities; second, the very big difference between AI chips and CPU chips is that the former is true. Finally, the construction of AI software stack also faces great challenges. For example, the development process from Huawei Kunpeng software stack to Shengteng software stack is very complicated and difficult.
She further summarized the trend of future chip design:
First, the single-thread and single-core performance will continue to improve in the future, and more definitions will appear for the instruction set. In addition, with the vigorous development of ARM, it will also provide a lot of space for future user customizations, which will be more Dadi improves the overall performance of the chip;
Second, in terms of memory storage, although there is no better technology yet, there is still much room for exploration in the redefinition of the memory interface and the entire memory architecture;
Third, breakthroughs in new standards and interface technologies. There will be new technical attempts in this area, but the results will be slower;
Fourth, flexible power management. With the advent of heterogeneous computing in the future, tuning technologies with low load and low power consumption will become very important.
"AI has many directions that can be developed in the future, and it is not a one-dimensional development direction. Although we still don’t know when the real AI era will come, I think computing power, collaboration, and application are the overall AI technology in the future. Three key dimensions of commercial use.”
Five academicians discuss the future of AI
The second academician roundtable was held in the afternoon of the same day, hosted by Chen Jie , academician of the Chinese Academy of Engineering and president of Tongji University . Chinese Academy of Engineering, radio and television technology expert Ding Wenhua , Chinese Academy of Engineering, Communications and Information Systems Specialist Wang Sha Fei , Chinese Academy of Sciences, laser and photonics technology expert Wang Lijun , a professor of Informatics Science University of Hamburg, Germany, Hamburg Academy of Sciences Zhang Jianwei four Academician, combined with the content of the keynote speech of the day, each expressed his prospects and suggestions for the future development of artificial intelligence in China.
Academician Ding Wenhua: The content of this conference is rich, which shows that China's artificial intelligence has made great progress in various fields. The content of this year's speech involves algorithms, applications, and resources, and everyone has achieved breakthroughs in their respective research directions. I believe that through the platform of Pengcheng Laboratory, high-end experts and talents in the field of artificial intelligence from the whole country and the world can be gathered to communicate and promote the development of artificial intelligence technology.
Academician Wang Shafei: I think the entire development of artificial intelligence still needs to go through a long stage. There are several challenges and problems:
- First, when artificial intelligence technology is applied to application scenarios, there are still many problems in intelligent reasoning. It is difficult for artificial intelligence to intelligently reason about unknown scenes or targets like humans.
- Second, the issue of interpretability. Now AI can calculate massive amounts of big data and achieve certain perception, but is the result correct? Maybe in the future we can improve this problem by adding people's experience.
I am very inspired by the reports of experts today. I think that through the efforts of my colleagues, artificial intelligence can break through the difficulties in basic research and get better applications.
Academician Wang Lijun: I mainly research laser chips. In recent years, with the development of artificial intelligence and information perception, I have also carried out research on communication and information perception (optoelectronic integrated chips). I have studied laser chips for decades and have some personal experience.
The development of domestic chips has been relatively slow in recent years, and they are still at a stage where they are limited by people. Why is there such a situation? There are several reasons:
First, the equipment cost required in the chip development process is very huge, and the average unit cannot afford it;
Second, in terms of time, it takes years or even more than ten years of hard work to make chips.
Third, in recent years, we have been pursuing results as soon as possible. Of course, the intention is good, but there are things that should respect the facts. Some government agencies may not be willing to invest in such large investments and slow results;
Fourth, researchers working on chips, especially young people, are more willing to invest in research that yields results and articles immediately.
Now our country is aware of these problems and has also taken some measures to overcome the chip problem. I believe that in a few years, our country will have a major breakthrough in the chip.
In addition, for information perception, I personally think that the next step is a very important direction in optoelectronics and hybrid integrated chips. It not only further integrates integrated circuit technology and integrated optics technology, but also integrates perception software and optical things into At the same time, the further improvement of reliability will greatly promote the AR industry.
Academician Zhang Jianwei: First of all, I would like to congratulate the Pengcheng Laboratory for such a great achievement in more than a year. After defining several important directions in the future, we will organize researchers from all walks of life to solve major issues of international people’s livelihood from a cross-cutting perspective. , Has become a very important platform for attracting the most high-end talents from industry, academia and research to gather exchanges and brainstorm.
After listening to your report today, I would like to emphasize a few more points:
First, lay the foundation. Today, some experts talked about multi-modal technology. From brain science multi-modal processing, chip multi-modal processing, image recognition, picture understanding, etc., multi-modal information processing has become a core technology of artificial intelligence, and it is also very worthwhile We further develop and research. The cross-modal learning project that I organized five years ago is the largest research project between China and Germany. It organizes and studies the multi-modal learning mechanism of people from brain science, psychology, artificial intelligence, robotics, etc. The algorithm is finally implemented with a robot.
Second, how will artificial intelligence be implemented next. I think that in addition to providing a basic platform for artificial intelligence, the next step is to truly integrate into demand and vertical fields, deepen integration, deepen and grow the processing chain, and create a world-class intellectual property and world-class market, so that the value of artificial intelligence Produce faster.
Third, public platform communication, ecological construction and social impact. Now local governments are paying more and more attention to open source and providing platforms for enterprises. I think that a new world-class artificial intelligence Demo prototype can be made in Shenzhen, a hot spot for entrepreneurship.
No comments:
Post a Comment