Dharma Institute released the top ten scientific and technological trends in 2022: AI for Science gave birth to a new paradigm of scientific research.

According to reports, "Ten Technical Trends of Dharma Institute 2022" adopts the analysis method of "quantitative divergence and qualitative convergence", and the whole analysis process is divided into two parts:

Dharma Institute analyzed 7.7 million public papers and 85,000 patents in 59 fields in recent three years/kloc-0, excavated hot areas and key technological breakthroughs, conducted in-depth interviews with nearly 100 scientists, and put forward ten scientific and technological trends that may come true in 2022, covering artificial intelligence, chips, computing and communication.

Specifically, the top ten technical trends are: artificial intelligence oriented to science, co-evolution of large and small models, silicon optical chips, green energy artificial intelligence, flexible sensing robots, high-precision medical navigation, global privacy computing, satellite-ground computing, cloud network convergence, XR Internet.

Dharma Institute believes that the path of computer science changing scientific research is from downstream to upstream. At first, computers were mainly used to analyze and summarize experimental data. Later, scientific calculation changed the way of scientific experiments. Artificial intelligence combined with high-performance computing began to use computer simulation experiments in areas with high experimental cost and difficulty to verify scientists' assumptions and accelerate the output of scientific research results, such as digital reactors for nuclear energy experiments, which can reduce experimental costs, improve safety and reduce nuclear waste.

In recent years, artificial intelligence has been proved to be able to discover scientific laws, not only in the field of applied science, but also in the field of basic science. For example, DeepMind uses artificial intelligence to help prove or put forward new mathematical theorems and help mathematicians form an intuition about complex mathematics.

Dharma Institute predicts that in the next three years, artificial intelligence technology will be widely used in applied science and become a research tool for some basic sciences.

Hua Xiansheng, head of the Urban Brain Laboratory of Ali Dharma Institute, said in an interview with InfoQ that using AI to help scientific research is mainly based on data and calculation, and AI ability is formed on the basis of data and calculation.

"In essence, AI has little difference between science and industry, and AI is also a tool to promote the development of the field. It's just that this field is a little different, and the threshold is relatively high, because this is what scientists should do, not what ordinary people and ordinary technicians can do. But in essence, because of the data in this field, algorithms can be designed to mine the' mystery' in the data to solve problems in this field. "

For practitioners, AI for Science needs AI experts to understand scientific problems and scientists to understand the principles of AI. "When AI is used in industry, it actually gradually moves from single-point technology to platformization. I think the future AI for Science will gradually become a platform. At this time, AI experts combine certain problems in a certain field, a certain discipline, and even a certain discipline to build a scientific research platform with scientists. At this time, scientists may have greater freedom and more powerful tools to do scientific research in batches and achieve richer and more important scientific breakthroughs. " Hua Xiansheng said.

Large-scale pre-training models such as Google's BERT, Open AI's GPT-3, Zhiyuan's Enlightenment, and Dharma Institute's M6 have made important progress, and the performance of large-scale models has advanced by leaps and bounds, providing a foundation for the development of downstream AI models. However, the large model training consumes too many resources, and the performance improvement brought by the increase of the number of parameters is disproportionate to the consumption improvement, which challenges the efficiency of the large model.

Yang Hongxia, a scientist at the Intelligent Computing Laboratory of Ali Dharma Institute, said in an interview with InfoQ that there are still several topics to be broken through in the pre-training mode:

Dharma Institute believes that the large-scale development of large-scale model parameters will enter a cooling-off period, and the cooperation between large-scale model and related small models will be the future development direction. The knowledge and cognitive reasoning ability precipitated by the large model are exported to the small model. The small model superimposes the perception, cognition, decision-making and execution ability of the vertical scene on the basis of the large model, and then feeds back the results of execution and learning to the large model, so that the knowledge and ability of the large model can evolve continuously and form an intelligent system with organic circulation. The more participants and beneficiaries, the faster the development of the model.

"The co-evolution of large and small models can also better serve more complex new scenes, such as virtual reality and digital people. It needs to be deployed and interacted simultaneously in the cloud. At the same time, the system is more flexible for protecting user data privacy, and users can maintain their own small models at different ends. " Yang Hongxia said to InfoQ.

Tang Jie, a professor of computer science in Tsinghua University and academic vice president of Beizhiyuan Artificial Intelligence Research Institute, said that the development of large-scale models does not rule out the possibility of further increase of model parameters in terms of cognitive intelligence, but parameter competition is not an end in itself, but the possibility of further improvement of performance. At the same time, the research of large-scale model pays attention to the original innovation of architecture, and further improves the cognitive intelligence ability of trillion-level model through methods such as continuous learning of model, increasing memory mechanism and breaking through triple knowledge representation. As far as the model itself is concerned, the new multi-modal, multi-language and programming-oriented model will also become the focus of research.

Dharma Institute predicts that in the next three years, based on the large-scale pre-training model, co-evolutionary intelligent systems will be explored in some fields. In the next five years, coevolutionary intelligent system will become the system standard, which will enable the whole society to easily acquire and contribute the capabilities of intelligent systems and take another big step towards general artificial intelligence.

The development of electronic chips is close to the limit of Moore's Law, and the progress of integration technology is close to saturation. The demand for data throughput in high-performance computing is increasing, and technical breakthroughs are urgently needed.

Photonic chips and electronic chips are different in technology. Using photons instead of electrons to transmit information can carry more information and transmit longer distances. There is less interference between photons, which provides two orders of magnitude higher computational density and two orders of magnitude lower energy consumption than electronic chips. Compared with quantum chips, photonic chips can continue the current computer system without changing the binary architecture. Photonic chip needs to be integrated with mature electronic chip technology, and silicon optical technology, which combines the advantages of photons and electronics, will be the mainstream form in the future.

Professor Peking University and Zhou Zhiping, the distinguished principal researcher of Shanghai Institute of Optics and Mechanics, said that Dharma Institute chose "silicon optical chip" as one of the scientific and technological trends in 2022 10, which confirmed the great application value of this technology in the field of information and communication. The further development of silicon optoelectronic chip is silicon-based optoelectronic chip: using the design method and manufacturing process of integrated circuit, micro-and nano-sized photonic, electronic and optoelectronic devices are heterogeneously integrated on the same silicon substrate, forming a complete and comprehensive new large-scale optoelectronic integrated chip. It more obviously reflects the continuous efforts of human society in nanotechnology and great interest in smaller equipment and more compact systems.

Dharma Institute predicts that photoelectric fusion is the development trend of chips in the future. Silicon photonics and silicon electronic chips complement each other, give full play to their respective advantages, and promote the continuous improvement of computing power. In the next three years, silicon optical chips will support high-speed information transmission in large data centers; In the next five to ten years, optical computing based on silicon optical chips will gradually replace some computing scenarios of electronic chips.

The large-scale development and utilization of green energy has become the main direction of energy development in the world today. With the trend of high proportion of green energy connected to the grid, it is difficult for the traditional power system to cope with the uncertainty of green energy in strong wind, rainstorm, lightning and other weather, and the ability to deal with complex faults in time.

In the process of operation monitoring, parameter verification and fault monitoring still need a lot of manual participation, so it is difficult to extract and identify fault features. Aiming at the challenges faced by large-scale green energy grid connection in terms of stability, operation and planning, a new generation of information technology based on artificial intelligence will provide technical guarantee and strong support for the efficient and stable operation of the energy system as a whole.

The deep integration of artificial intelligence and energy and electricity will promote large-scale new energy power generation, grid connection, transmission, consumption and safe operation, and complete the upgrading of energy system.

On Tuesday, the chief system architect of China Electric Power Research Institute believed that the intelligent regulation and operation deduction of the new power system will be inseparable from AI technology. With the support of AI technology, a number of digital twins with interactive physical power grid and IT applications will be built, and each digital twin can solve a certain scene or a certain aspect of power grid operation problems. In this way, when there are enough twins to form a digital twin system for power grid regulation and control, and solve various problems in power grid operation, intelligent regulation and control can be realized.

Dharma Institute predicts that in the next three years, artificial intelligence technology will help the power system achieve large-scale green energy consumption, energy supply can be interconnected and mutually beneficial in time and space dimensions, network sources will develop harmoniously and flexibly, and the power system will operate safely, efficiently and stably.

Robots are masters of technology. In the past, with the integration of hardware, network, artificial intelligence and cloud computing, the maturity of technology has made great progress, and robots are developing towards multi-task, adaptive and cooperative routes.

Flexible robot is an important breakthrough representative, which has the characteristics of softness, flexibility, programmability and expansibility. Combined with flexible electronics, force sensing and control technology, it can adapt to various working environments and make adjustments in different tasks. In recent years, the flexible robot combined with artificial intelligence technology has made the robot have the perception ability, improved universality and autonomy, and reduced the dependence on pre-programming.

Flexible sensing robot increases its ability to sense the environment (including force, vision, sound, etc.). ), it enhances its ability to migrate tasks, and it is no longer necessary to exhaust all kinds of possibilities like traditional robots, and it can perform tasks that rely on perception (such as medical surgery), expanding the applicable scenarios of robots. Another advantage is the adaptability in the task, which can respond to sudden changes in time, complete the task accurately and avoid problems.

Dharma Institute predicts that in the next five years, flexible robots will fully combine the intelligent perception ability brought by deep learning, face a wide range of scenarios, and gradually replace traditional industrial robots and become the main equipment on the production line. At the same time, it has been commercialized in the field of service robots, which has advantages in scene, experience and cost, and has begun to be applied on a large scale.

Traditional medical treatment relies on the experience of doctors, just like manual road finding, and the results are uneven. The deep integration of artificial intelligence and precision medicine, and the organic combination of expert experience and new auxiliary diagnosis technology will become a high-precision navigation system for clinical medicine, providing automatic guidance for doctors, helping medical decisions to be faster and more accurate, and realizing quantification, computability, predictability and prevention of major diseases.

It is expected that in the next three years, people-centered precision medical care will become the main direction, and artificial intelligence will fully penetrate all aspects of disease prevention and diagnosis and treatment, and become a high-precision navigation collaboration for disease prevention and diagnosis and treatment. With the further development of causal reasoning, interpretability is expected to achieve a breakthrough, and artificial intelligence will provide strong technical support for disease prevention and early diagnosis and treatment.

Data security protection and data circulation are two dilemmas in the digital age, and the solution is privacy computing. In the past, due to performance bottlenecks, insufficient technical trust, and inconsistent standards, privacy computing can only be applied to scenes with small data. With the integration and development of special chip, encryption algorithm, white box and data trust technology, privacy computing is expected to leapfrog to mass data protection, and data sources will expand to the whole world, stimulating new productivity in the digital age.

Ren Kui, a professor at Zhejiang University and dean of the School of Cyberspace Security of Zhejiang University, said that privacy computing is not a single technology, but a unified name, including secure multi-party computing first proposed by 1982, and later homomorphic encryption, trusted computing, and differential privacy. However, privacy computing did not have much practical value at an earlier time, such as homomorphic encryption, which is good in theory, but it is too expensive to use in practice. Now with the acceleration of hardware and the innovation of software, the practical trend is gradually seen, and of course there is still a process.

Dharma Institute predicts that in the next three years, global privacy computing technology will make new breakthroughs in performance and interpretability, or data trust institutions will provide data sharing services based on privacy computing.

Digital services based on ground network and computing are limited to densely populated areas, and no man's land such as deep space, ocean and desert is still a blank area for services. High-low orbit satellite communication and ground mobile communication will be seamlessly connected to form an integrated three-dimensional network of air, land and sea. Because computing moves with the network, satellite-to-earth computing will integrate satellite system, air network, ground communication and cloud computing, become a new computing architecture, and expand the space of digital services.

Zhang Ming, head of XG Laboratory of Ali Dharma Institute, believes that the successful commercialization and large-scale development of satellite-earth computing still involves many breakthroughs in core technologies.

Taking LEO satellite terminal as an example, firstly, we should be guided by the scene demand and commercial value, and secondly, we should design high-performance, low-cost and multi-scene commercial products from the perspective of technological breakthrough and solving engineering problems. For example, in key technologies, how to design a new millimeter-wave phased array antenna and the corresponding beamforming control algorithm in a low-cost way to meet the performance requirements; How to design a new satellite-ground communication protocol to meet the needs of multi-users, mobility and complex dynamic services of satellite Internet; In addition, in terms of terminal integration and optimization, there are still many engineering problems that need to be broken through and solved to meet various needs in different scenarios of land, sea and air.

Dharma Institute predicts that in the next three years, the number of low-orbit satellites will usher in explosive growth, and together with high-orbit satellites, it will form the satellite Internet. In the next five years, satellite Internet and terrestrial network will be seamlessly integrated to form ubiquitous Internet, and satellites and their terrestrial systems will become new computing nodes and play a role in various digital scenes.

The development of new network technology will promote the development of cloud computing to a new computing system with cloud network integration, and realize the professional division of labor at the cloud network: the cloud will be the brain, responsible for centralized computing and global data processing; As a connection, the network integrates various network forms through the cloud to form a network with low delay and wide coverage; As an interactive interface, the terminal presents various forms, which can provide a thin, lasting and immersive experience. Cloud-network convergence will promote the birth of new applications such as high-precision industrial simulation, real-time industrial quality inspection and virtual-real integration space.

Dharma Institute predicts that in the next two years, there will be a large number of application scenarios running in the cloud-network convergence system, accompanied by more new devices born according to the cloud, bringing more extreme and richer user experience.

With the development of technologies such as end-cloud collaborative computing, network communication and digital twinning, the XR (Future Virtual and Real Integration) Internet with immersive experience as the core will usher in an explosive period. Glasses are expected to become a new human-computer interaction interface, which will promote the formation of XR Internet, which is different from the flat Internet, and give birth to a brand-new industrial ecology from components, equipment, operating systems to applications. XR Internet will reshape the form of digital application and change the interactive way of entertainment, social interaction, work, shopping, education, medical care and other scenes.

Dharma Institute predicts that a new generation of XR glasses will be produced in the next three years, integrating AR and VR technologies, and using technologies such as end-cloud collaborative computing, optics and perspective to make the shape and weight close to ordinary glasses. XR glasses will become the key entrance of the Internet and have been widely popularized.