Shulou(Shulou.com)06/02 Report--

Https://www.toutiao.com/a6652872822005694987/

1. The basic Theory of New Generation artificial Intelligence

Focus on the major scientific frontiers of artificial intelligence, focus on breaking through the bottlenecks of basic mechanisms, models and algorithms of artificial intelligence, and focus on the layout of a new generation of basic theory of artificial intelligence that may lead to changes in the paradigm of artificial intelligence. to provide a strong scientific reserve for the sustainable development and deep application of artificial intelligence.

1.1 New Generation Neural Network Model

Draw lessons from neural cognitive mechanism and mathematical methods of machine learning, carry out the research on new theories and methods, such as nonlinear mapping of neural network model, automatic evolution of network structure, functional alienation of neurons and modules, small sample learning / weak label / unlabeled sample learning, interpretable, and so on. Essentially enhance the scope and ability of deep neural networks to support and solve real artificial intelligence problems.

1.2 Adaptive perception for open environment

In view of the problem that the transformation of application scenarios can easily lead to a sharp decline in the performance of intelligent systems, hierarchical network structures with strong adaptability, machine learning strategies that can learn continuously and general efficiency measurement methods are developed. break through the difficulties such as unsupervised learning, empirical memory utilization, implicit knowledge discovery and guidance and attention selection, and promote the formation of universal perceptual intelligence in open environment and changing scenarios.

1.3 Cross-media causal inference

This paper studies a new machine learning method based on the formation of cross-media human common sense knowledge, and carries on the bottom-up deep abstraction and induction of cross-media data with the support of common sense knowledge, and effectively controls the top-down deduction and reasoning of uncertainty. establish a new model and method for the coordination and supplement of logical reasoning, inductive reasoning and intuitive epiphany. Realize the leap of cross-media from intelligent relevance analysis to causal inference supported by common sense knowledge.

1.4 Game decision under the condition of incomplete information

According to the game characteristics of human economic activities and man-machine confrontation under the condition of incomplete information, and combined with the progress in the fields of machine learning, cybernetics and game theory, this paper studies the dynamic mechanism and optimal decision-making model of game confrontation in uncertain and complex environment, and integrates confrontation learning and reinforcement learning with dynamic game theory. Realize the task-oriented general intelligent basic model and dynamic game decision-making theory in incomplete information environment.

1.5 the mechanism and calculation method of group intelligence emergence

This paper studies the organization mode and incentive mechanism of large-scale group cooperation in an open, dynamic and complex environment, establishes a compound incentive algorithm that can be expressed, calculated and regulated, and explores the emergence mechanism and evolution law of individual contributions converging into swarm intelligence. break through the global goal-oriented swarm intelligence evolution method and spatio-temporal sensitive swarm intelligence collaboration to achieve predictable, guided and sustainable swarm intelligence emergence.

1.6 people-in-the-loop hybrid enhanced intelligence

Task modeling, environmental modeling and human behavior modeling under uncertainty, vulnerability and openness are studied, and human-in-the-loop machine learning methods and hybrid enhanced intelligent evaluation methods are developed. the advanced cognitive mechanism of human analysis and response to complex problems is closely coupled with the machine intelligence system to effectively avoid the decision-making risk and system out of control caused by the limitations of artificial intelligence technology. Realize the two-way cooperation between man and machine and the convergence of solving complex problems.

1.7 Man-machine-object cooperative control method in complex manufacturing environment

Facing the complex multi-dimensional man-machine-object coordination problem in discrete manufacturing and process industry, the distributed networked collaborative control method across layers and domains is studied, which breaks through the theory of ternary collaborative decision-making and optimization, realizes the virtual-real integration and dynamic scheduling of man-machine, and explores the reconstruction of unmanned machining production line and intelligent interaction between man and machine. It provides theoretical and methodological support for the exploration of intelligent factory development model and the establishment of standard system.

two。 Key common technologies for major requirements

Centering on the urgent need to enhance the international competitiveness of China's artificial intelligence, facing the major demand, we should break through the key common technologies of the new generation of artificial intelligence, with algorithms as the core and based on data and hardware. we will comprehensively enhance the capabilities of perceptual recognition, knowledge computing, cognitive reasoning, collaborative control and operation, and human-computer interaction to form an open, compatible, stable and mature technology system.

2.1 generalizable domain knowledge learning and computing engine

To meet the needs of cross-border integration of new business type and knowledge innovation services, to overcome the key technologies needed for the establishment of large-scale and comprehensive knowledge centers. Break through the core technologies such as knowledge processing, in-depth search and visual interaction, and form the capabilities of concept recognition, entity discovery, attribute prediction, knowledge evolution and relationship mining, so as to realize the automatic acquisition of knowledge with continuous growth. form independent induction and learning ability from data to knowledge and from knowledge to service. Service verification is carried out in 1-2 knowledge-intensive areas to reach or exceed the average Q & A service level of domain experts.

2.2 Cross-media analysis and reasoning technology system

To meet the major requirements of cross-media content regulation, situation analysis and cross-modal medical analysis, this paper studies the unified representation theory, model and acquisition method of cross-media multiple knowledge. Build a knowledge graph and analytical reasoning technology with a level of more than one billion to adapt to the evolution of cross-media content, and establish a generalization mechanism from directional reasoning to general reasoning. Backtracking and interpretable cross-media intelligent reasoning is realized in 1-2 typical application scenarios, and the accuracy exceeds the level of intermediate experts in the field.

2.3 active scene perception under Cognitive tasks

Aiming at the cognitive tasks such as target search, scene analysis and interpretation in complex environment, the technologies of active visual perception, 3D modeling and location of natural scene are studied. Acoustic environment detection in noisy scenes and speech active perception technology based on auditory feedback mechanism are studied. Visual and auditory collaborative cognitive technology for actively discovering new targets and their attribute knowledge from natural scenes is studied. The experimental platform of typical scene is established and the function is verified.

2.4 Research on the aggregation of Group Intelligence stimulation for Group Software Development

For large-scale complex group intelligence innovation activities such as group software development, this paper studies the technologies such as coordination and evolution of group intelligence community, decomposition and adaptation of group intelligence tasks, analysis and evaluation, quality control and reuse fusion of group intelligence innovation products, and studies the technology of code tagging, test verification and defect repair of group intelligence software products. We will study the mechanism and technology of group intelligence stimulation and convergence in the group intelligence open source community, promote the formation of a million-scale group intelligence innovation and talent training ecology oriented to specific fields, and effectively promote the establishment of artificial intelligence technology and application ecology.

2.5 Research on hardware and software technology of man-machine cooperation

Facing the application scenarios of human-computer cooperation such as intelligent manufacturing and autopilot, a man-machine cooperation technology platform with the integration of software and hardware is studied and constructed. Study the models and methods that adapt to real-world situational understanding and collaborative decision-making; study new learning methods that mix human intuition, experience and behavior from man-machine cooperation; develop a new hybrid computing architecture and intelligent software and hardware that can naturally understand the environment and situation and deal with large-scale knowledge.

2.6 Autonomous intelligence precise perception and control of unmanned system

According to the development requirements of autonomous intelligence such as sea, land, air and sky unmanned platform, the collaborative perception method based on multi-sensor information fusion in unconstrained environment is studied, and the semantic modeling and understanding methods of large-scale scene are studied. realize map construction, thorough perception and dynamic cognition in complex environment; study the methods of fast and accurate segmentation, detection, location, tracking and recognition of multi-source heterogeneous perceptual objects in complex scenes. Establish or use the existing autonomous intelligence system for technical verification to achieve natural, accurate and safe interaction and precise control in the autonomous intelligence unmanned system.

2.7 dexterous and precise operation learning of autonomous agents

According to the demand of autonomous operation in complex unmanned production system, the teaching and efficient demonstration of complex dexterous and accurate operation skills based on intelligent human-computer interaction are studied. Machine learning and skill generation of complex skills such as grasping, alignment, approach, loading and other complex skills are studied; dexterous motion planning and coordinated control of autonomous agents are studied to realize the motion mapping from skill to dexterous operation. This paper studies the representation method of multi-level operation skills to realize the knowledge-based expression of complex skills, and carries on the technical verification of dexterous operation skills learning around typical scenes such as precision assembly.

3. Intelligent Chip and system

Focusing on the key links of the development of artificial intelligence industry and application ecological infrastructure, this paper focuses on new sensing devices and systems from the point of view of artificial intelligence innovation platform and basic support. the key technical standards of artificial neural network and artificial intelligence open source open platform.

3.1 New sensing devices and chips

This paper studies the signal processing and information processing mechanism of biological perception channels, such as vision, hearing, touch, smell, etc., and develops new sensing devices, chips and corresponding neural network perceptual information representation, processing, analysis and recognition algorithm models. develop a biological-like, performance-surpassing biological perception system and achieve functional verification.

3.2 key standards and verification chips for neural network processors

The design of neural network computing instruction set supporting training and reasoning, the formulation of neural network representation and compression standards, the development of efficient basic algorithm library and development interface standards, and the realization of supporting development tool chain. establish an open chip platform standard that does not depend on the specific chip implementation, and realize the unification of software and hardware system interface. Implement verification chips and sample applications that support the above instruction sets, algorithm libraries, standards and development interfaces.

3.3 artificial Intelligence Open Source Open basic platform and Intelligent operating system prototype

Research on intelligent hardware resource management technologies such as intelligent sensors, intelligent processing chips and intelligent controllers, and develop an open source and open foundation platform of artificial intelligence that supports a variety of heterogeneous hardware. The intelligent software such as intelligent algorithm and knowledge base and data resource management technology are studied, and the general open source algorithm base, model base and basic software platform of human-computer interaction are developed. Support the distributed allocation and scheduling of large-scale intelligent tasks, establish an open source artificial intelligence ecology that encourages innovation, organic integration and rapid application, and support the development of basic software and core hardware such as intelligent operating systems.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.