In the field of artificial intelligence, the basic capabilities of linguistic macromodeling, graphic macromodeling and even multimodal macromodeling of artificial intelligence have been fully demonstrated. For example, Alibaba Dharma Academy announced the latest progress of the multimodal big model M6, with parameters leaping from trillions to 10 trillion; Pengcheng Laboratory and Baidu jointly released the world's first knowledge-enhanced 100 billion big model - Pengcheng - Baidu - Wenshen, with a parameter scale of 260 billion.
Not only that, the cross-fertilization of AI with other scientific fields also rubbed sparks. On Science's recently announced list of scientific breakthroughs for 2021, AlphaFold and RoseTTA-fold, two AI-based techniques for predicting protein structure, topped the list.
In the field of human-computer interaction, while Zuckerberg renamed Facebook "Meta," Tesla and SpaceX CEO Elon Musk is focusing on brain-computer interfaces. Musk believes that brain-computer interface devices will be more likely to change the world, to help quadriplegic or physically challenged people to live and work better, "complex brain-computer interface devices can allow you to fully immerse yourself in virtual reality". In addition, in May, Stanford University developed an intracortical brain-computer interface system that can decode a paralyzed person's imagined handwriting movements from neural activity in the motor cortex and convert them into text.
In the field of supercomputing, the most noteworthy is that in November this year, China's supercomputing application team won the Gordon Bell Prize, the highest award in the field of international high-performance computing, for its "real-time simulation of ultra-large-scale quantum random circuits".
In terms of open source, the RISC-V open source instruction set and its ecosystem are rapidly emerging; the openEuler operating system open source community led by Huawei, with the participation of the Institute of Software of the Chinese Academy of Sciences, Kirin Software and others, has brought together 7,000 active developers, completed more than 8,000 self-maintained open source packages, and spawned more than 10 vendors' commercial distributions. ......
Looking back to 2021, Information Technology Edition invites industry experts to summarize the development of the above four areas and look forward to future trends.
Author Zhang Shuanghu
AlphaFold may be the "No. 1" in artificial intelligence (AI) in 2021.
Recently, Science magazine published its list of scientific breakthroughs for 2021, and AlphaFold and RoseTTA-fold, two AI-based techniques for predicting protein structure, topped the list.
A few days earlier, AlphaGo and AlphaFold were also on the list of the "2021 Top 10 Global Engineering Achievements (major achievements in engineering science and technology that have been validated by global practice in the past five years and have global impact)" selected by the journal of the Chinese Academy of Engineering.
In an interview with China Science Daily, several experts looked back at this year's achievements in the field of artificial intelligence and talked about AlphaFold.
"AlphaFold, which is oriented toward scientific discovery, and the AI development ecology that is being constructed in China can't be left out of the conversation." Wu Fei, director of the Institute of Artificial Intelligence at Zhejiang University, told China Science Daily.
Wang Jinqiao, a researcher at the State Key Laboratory of Pattern Recognition at the Institute of Automation of the Chinese Academy of Sciences, nominated "new crown diagnosis with AI," "the integration of AI with biology, pharmaceuticals, materials and other sciences (AI for Science)" and "Tri-modal large model Zidong Taichu".
In the medical field, AI recognition of coughing has long been used for pneumonia, asthma, Alzheimer's disease and other disease detection. Researchers at the Massachusetts Institute of Technology (MIT) have developed an AI model that can identify new crown patients by analyzing cough recordings, and the accuracy of identifying new crown patients coughing is 98.5%, of which identifying asymptomatic infected people is as high as 100%. A few days ago, it was reported that the model has been used to identify the Omicron virus.
"For the first time, Zidong Taichu has realized the unified expression of graphic-text-sound semantics, with both cross-modal comprehension and generation capabilities." Wang Jinqiao said, "The 'All-Media Multimodal Large Model R&D Program' currently released with Xinhua News Agency*** to achieve unified modeling of all-media data comprehension and generation, and to build a full-stack localized media AI platform has been exploratively applied to scenarios such as textile and automotive industry quality inspection. "
On December 7, the official website of the Ministry of Science and Technology announced three letters to support Harbin, Shenyang, Zhengzhou, three places to build a new generation of national pilot areas for the innovative development of artificial intelligence. So far, China has 18 national new generation of artificial intelligence innovation and development pilot area, which will lead to drive China's artificial intelligence innovation and development.
"China is promoting the development of AI ecology and building a good ecology." Wu Fei said, "There are 15 national new-generation AI development and innovation platforms, 18 national new-generation AI innovation and development pilot zones, eight AI innovation and application pilot zones, and talent cultivation carriers such as undergraduate majors in AI and cross-disciplines set up by higher education institutions."
"One is the big model, and the other is the combination of AI and basic disciplines." Sun Maosong told China Science News, "The basic capabilities of linguistic big models, graphic big models and even multimodal big models have been fully demonstrated, establishing its status as a basic soft facility for intelligent information processing. At the same time, it is not simply scaled up, but challenges both the ability to integrate digital resources and computing power. Although its limitations are also obvious, some of the 'peculiar' properties it exhibits (e.g., sample less learning, deep double descent, cue-based task tuning, etc.) have led scholars to the expectation that the mega-parameter scaling may trigger qualitative changes, thus laying the groundwork for new breakthroughs."
This year, the field of artificial intelligence from the "big refining model" to "refining big model" stage, from the scale of hundreds of billions to trillions, in the field of big model, it seems that there is no biggest, only bigger.
In March, Beijing Zhiyuan Artificial Intelligence Research Institute released China's first super-large-scale artificial intelligence model "Wudao 1.0". In June, Zhiyuan rewrote its own record, releasing Wudao 2.0, with a parameter scale of 1.75 trillion; in September, Wave Artificial Intelligence Research Institute launched a Chinese giant language model -- Source 1.0, with a parameter scale of 1.75 trillion; and Wave Artificial Intelligence Research Institute launched a Chinese giant language model -- Source 1.0, with a parameter scale of 1.75 trillion. -Source 1.0, with a parameter count of 245.7 billion; in November, Alibaba's Darmo Academy announced the latest progress of the multimodal macromodel M6, with parameters jumping from trillions to 10 trillion; in December, Pengcheng Labs and Baidu jointly released the world's first knowledge-enhanced 100-billion macromodel - the -Pengcheng-Baidu-Wenxin, with a parameter scale of 260 billion.
Correspondingly, recently Racer and ETH Zurich proposed Persia, a new recommender system that supports model training with up to 100 trillion parameters.
On the other hand, AI continues to take over in basic disciplines.
July, DeepMind's artificial intelligence program Alphafold2 research results and topped the "Nature", in the field of structural biology research, artificial intelligence or lead biology, medicine and pharmacy into a new world; in November, the United States, researchers at the University of Southern California, through the brain-computer connection device, let the monkeys play the game and the treadmill, so as to carry out neural activity data research; and in December, a machine learning framework developed by DeepMind has helped uncover two new conjectures in pure mathematics, demonstrating the potential of machine learning to support math research.
"This year, AI has also made considerable achievements in applications across industries." Sun Maosong said, "The combination of AI with basic disciplines has shown great potential, with several top papers published, and has shown some kind of strong trend that 'AI + basic science' holds great promise."
Author Zhang Shuanghu
Brain-computer interfaces, AR glasses, intelligent voice, myoelectric bracelets, spatial gesture recognition ......2021 In the year of 21, the field of human-computer interaction is in the air, from basic research to the application of landing. Whether it is intelligent health, meta-universe, or the booming development of the field of autonomous driving, it seems to indicate that human-computer interaction is standing at the door of industrialization landing.
"The high-throughput ultra-flexible neuroelectrode we developed has passed the scientific research clinical ethical approval, and is about to carry out human clinical trials of brain-computer interface." Tao Hu, deputy director of the Shanghai Institute of Microsystems of the Chinese Academy of Sciences and deputy director of the Joint State Key Laboratory of Sensing Technology, told China Science Bulletin, "Safe and stable large-scale acquisition of neuronal signals from the human brain and closed-loop regulation will realize the repair of patients' perceptual and motor functions."
Brain-computer interface technology is bringing more and more convenience to patients. In May, Stanford researchers published a cover paper in Nature developing an intracortical brain-computer interface system that can decode the imagined handwriting movements of a paralyzed patient from neural activity in the motor cortex and convert them into text. With the system, subjects (paralyzed by spinal cord loss) could type nearly 100 characters per minute with an auto-corrected offline accuracy rate of more than 99 percent.
Not long ago, Musk said he hoped to use Neuralink's microchip device on humans next year. The chip would be used to treat spinal cord injuries, Parkinson's disease and other brain and neurological disorders. The technology is currently awaiting approval from the U.S. Food and Drug Administration.
"The field of brain-computer interfaces has accumulated considerable technology and is expected to be a powerful tool for solving brain diseases." Tao Hu said, "Everyone is seizing the first opportunity for clinical application, and the technology may be realized for on-the-ground application next year. It is expected that within two or three years, unicorn enterprises comparable to Musk Neuralink will appear in China."
"Human-computer interaction will lead to a new trillion-dollar market." This judgment by Yan Qun, a special professor at Fuzhou University, also encapsulates the huge market of the meta-universe.
Some people call 2021 "the year of meta-universe", while others think it's just "new wine in old bottles". But no matter what, meta-universe is already an indispensable topic in the field of human-computer interaction this year.
"Meta-universe is a synthesis of virtual reality, augmented reality and mixed reality, and it is actually not something new." Liu Wei, director of the Human-Computer Interaction and Cognitive Engineering Laboratory at Beijing University of Posts and Telecommunications, told China Science Bulletin, "The meta-universe is the direction of development for the real world and the virtual world across the future, but there are still some technical problems that have not been well solved."
In the real world, the problems of human-computer interaction and the mixing of human-computer environment systems have not been well solved. Objective data, subjective information and knowledge are still not perfectly blended in real-world human-computer interaction, whether in the input, processing or output process.
Liu Wei believes that, whether in the real world or the virtual world, both human and machine decision-making have "fast decision-making" and "slow decision-making" processes. Human decision-making sometimes rely on logical decision-making more, sometimes intuitive decision-making more, this "mixed decision-making" constantly changing, and it is difficult to find the pattern of change. This aspect of the problem of machine decision-making has not yet been resolved.
"The meta-universe is still in the early stages of the pie." said Liu Wei. Liu Wei said, "because its underlying mechanism has not been solved -- people in the real world failed to perfectly solve the problem of human-computer interaction, brought to the meta-universe can not be solved."
When it comes to human-computer interaction, Liu Wei thinks the second issue that can't be left out of the conversation is the "complexity domain".
"This year's Nobel Prize in Physics also went to the author of a model for predicting climate change in complex systems." Liu Wei said, "Human-computer interaction is also a complex system, which includes both repetitive problems and messy, cross-domain synergies."
According to Liu Wei, from the perspective of intelligence, a complex system includes three important components: the person, the equipment (artifacts), and the environment. This is actually a number of things interact with each other, intertwined, both entangled and overlapping "man-machine ring system" problem.
"In human-computer interaction, machines are strong in dealing with 'complex' issues, and people are good at managing 'miscellaneous' things - cross-domain collaboration, things cross-domain synergy, balance between things, etc. Because people have not yet found the simple operation law of complex things, so to solve all the problems of intelligent products and intelligent systems, we have to find their combination, integration and interaction points in the system of human, machine and environment. And, people have to be in a dominant position in this system."
The third phenomenon in the field of human-computer interaction that has attracted Liu Wei's attention is that "artificial intelligence has helped mathematicians discover some laws." "Recently, DeepMind developed a machine learning framework that can help mathematicians discover new conjectures and theorems." Liu Wei said, "AI is a basic mathematical tool, and at the same time, mathematics reflects some basic laws. If AI can help mathematicians deal with some mathematical problems, then people will have a better understanding of the simple laws of complex systems, and new breakthroughs in human-computer interaction will be possible."
Author Zhang Yunquan (researcher at the Institute of Computing Technology, Chinese Academy of Sciences)
This year has been a bumper year for supercomputing applications in China.
At the Global Supercomputing Conference (SC21) held in the United States in mid-November, China's supercomputing application team won the highest international academic award in the field of high-performance computing applications with its groundbreaking simulation of quantum circuits based on a new Shenwei system ("real-time simulation of ultra-large-scale quantum random circuits") -- "Gordon Bell".
The system was awarded the Gordon Bell Prize for its pioneering simulation of quantum circuits.
At the same time, Tsinghua University's supercomputing team once again won the SC 21 Student Supercomputing Competition, making it the fourth consecutive champion of the SC competition. These achievements in scalability and performance tuning of large-scale application software show that China's development in parallel software is on the rise.
Back to the drive of supercomputing to the industry, we would like to reintroduce the term "arithmetic economy". As early as 2018, we put forward the concept of "arithmetic economy", that supercomputing as the core of the arithmetic economy will become a representative indicator of the degree of development of the digital economy of a place and the main means of transformation of old and new kinetic energy.
Considering the development trend in recent years, we believe that the current development trend of high-performance computing has fully demonstrated that, with the integration of supercomputing and cloud computing, big data, AI innovation, arithmetic has become the key to the development of the entire digital information society, and arithmetic economy has already ascended to the stage of history.
Through a comprehensive analysis of the current situation of China's high-performance computer development in 2021, it can be summarized that the current high-performance computing is showing the following characteristics.
First, HPC and cloud computing have been y integrated. High-performance computing is usually based on MPI, high-efficiency communication, heterogeneous computing and other technologies, favoring exclusive operation, while cloud computing has elastic deployment capabilities and fault-tolerant capabilities, support for virtualization, unified scheduling of resources and elastic system configuration.
With the development of technology, supercomputing and container cloud are converging and innovating, high-performance cloud has become a new product and service, AWS, AliCloud, Tencent, Baidu, and the representative of the commercial supercomputing, "Beilong Super Cloud", have launched high-performance cloud services and products based on supercomputing and cloud computing technology.
Secondly, the application of supercomputing from the past high precision to a broader and wider direction. With the development of supercomputers, especially the continuous decline in the cost of use, its application areas from the national strategic significance of precision research and development, information security, oil exploration, aerospace and "high cold" scientific computing to a wider range of the main battlefield of the national economy to expand rapidly, such as pharmaceuticals, gene sequencing, animation rendering, digital Movies, data mining, financial analysis and Internet services, etc., can be said to have penetrated into all walks of life in the national economy.
From the HPC TOP100 list in recent years, supercomputing systems used to focus on scientific computing, government, energy, power, meteorology, etc. In the past five years, supercomputing systems deployed by Internet companies have accounted for a significant proportion of the total, with the main applications being cloud computing, machine learning, artificial intelligence, big data analysis, and short videos. The sharp rise in demand for computing in these areas suggests that supercomputing is converging with Internet technology.
From the Linpack performance share of the HPC TOP100 list, arithmetic service occupies the first place with 46%; supercomputing centers account for 24%, ranking the second; artificial intelligence, cloud computing and short video follow with 9%, 5% and 4% respectively.
It can be seen that the continued increase in the share of artificial intelligence has a lot to do with the rapid rise of algorithms and applications such as machine learning and the widespread use of deep learning algorithms in big data. Internet companies have rediscovered the value of supercomputers, especially GPU-accelerated heterogeneous supercomputers, through deep learning algorithms and have invested heavily in building new systems.
In summary, the current arithmetic services, supercomputing centers, artificial intelligence, scientific computing and other fields are the main users of high-performance computing, the Internet, big data, especially in the field of AI strong growth.
Once again, the national level has developed a strategic arithmetic layout plan. In May of this year, the National Development and Reform Commission and other four departments jointly issued the "National Integrated Big Data Center Collaborative Innovation System Arithmetic Hub Implementation Plan", proposing the construction of national hub nodes of the national arithmetic network in Beijing-Tianjin-Hebei, the Yangtze River Delta, the Guangdong-Hong Kong-Macao Greater Bay Area, Chengdu-Chongqing, and Guizhou, Inner Mongolia, Gansu, Ningxia, and launching the implementation of the "East Counts West" project, which seeks to bring the data in the eastern part of the country to the west. The project aims to send data from the east to the west for storage and computation, and at the same time establish computing nodes in the west to improve the unbalanced layout of digital infrastructure, effectively optimize the layout structure of data centers, upgrade computing power, and build a national computing power network system.
Finally, the arithmetic demand of artificial intelligence has become the main driving force of arithmetic development. Machine learning, deep learning and other algorithmic innovations and through the Internet of Things, sensors, smart phones, smart devices, Internet technology to collect big data, as well as supercomputers, cloud computing and other components of the super computing power, is recognized as the artificial intelligence era of the "Troika", *** with the latest round of artificial intelligence revolution.
In the context of the booming development of artificial intelligence, virtualized cloud computing to high-performance container cloud computing evolution, big data and parallel computing, machine learning integration and innovation has become the latest direction of industrial development.
In addition, in terms of intelligent computing evaluation, China has put forward a number of benchmarking programs, including AIPerf 500, which is a powerful complement to the traditional Linpack test standards.
These developments indicate that supercomputing technology is penetrating the industry at an accelerated rate, and we have entered an era of AI that relies on arithmetic power, which is one of the inevitable trends for future development. As users' demand for computing power continues to grow, the computing power economy will certainly occupy an important position in the future development of society.
Author Wu Yanjun (researcher at the Institute of Software Research, Chinese Academy of Sciences)
The remarkable development of open source is not just this year. A lot of important things have happened in open source in recent years.
For example, the rapid rise of the RISC-V open source instruction set and its ecology. This is the same as the birth of Linux in the early 1990s. At the time, UNIX and Windows were the norm, and few could have predicted the Linux-based operating systems that are now part of every aspect of people's lives.
Today, more than 80% of the apps that people use every day run on Android with Linux as the kernel, and there is a high probability that the operating system running on the back-end servers that support their business is also a Linux distribution.
So it's just as likely that RISC-V today is undervalued as immature and difficult to compete with ARM and X86. But perhaps in the future RISC-V, like Linux, will eventually become the dominant instruction set ecosystem worldwide, with products in every aspect.
In 2020 alone, RISC-V International (RVI, the new name of the RISC-V Foundation after it moved to Switzerland) saw its membership grow by 133 percent. In fact, RVI's move to Switzerland is significant in its own right, and is a classic example of the open source world not "choosing sides" in the face of competition from major countries, and is worthy of reference for other open source foundations around the world.
In China, at the end of 2019, Huawei led the openEuler operating system open source community with the participation of the Institute of Software of the Chinese Academy of Sciences and Kirin Software. In just two years, the community has brought together 7,000 active developers, completed more than 8,000 self-maintained open source software packages, and spawned commercial distributions from more than 10 vendors.
This is the first real "root community" in the field of basic software in China, although there is still a gap with Debian and Fedora, which have a history of more than 20 years, but it is an important step forward, and there is finally a domestically-led new platform that can be accumulated for a long period of time in terms of academic research, technological research and development, and industrial innovation. The platform is a new one for academic research, technology development and industrial innovation.
Meanwhile, Huawei launched HarmonyOS, an operating system for Hongmeng, after encountering the overseas cut-off of the Android operating system GMS (Google Mobile Services), and launched OpenHarmony, an open source project under the Open Atom Open Source Foundation.
Currently OpenHarmony has attracted the participation of a large number of domestic vendors in a short period of time, and also laterally
OpenHarmony is an open source project that has attracted the participation of many domestic vendors in a short period of time, which also reflects the strong demand for a new generation of operating systems for the Internet of Everything from the domestic industry. Although there is still a gap between OpenHarmony and Android in terms of ecological scale and technological integrity, OpenHarmony has, after all, taken the first step towards building its own ecosystem.
This is equivalent to a boundary for the fair use of the source code, that is, the fair use is limited to the interface, once deep into the implementation of the interface code, you need to comply with the relevant license. This is an important reference for the legal definition of open source intellectual property.
In May of this year, the 2021 China Open Source Development Blue Book was released. It not only systematically sorted out the current situation of China's open source talents, projects, communities, organizations, education, and business, and gave development suggestions, but also provided references for the relevant management departments of the national government to formulate open source policies and layout open source strategies, and provided more case references and data support for scientific research institutes, technology enterprises, and open source practitioners.
Whether it is the development of open source software to the open source hardware and software ecosystem around the open instruction set, or open source has a strict legal boundary constraints, or the domestic leading enterprises are trying to open source to explore the solution to the problem of "neck", and has achieved certain results ... ... ...Many cases point to one direction - the trend of open source is unstoppable. Because it comes from the nature of human beings to share knowledge and create collaboratively, it is also an important mode for human civilization to be passed on in the digital age.
Of course, it is undeniable that there are still many problems with open source, such as the security of the open source software supply chain. Here the security of both the traditional sense of software quality, security vulnerabilities, but also open source software can not be sustained effective maintenance of the problem (such as OpenSSL in the emergence of HeartBleed problem only two part-time maintainers, log4j problems only three part-time maintainers), but also the competition between the big countries led to the "cut-off" problem (such as GitHub), the problem of the "supply" problem (such as GitHub). There are also problems with "cut-offs" due to competition from larger countries (e.g., GitHub had restricted access to Iranian developers).
With the concentration of open source software on commercial platforms like GitHub, this problem will become more pronounced, and even a major risk. Open source software, which is supposed to be the intellectual property of all mankind, could be turned into a weapon of "long arm". In order to avoid this problem, public*** infrastructure such as open source code hosting platforms and open source software build and distribution platforms need to be "decentralized". The world needs multiple open source software infrastructures to minimize the threat to the open source community from political forces.
For China, as open source software becomes an important part of the support for many major infrastructures such as scientific research and industry, it is important to have an infrastructure for open source software itself, which has the functions of code hosting, compiling, building, testing, releasing, and operation and maintenance, to ensure the security and continuity of the supply of open source software, and to enhance the confidence of various industries in using open source software.
In the future, core technology innovation and open source contribution to lead the development of domestic enterprises will become the new driving force, or open source business in China to another climax.