On August 1, 1958, the "103 machine" was successfully commissioned, marking the birth of the first modern electronic computer in mainland China. China has since "owned" computers.
Today, generative AI has become a new generation of productivity tools. Looking back on the history of nearly 70 years, it is the research and development years of a large number of scientific and technological personnel in the field of computing, and it is also a long road intertwined with ideals and struggles. In this process, countless scientific research figures and their deeds have become a source of motivation for future generations to continue to explore and innovate.
With the change of stars, many changes have taken place in the "computing" we face and understand, and intelligent computing has a more dimensional technical subset, richer connotation, and broader applications...... Most of the "Intelligent Computing" we discuss aims at the world's scientific and technological frontiers and major national strategic needs, studies multi-level computing problems such as devices and chips, advanced computers, software and systems, platforms and applications in the field of computing, and provides advanced computing chips, powerful computing capabilities, and efficient and intelligent computing platforms for the scientific and technological innovation system and industrial development system.
As an important foundation and technology pillar of various industries, and one of the most actively used areas in the industry today, intelligent computing is reshaping our world in a unique way. From electronics and biological sciences to materials science, from computing power development to space exploration, the application of intelligent computing is ubiquitous. As AI technology matures, intelligent computing has become an important force for scientific discovery, technological innovation, and business change.
MIT Technology Review China and DeepTech jointly launched the "China Intelligent Computing Innovators" to recognize researchers, engineers, and industry practitioners who have made outstanding contributions to the field of intelligent computing, not only making breakthroughs in academic research, but also demonstrating outstanding role models in technology promotion and commercial applications.
After a year-long solicitation, nomination and review process, in April 2024, MIT Technology Review China × DeepTech officially released the list of "2023 China Intelligent Computing Innovators", which is the technology and talent benchmark of the intelligent computing industry, representing our dedication to cutting-edge technology, the importance of intelligent computing ecology, and the value of technology for good.
Figure丨MIT Technology Review China × DeepTech released the "2023 China Intelligent Computing Innovators" (source: DeepTech)
The field of computing itself is highly interdisciplinary, and combined with the fields of the two inductees, we found that the research directions of the selected candidates cover multiple dimensions from basic scientific research to applied technology, including AI-enabled scientific discovery (AI4S), machine learning, advanced/new computing paradigms (quantum computing, photonic computing, biologically inspired computing, etc.), security and trustworthiness, terminal intelligent applications, architecture, software and hardware development, etc. These studies are not only academically forward-looking, but also gradually show their application potential and value in practical applications.
The inductees in the AI-Enabled Scientific Discovery (AI4S) field focus on applying AI to scientific discovery, including but not limited to biology, materials science, physics, and more. Their work often involves developing new algorithms or improving existing ones to discover new content as well as processing and analyzing large and complex data sets, with interdisciplinary collaboration and algorithmic innovation characterizing innovative work in this field. For example, AI models can be used to predict protein structures and accelerate drug discovery, large amounts of biomedical data can be used to train models to identify disease markers or predict disease risk, and machine learning can be used to optimize the synthesis process of new materials to improve material performance.
The enduring topics in the field of intelligent computing include artificial intelligence, new computing infrastructure, and important issues of security and trustworthiness. In particular, the emergence of large models in 2023 has also brought higher requirements for computing power, and has also stimulated the emergence of more new computing paradigms and more innovative entrepreneurs.
Among them, there are still problems of data security and privacy security, and "protectors" can still be seen, who are committed to improving the security and reliability of intelligent computing systems, developing intelligent systems that can resist cyber attacks, researching new encryption technologies to protect communication security, and designing algorithms to protect the privacy of user data without sacrificing computing efficiency, etc.
Among the selected candidates are entrepreneurs and entrepreneurs, who are engaged in the application forms and capabilities of intelligent computing, including the application of mobile phones, computers, vehicles, etc., the integration of man-machine-things, and the artificial intelligence technology and embodied intelligence of higher-dimensional, more autonomous, and multimodal systems in the future.
The R&D and application of these new technologies are inseparable from the developers of the underlying infrastructure, software, and hardware platforms, who develop hardware and software architectures that can support efficient and intelligent computing, develop new intelligent algorithms, improve the processing speed and accuracy of computing tasks, and develop domestic independent and controllable hardware systems, etc., so that the underlying foundation of computing is more solid, and the applications of the upper layer can flourish and grow.
The selection of the 2023 China Intelligent Computing Innovators is not only a review of the achievements in the field of intelligent computing in the past year, but also a look forward to the future. We look forward to seeing more inspiration and technologies emerge in the computing field, and we look forward to seeing the practical application of new technologies and the wider business potential and economic value.
The list of selected candidates is as follows:
*The following rankings are listed in no particular order, in alphabetical order only
Machine learning was used to extract the information from the night hyperspectral imaging and reconstruct the daytime scene, which successfully overcame the "ghost" effect, and then proposed and experimentally realized the "hyperspectral thermal radar" for the first time.
Application scenarios: unmanned driving, intelligent medical care, intelligent agriculture, national defense, etc
On the one hand, quantum information algorithms are used to break through the optical diffraction limit and improve imaging accuracy, which has nearly 10 times the accuracy of diffraction-limited optical imaging (such as astronomical imaging, ultra-long-distance machine vision, etc.). On the other hand, Bao Fanglin led the team to use machine learning, combined with infrared physics, information theory, etc., to overcome the "ghost" effect in thermal imaging, and developed a subversive night vision technology - TeX vision, and developed TeX vision-based detection and ranging, that is, thermal radar. Thermal radar can allow machines to see the same clear scene as in the day at night, greatly improving machine vision under low visibility, and has great potential to obtain as wide social applications as microwave radar, lidar and other technologies, and equip robots and unmanned cars with a pair of eyes that can see through the darkness.
He is engaged in the basic theoretical research of machine learning, develops deep learning theory, distributed machine learning algorithms, and trusted machine learning methods, and provides a series of technical solutions for the realization of accurate, efficient and safe machine learning by using the mathematical mechanism of machine learning.
Application scenarios: intelligent algorithm security, etc
Chen Wei has been engaged in the research of mathematical mechanism and optimization technology of machine learning for a long time, and has made a series of achievements in deep learning mechanism, distributed optimization, security and trustworthiness applications, etc., and is committed to overcoming the basic theory of intelligent algorithm security and solving the security and trustworthiness problems of intelligent computing. Differential privacy protection for deep learning, the projection gradient perturbation method and theory have been transformed into Microsoft cloud service Azure, and the international market share of the product exceeds 20%. Aiming at the optimization techniques commonly used in adaptive pre-regulator and momentum method, the hyperparameter condition of asymptotic convergence to the maximum interval solution is proved, and the implicit regular theory of deep learning is developed. For distributed machine learning training, the proposed "delay-compensation" mechanism has been integrated into ONNX, an open source platform for distributed learning training, and has become one of the standard collaborative mechanisms for distributed machine learning, which has been used by nearly 40 enterprises at home and abroad, and the proposed large-scale decision tree model aggregation method constitutes the core technology of LightGBM, an open source platform for distributed decision trees, which has been used by more than 6,000 open source projects.
The programming and compilation optimization technology and data center programming and compilation technology dedicated to the field of focus have important practical significance and economic value for building the software ecology of domestic processor chips.
Application scenarios: artificial intelligence chips, supercomputing, etc
Cui Huimin has been engaged in the research of compilation software for domestic processor chips for a long time, and has made a series of scientific research achievements in domestic chip programming language, compilation technology, and cross-platform system software stack. In terms of heterogeneous programming and compilation, Cui Huimin focuses on domain-specific programming and compilation optimization technologies, including compilation and optimization for AI and communication, aiming to solve the programming problems brought by heterogeneity to programmers and give full play to the processing potential of domain-specific chips. Heterogeneous acceleration chips such as artificial intelligence chips, network processing chips, and Shenwei processors have made breakthroughs in a number of key technologies in the direction of programming and compilation, and have been applied to industry-leading artificial intelligence chips and supercomputing chips, which have been highly praised by the processor chip team and have played an important role in promoting the ecological construction of domestic chips.
For the first time, the security weaknesses of a number of internationally important cryptographic algorithms were discovered, and the quantum guessing decision attack method of stream cryptographic algorithms was proposed for the first time, which had a significant impact on the formulation and application of standard cryptographic algorithms in scenarios such as mobile communication and the Internet of Things.
Application scenarios: mobile communications, Internet of Things, etc
Ding Lin has long been committed to the research of cryptanalysis and decipherment under classical and quantum computing conditions, and has discovered for the first time the security weaknesses in many internationally important cryptographic algorithms such as the international 5G candidate encryption algorithm SNOW-V, the international GPRS standard encryption algorithm GEA-1/GEA-2 and the ISO/IEC standard encryption algorithm Grain-128a, and has achieved a number of cryptanalysis and decipherment results with the best computational complexity in the world For the first time, the algorithm proposes a quantum guessing decision attack method based on stream cryptography algorithms, which proves that cryptographic algorithms such as ISO/IEC standard encryption algorithm SNOW 2.0 are insecure under quantum computing conditions, and its research results have a significant impact on the formulation and application of standard cryptographic algorithms for mobile communications, Internet of Things and other scenarios, and have important academic value for enriching cryptanalysis theories under classical and quantum computing conditions.
Promote the research and development of fine-grained visual semantic models, vision and language, and promote the large-scale application of the latest research results of visual basic models and multimodal pre-training in multiple scenarios.
Application scenarios: multimedia information retrieval, AI intelligent generation, embodied intelligence, etc
Promoted the development of Microsoft's Xiaoice cross-modal poetry creation model, generated a total of 1 million original poems, and published the world's first collection of poems created by artificial intelligence. In addition, the image super-resolution and high-definition image stylization technology based on Visual Transformer have been successfully integrated into Microsoft OfficePlus and Designer products and have maintained stable operation. By improving image quality in real time and efficiently generating diverse images, the product experience in the field of office automation has been greatly improved, and better services have been provided for hundreds of millions of users. The video enhancement technology developed by Fu Jianlong and his team has also been successfully applied to Microsoft's Edge browser, serving 280 million users around the world.
Systematic research on swarm intelligence, crowd-intelligent perception computing, and human-machine-object fusion crowd computing will help improve system performance, optimize resource utilization, improve efficiency and quality, and provide theoretical and key technical support for the future human-machine-object cross-domain integration intelligent application and the construction of a human-machine harmonious symbiosis society.
Application scenarios: smart city, social governance, etc
Focusing on the core challenges such as the difficulty of dynamic coordination of edge-end resources, the weak adaptation of perception ability scenarios, and the poor context evolution of computing models, Guo Bin innovatively proposed the theory and method of edge-end integrated sensor-computing collaborative adaptive evolution with the characteristics of "organic linkage of resources, dynamic adaptation of capabilities, and automatic evolution of models", and took the lead in proposing and promoting the development of the emerging discipline of human-machine-object integration crowd intelligence computing in the world. His work explores the basic mechanisms, computational models, core algorithms and system platforms such as swarm intelligence perception and cognition, swarm intelligence collaboration and aggregation, swarm intelligence learning and evolution, and provides support for the construction of a crowd intelligence space with self-organization, self-learning, self-adaptation and self-evolution capabilities.
Artificial intelligence "encounters" multimodal remote sensing, innovates a new paradigm of multimodal AI remote sensing large model, unlocks the application potential of remote sensing big data, and promotes innovative progress in earth observation and sustainable development.
Application scenarios: surface element perception, disaster response, precision agriculture, smart city, climate change, etc
Danfeng Hong has developed a series of world-leading AI model methods, such as SpectralGPT, MiniGCN, MDL-RS, etc., which effectively bridge the gap between remote sensing big data and high-performance computing capabilities, and stimulate the potential of efficient extraction and analysis of multimodal remote sensing big data. On this basis, Hong Danfeng is committed to the development of a full-link multimodal remote sensing big data intelligent interpretation system with the design of a general multimodal AI model as the core, integrating big data, large computing power, large models, and large applications, which has greatly promoted the progress of intelligent computing in the field of earth science. The relevant research results have been successfully applied in many geoscience scenarios, such as disaster response, climate change, etc., and have been widely recognized by experts and scholars in the same field.
The scale of molecular dynamics simulation with first-principles accuracy has been expanded to 10 billion atoms, and the computational efficiency has been increased by 4 orders of magnitude, and the combination of intelligent supercomputing and physical models has been realized for the first time in the world.
Application scenarios: new material research and development, energy and chemical industry, drug research and development, etc
Weile Jia has been committed to the field of high-performance computing for a long time, and his main research interest is high-performance computing problems in AI-driven microscale simulations. In 2020, Jia Weile and his team expanded the scale of molecular dynamics simulation with first-principles accuracy to 100 million atoms for the first time in the world, and the computational efficiency was increased by more than 1,000 times, which was the first time in the world to realize the combination of intelligent supercomputing and physical models. In 2022, the AI model compression method was further developed, extending first-principles precision molecular dynamics simulations to tens of billions of atoms and 10 nanoseconds per day, an improvement of 4 orders of magnitude over the previous best work in related fields. In addition, he led the team to develop a large-scale sparse matrix solver dedicated to the field of first-principles computing, which has the highest peak value (65PFLOPS) and computational efficiency (5%) in the world on domestic supercomputing, and worked with collaborators to promote the simulation of complex metal heterojunctions to a scale of 2.5 million atoms.
Helping humans better understand the brain through AI innovation, and using those understandings to improve brain health and design brain-inspired AI.
Application scenarios: artificial intelligence, healthcare, embodied intelligence, etc
Dongsheng Li focuses on the intersection of AI and brain science, as well as AI and healthcare. In terms of understanding the brain through intelligent computing, Li Dongsheng led the team to explore EEG signal analysis and other directions, and built the first cross-dataset EEG pre-training model, which has played an important role in human brain signal understanding, encephalopathy diagnosis and encephalopathy mechanism analysis. In terms of brain-inspired AI, he and his team have carried out research on brain-inspired neural networks and brain-inspired machine learning algorithms, and the designed CircuitNet neural network provides a new infrastructure for machine learning, and the proposed unified behavior modeling framework provides a new technical route for the study of embodied intelligence. In addition, he has used machine learning innovations to understand the pathogenesis of diseases, discover effective drugs, and improve medical efficiency.
Focusing on AI-driven scientific research, using AI technology to analyze and understand biological data, through the study of multi-omics data, in-depth exploration of biological mechanisms, so as to facilitate drug discovery.
Application scenarios: precision medicine, genetics research, etc
Xiangtao Li led the team to develop a deep learning model for single-cell transcriptomics to improve the clustering performance of complex high-dimensional sparse spatial data. They have successfully solved the problem that traditional clustering algorithms are difficult to accurately measure and identify the distribution and structure of complex biological data, and this progress has advanced the analysis of cellular data and the understanding of gene regulatory mechanisms. In addition, the team conducted in-depth research on the prevalent data loss and noise problems in cancer genomics data. His work has helped screen out variant genes that are highly associated with cancer, providing important theoretical support for cancer research. At the same time, he also explored a deep sequence language learning method based on proteomics data, and improved the characterization of biological sequences, which not only considers the sequential coding of sequences, but also makes full use of the biological features in the sequences, providing a new perspective for the analysis of biological data.
Following the first principles, this paper focuses on improving the functions of the current AI paradigm represented by deep learning from the basic methodology level, and then explores the limit state form of machine intelligence based on mathematical theory and its implementation path.
Application scenarios: artificial intelligence, industry, medical, etc
Liu Bin has fundamentally improved the function of the deep learning paradigm to improve its ability to deal with practical problems in the complex and dynamic physical world, including efficient continuous learning ability, uncertainty quantitative reasoning ability, independent exploration learning and decision-making ability, mechanism knowledge and data fusion ability, small sample learning ability, etc. In addition, modern AI algorithms, such as deep learning and reinforcement learning, are used as tools to solve long-standing computing problems in mathematics, physics, and statistics.
Facing the characteristics of omics data, the new paradigm of interdisciplinary research and transformation of AI-driven omics effective analysis and precise intervention is innovatively developed, which provides guidance for precision medicine research on complex diseases.
Application scenarios: drug discovery, immunotherapy, cell therapy, gene editing, etc
Qi Liu is committed to developing a research paradigm of cross-integration of artificial intelligence and bioomics ("AI for Omics"), using omics AI to empower data-driven precision medicine research and translation. Focusing on the effective intervention of complex diseases, it has developed a series of AI-enabled computing platforms for small/large molecule intervention and gene editing intervention, including the development of PanPep, a computing system based on meta-learning for antigen-TCR specific recognition, which effectively solves the problem of long-tail data distribution and small samples in this field, and provides methodological support for neoantigen recognition, TCR-T therapy and cell therapy in the field of tumor immunotherapy.
It is committed to the field of computer hearing for medical health, and explores new ways to use audio signals as a new digital phenotype and non-drug intervention to assist diagnosis.
Application scenarios: smart healthcare, etc
Qian Kun has long been committed to the research in the fields of affective computing, computer hearing and brain science, and has continued to deepen his research direction in the research direction of "computer hearing for medical health". He established the world's first computer auditory medicine database platform, Voice of the Body, which released the first publicly available and accurately labeled heart sound database collected by a single institution, Shenzhen Heart Sound Database (HSS), and the first public somatization disorder speech database, Shenzhen Somatic Speech Corpus, and is working with many top hospitals in China to use unified standards to collect, annotate, and build subsequent databases.
In addition, based on the physiological structure of the human ear hearing system and artificial intelligence technology, he has constructed a brain-like hearing model, which approximates or even surpasses the acoustic signal processing function of the human ear hearing system, which can effectively solve the problems that traditional acoustic features are difficult to simulate the nonlinear perception and nonlinear amplitude gain characteristics of the human ear hearing system on speech, and the deep learning model has poor interpretability.
Focusing on the modeling and learning of complex graph data, the theory of heterogeneous graph modeling and representation is established, the trusted graph neural network model is proposed, and the open source graph learning platform is developed to promote the wide application of graph intelligence.
Application scenarios: finance, e-commerce, telecommunications, energy, etc
In view of the challenge of difficult representation of structural knowledge of complex graph data, Ishikawa established the theory of heterogeneous graph modeling and representation, which broke through the problem that homogeneous graphs cannot effectively model complex systems, and led the transformation of complex system modeling and analysis methods from homogeneous graphs to heterogeneous graphs. In order to solve the challenge that the knowledge of the trusted structure of graph data is difficult to guarantee, a trusted graph neural network model is proposed to alleviate the black-box learning dilemma of graph neural network and explore the implementation of trustworthy artificial intelligence. In addition, in view of the challenge that the actual structure of industrial graph data is difficult to use, Ishikawa proposed a graph modeling and mining technology for typical applications (recommender system and network security), and built an independent and controllable graph machine learning platform, breaking through the monopoly of foreign platforms and promoting the industrial application of graph intelligence technology.
A credible biomedical model has been designed to conduct top-down, micro-to-macro in-depth research on cross-scale and cross-modal biomedical data, and the relevant results have been successfully applied to clinical diagnosis and treatment.
Application scenarios: intelligent clinical diagnosis and treatment, etc
Wang Haishuai proposed an efficient representation learning algorithm for graph and time series biomedical data based on interpretability features, which identified the spatial correlation and temporal interaction between data, and solved the problems of trusted learning such as spatiotemporal evolution data modeling and interpretability of biomedical big data. In addition, he also explored the structural correlation of cross-modal data representation, filled the gap of cross-scale multimodal biomedical data analysis from genotype to phenotype, solved the problem of insufficient development and utilization of biomedical big data and low clinical decision-making efficiency, clarified the pathogenesis of diseases from multiple levels, and improved the ability of early screening and intervention of diseases. The research results have been successfully applied to autism screening, early screening for depression, estimation of glomerular filtration rate and prediction of kidney disease grading, segmentation of liver lesions, rapid diagnosis of cerebral palsy, etc., and have been put into clinical use in Veterans Hospital, Barnes-Jewish Hospital, the Second Affiliated Hospital of Zhejiang University, and the Affiliated Children's Hospital.
It is committed to the intelligent computing and genetic construction of energy information materials, and greatly reduces the R&D cycle and R&D cost of new materials.
Application scenarios: materials, energy, etc
Aiming at the major strategic goals of the national "dual carbon plan", Wang Hui is committed to the digital and intelligent research and development and industrial application of green and environmentally friendly new energy information materials. The new paradigm of scientific research based on multi-scale simulation, high-throughput computing and machine learning, focusing on experimental science guided by intelligent computing, aims to construct a gene database intrinsically related to "microstructure-macroscopic physical properties", and comprehensively accelerate the process from new material design to engineering application. His team focuses on the research of new green solid-state refrigeration, proposes an efficient calculation method for multi-physics regulation order parameter transformation, greatly improves the prediction accuracy and evaluation process of solid-state refrigeration performance, explores mathematical descriptors and machine learning models, establishes data-driven high-throughput computing simulation processes and databases, and provides a genetic blueprint for the design and development of zero-emission and energy-efficient new refrigeration materials, which is expected to accelerate the pace of cultivating disruptive technologies and their applications, leading the green refrigeration technology revolution and serving the transformation and upgrading of the traditional refrigeration industry.
Breaking through the physical limits, we are committed to researching innovative in-memory computing technologies and chips with higher parallelism and higher energy efficiency.
Application scenarios: device-side and edge-side AI scenarios
In 2022, Wang Shaodi led Zhicun Technology to take the lead in realizing the mass production of the world's first in-memory computing SoC chip WTM2101. After entering the market for about a year, WTM2101 has landed AR glasses, smart watches, TWS earphones and other market products, with revenue of more than 10 million. In addition, WTM2101 chips will also be implemented in medical, industrial and other application scenarios, enabling the improvement of end-side AI capabilities. At present, the WTM-8 series of edge-side computing power chips independently developed by Zhicun Technology is also about to be mass-produced, which can provide at least 24Tops computing power and consume only 10% of similar solutions in the market, which will help mobile devices achieve higher performance image processing and spatial computing.
Deeply cultivating voice technology, building Xiaomi's full-stack self-developed voice technology, focusing on multi-device human-to-human interactive acoustic technology and human-computer interactive voice technology, providing users with a more natural, free, and stress-free intelligent voice interaction mode, allowing users to enjoy the better life brought by "sound energy intelligent computing".
Application scenarios: mobile phones, IoT devices, smart cars, and other terminal devices
Wang Yujun set up the acoustic voice team of Xiaomi Group from 0 to 1, established an interactive acoustic voice technology system based on hardware, software and algorithms, and gave smart devices "mouths" and "ears". It provides self-developed sound algorithms for Xiaomi's 150 million mobile phones and AIoT devices, covering 79 categories and 5,312 smart terminals, and a self-developed technology system covering speech recognition, speech synthesis, voice pickup and noise reduction, sound perception, acoustic measurement, and sound playback algorithms, as well as acoustic voice algorithms in 6 major directions and 24 minor directions, and provides an average of 1.26 billion online voice services per day for "Xiao Ai", bringing users a more extreme intelligent voice interaction experience. After 5 years of technical practice, we worked with the team to create Xiaomi's "personalized and emotional sound interaction" system.
He has long been committed to solving intelligent computing problems in massive biological sequence analysis, combining interpretable artificial intelligence methods to learn complex key patterns of biological sequences, and trying to capture molecular functional associations from the sequence level and resolve life codons.
Application scenarios: clinical diagnosis, drug screening, etc
Focusing on the four-step route of "extraction-fusion-optimization-learning", Wei Leyi proposed a sequence representation method driven by expert knowledge and data, as well as a multi-scale deep learning model for biological sequence semantic learning. In addition, based on the idea of transfer learning, he designed a meta-learning strategy with mutual information maximization, which broke through the bottleneck of insufficient sample size of target task labeling. Through the joint drive of biological knowledge and deep learning, the combination of drugs and targets and sensitivity detection are realized efficiently and accurately, and the biological mechanism is innovatively introduced into modeling, which alleviates the problem of unexplainable models, and related technologies have been applied to many hospitals, and the developed cancer prognosis evaluation system has been applied to the treatment of nasopharyngeal cancer and breast cancer.
Using artificial intelligence technology theories and methods, mining the characteristics of the human brain, exploring the brain mechanism of cognitive function, and benefiting from the inspiration and inspiration obtained from brain science exploration, developing new artificial intelligence theoretical methods to achieve two-way promotion of information science and brain science.
Application scenarios: industrial, medical, etc
Wu Xia focuses on the research of brain data analysis based on artificial intelligence, and takes the China Brain Project as an opportunity to propose a series of algorithms for intelligent extraction of spatio-temporal brain features according to the characteristics of brain data, so as to realize the monitoring and evaluation of brain status of people in special positions. In addition, in order to solve the problem of "how to characterize the brain based on complex brain function data", a series of artificial intelligence analysis algorithms were constructed for the extraction of neurological markers of cognitive function based on the characteristics of neuroimaging data. It also applies AI-related theories and algorithms to brain diseases such as depression, autism, and Alzheimer's disease. It not only deepens the understanding of the neural mechanism of brain diseases, but also helps to promote the proposal and development of new diagnosis and detection methods.
This paper clarifies the reasons for the stronger robustness of type 2 fuzzy system, and proposes a variety of efficient training methods for fuzzy system, which are successfully applied to intelligent control, intelligent decision-making, brain-computer interface, etc.
Application scenarios: intelligent control, affective computing, brain-computer interface, etc
Wu Dongrui is one of the earliest scholars in the world to conduct research on type 2 fuzzy systems, and systematically studies the robustness, continuity, computational cost, and optimization methods of interval type 2 fuzzy controllers. For the first time, Type 2 fuzzy logic was used to completely realize the perceptual computer proposed by Professor Zadeh of the University of California, Berkeley, the founder of fuzzy logic, so that people can interact with the computer through natural vocabulary and be applied to various decision-making scenarios such as weapon system evaluation. In view of the challenge that the existing fuzzy system can only process low-dimensional small data, this paper draws on the methods of dropout and batch normalization in deep learning, proposes a fuzzy system training method that can efficiently process high-dimensional big data, and establishes the Python open-source toolbox PyTSK.
It took the lead in proposing the "chip-based high-dimensional optical neural network" and the high-dimensional optical computing architecture, and developed the high-dimensional optical computing chip and optoelectronic hybrid computing board, and the computing power has exceeded the 100 Tops level, providing a new technical approach and method for the development of ultra-large computing power photonic computing prototypes.
Application scenarios: photonic computing, etc
Focusing on optical artificial intelligence and large-capacity optical interconnection, Xie Peng focused on the research of integrated optical chips and intelligent optical information processing, combined with flat on-chip multi-wavelength light source system, large broadband optical computing chip and technology, multi-dimensional information reuse technology and photoelectric conversion module, etc., to develop a chip-based high-dimensional optical neural network system, and developed a new intelligent photonic processor with a high-dimensional optical computing architecture, with a single-chip computing power has reached the 100-Tops level, and is expected to push the computing power to 1,000-Tops It provides a new idea for expanding the computing power boundary of photonic computing, which is expected to increase the computing speed by 3-4 orders of magnitude and reduce the power consumption by 2 orders of magnitude. Beijing Core Computing Technology Co., Ltd. has realized the technological transformation and industrial implementation of relevant scientific and technological achievements, and provided key technical support for user units.
Through the overall security design of the operating system, 5G network, and cloud-native system, the integrated security of the cloud and the edge is enhanced, the secure intelligent computing is realized, and the security permissions of the cloud-native system are automatically configured using large models combined with program analysis technology around the world.
Application scenarios: cloud-native security, 5G core network security, network security, etc
Xue Hui is committed to integrating the security and defense of the cloud edge, and abstractly planning the security integration of the operating system, 5G network, and cloud native system. Considering the integration of operating system resource isolation, 5G network slicing resource isolation, cloud native system security configuration isolation, etc., the formal proof of key components is made, and the constraints of security and network computing power are jointly solved. Give full play to the computing power scheduling technology brought by the 5G network, and realize the safe scheduling of computing power in mobile scenarios through the combination of 5G network measurement technology and cloud native technology.
Classical and emerging technologies such as operations research, theoretical computer science, machine learning, and quantum computing are intersected to form (quantum) machine learning solutions for combinatorial optimization problems, and are used in scenarios such as chip design and drug design.
Application scenarios: chip design, drug design, etc
Yan's research focuses on new artificial intelligence solvers for operations research optimization problems, using machine learning and other methods to train models, automatically obtain relevant solution strategies or form inference models, and at the same time, introduce front-end time series prediction and quantum computing technologies into the solution. Yan Junchi researches machine learning theories and methods for combinatorial optimization problems, and responds to the application of EDA and drug design in the three leading industries of artificial intelligence, biomedicine, and integrated circuits in Shanghai and even the Yangtze River Delta. At the same time, through machine learning, the front-end perception, prediction and back-end decision-making are further coordinated to form an end-to-end prediction-decision-making paradigm, and further drive the end-to-end paradigm of autonomous driving. In addition, quantum computing is introduced into the solution of combinatorial optimization problems, especially quantum machine learning models and algorithms are proposed to deal with more general constraint combinatorial optimization problems.
The methods developed by him and his collaborators such as trRosetta, I-TASSER, COACH, BioLiP, etc., have become important methods in the field of structural bioinformatics, and are widely used by experimental scientists such as structural biologists, effectively reducing experimental costs and promoting the development of structural biology and other fields.
Application scenarios: protein and RNA structure prediction, etc
Jianyi Yang is engaged in the field of structural bioinformatics, focusing on protein and RNA structure prediction. Based on the intersection of artificial intelligence and mathematics, he cooperated with Professor David Baker of the University of Washington, Professor Yang Zhang of the University of Michigan and other authoritative experts to jointly develop a series of new methods with important international influence, including the development of the first protein structure prediction method with an accuracy beyond AlphaFold1, trRosetta, and a series of important promotions, and the application of transformer network to develop an automatic RNA tertiary structure prediction method trRosettaRNA, which improves the accuracy of structure prediction, has made an important contribution to the development of I-TASSER, a protein homology modeling method.
Focus on the use of high-performance computing technology research to make AI large models run faster, more efficiently, and at a lower cost.
Application scenarios: cloud computing, chip design, biomedicine, autonomous driving, etc
You Yang has proposed a series of methods, including the LARS optimizer, the LAMB optimizer, and the DATE framework, which enable AI training to scale to thousands of processors without loss of accuracy, and break the ImageNet and BERT world records for training speed. To further push the limits of scalability, Yang You designed a new method for efficient memory automatic distribution and led the establishment of Colossal-AI, a next-generation large-scale AI system, one of the fastest-growing open source projects in the field of scalable AI. With Colossal-AI as the core product, You Yang led the establishment of Luchen Technology, which successfully achieved industrialization and cooperated with a number of Fortune 500 companies/China, Southeast Asian technology giants, national research institutions in Asia, overseas supercomputing centers, etc., to promote the commercialization of AI large models.
TableGPT, the first structured data model, has been independently developed, which not only demonstrates highly accurate quantitative and quantitative analysis capabilities, but also supports seamless integration into the infrastructure of enterprises such as databases and data warehouses.
Application scenarios: industrial manufacturing, financial securities, civil aviation transportation, etc
Zhao Junbo proposed and presided over the development of TableGPT, which was released in 2023 and is regarded as the world's first large-scale model product that connects relational databases and data warehouses. Compared with the general large model, TableGPT has achieved a complete closed-loop from "seeing business data" to "understanding business logic" to "chatting about business in real time" with its unique database docking capabilities. TableGPT has broken through a number of key technical difficulties, and its system functionality, robustness, and usability have significant advantages, and it has accumulated many successful landing cases in industrial manufacturing, financial securities, civil aviation transportation and other fields, bringing certain economic benefits.
Solve the sustainable development problems of the power and energy industry with intelligent computing technology, including monitoring carbon emissions, reducing carbon emissions, improving energy efficiency, and promoting the use of renewable energy.
Application scenarios: power energy systems, etc
Focusing on "AI Helps Low-carbon Energy Transition", Zhao Junhua solves key scientific problems in the field of energy and environment based on intelligent computing technology, providing new perspectives and solutions for intelligent management and low-carbon optimization of energy systems. In 2023, he led the development of the first carbon satellite-based carbon measurement AI model, which integrates multi-source data for carbon emission monitoring, greatly improving the accuracy of emission estimation for specific emission sources.
In the direction of power system optimal scheduling and planning, his team proposed a two-stage low-carbon power system dispatching model based on zero-sum revenue data envelopment analysis to improve the optimization results and help the power system reduce carbon emissions. On the issue of co-optimization of the electricity-carbon market, he and his team and the State Grid Corporation of China jointly developed an artificial intelligence-driven co-simulation technology for the electricity-carbon market, which provides a scientific basis for the design of the mechanism and rules of the two markets. In 2024, he cooperated with Galaxy International and China Southern Power Grid Corporation to successfully list the first green bond based on high-precision carbon data certification on the Hong Kong Stock Exchange, successfully realizing the application of corporate carbon metering technology in the field of green finance.
Using cutting-edge computing technologies such as big data and artificial intelligence, we will intersect research on key issues in the field of management and social sciences such as human resource management and labor economics, and serve people's better career development through forward-looking interdisciplinary basic research and innovative industrial applications.
Application scenarios: intelligent recruitment, enterprise digital and intelligent transformation, etc
Zhu Hengshu has been engaged in the interdisciplinary scientific direction of "talent and management computing" for a long time, and he led the team to propose a deep sequence analysis technology for talent career prediction and a personalized recommendation technology for enterprise talent development needs. In addition, the market-driven skill value evaluation method and labor market trend dynamic analysis technology are proposed, and the fine-grained large-scale labor market analysis application is realized.
Based on the above achievements, he presided over the development and implementation of the world's first big data intelligent talent management solution, as well as a number of innovative commercial products such as bilateral reciprocal recruitment recommendation system, which have served hundreds of millions of Internet users and tens of millions of enterprises, and achieved significant social and economic benefits.
This paper proposes a new concept of framework nucleic acid intelligent diagnosis molecular machine with molecular primitive synergistic effect based on framework nucleic acid and nucleic acid molecular motifs as structural units, nucleic acid hybridization, molecular recognition, and nucleic acid molecular network.
Application scenarios: intelligent biosensing, biocomputing, disease diagnosis, etc
Zuo Xiaolei focuses on the design methods of framework nucleic acids and nucleic acid molecular motifs, and the synthesis technology of high-throughput standardized framework nucleic acids and nucleic acid molecular motifs in order to solve challenges such as poor biomolecular perception ability (biomolecular recognition performance, etc.) and insufficient information processing ability (computational efficiency of biomolecules, etc.).
By using the self-assembly behavior and logical operation ability of nucleic acid molecules, a new biocomputing framework with nanostructures, intelligent molecular networks and storage and computing integration with various structures and functions can be accurately designed is constructed, so as to provide new technologies for human beings to simulate the information computing function of life matter and explore the basic laws of information computing in living systems. This will provide technical support for improving the information perception ability, information processing ability and intelligent behavior of biological computers, and has application potential in the fields of live cell information storage, intelligent live cell biocomputing, micro-nano robots, and nano intelligent diagnosis and treatment.