Generative artificial intelligence is becoming a new technology base for informatization, digitization and intelligence, especially the generative model represented by the large language model has a strong knowledge coding and storage ability, text and ** comprehension and generation ability, as well as the reasoning ability of complex tasks, which is considered to be one of the technical routes that may realize general artificial intelligence, and is expected to become an important technology to promote a new round of scientific and technological change and economic development, and have a far-reaching impact on science and society.
At present, generative artificial intelligence is in a critical period of rapid change and update iteration, and is facing challenges such as unclear basic theoretical mechanism of intelligence emergence, high computational cost, need to improve the security of generated content, and difficulty in automatic evaluation. It is of great significance to enhance the national strategic position and international competitiveness by laying out the basic theoretical innovation and key technological breakthroughs of generative AI, driving systematic and original basic research, improving the full-cycle security of models from different levels such as pre-training data governance, alignment strategies, and knowledge enhancement, and promoting the development and application of generative AI technology.
1. Scientific objectives
This project explores the intelligent emergence mechanism of large models and improves their ability to deal with complex tasks, strives to solve the security and trustworthiness problems existing in the generated content, explores the knowledge fusion and innovative application of large models, enhances user experience and promotes the development of new technologies.
2. Funding research directions
Facing the frontier scientific direction of generative artificial intelligence and the major national strategic needs, in view of the current technical bottlenecks and challenges, the research is carried out from the six levels of "theoretical mechanism, architecture design, computing efficiency, security alignment, evaluation methods, and typical applications", focusing on the basic theory and key technologies of efficient and credible models. This project intends to fund the following research directions:
(1) Research on the mechanism of intelligent emergence of generative models
In view of the capabilities of intelligence emergence, situational Xi, instruction fine-tuning, and chain of thought in generative AI, effective analysis tools are designed to explore its mathematical mechanism, and on this basis, the generalization, multi-step reasoning, and complex task decomposition capabilities of the model are further improved.
(2) Research on new efficient neural network architecture for generative models
In order to solve the complexity of transformer architecture in generative artificial intelligence, this paper studies the efficient model architecture and cross-modal unified architecture, optimizes the main modules such as self-attention, improves the model processing ability and efficiency in long sequence and multimodal scenarios, and explores new architecture design methods such as non-transformer efficient architecture and recombinable modular architecture.
(3) Research on efficient training and reasoning methods for large models
This paper studies the general and efficient training and inference methods of large models of various scales (especially in the allocation of low computing power resources, or models with more than 10 billion or 100 billion parameters), proposes a distributed training algorithm with low communication and high concurrency that adapts to domestic software and hardware, improves the inference efficiency of the model under the condition of low resources, and explores the collaborative inference mechanism of the model cloud and edge, so as to ensure the performance and response speed.
(4) Research on the values and security alignment strategies of large models
This paper studies a sustainable, highly generalized, and strongly adversarial large model alignment technology that conforms to human value preferences, so as to realize the alignment and transcendence of safety ethics and cognitive reasoning ability, improve the harmlessness and effectiveness of large model applications, and enhance the high stability and application reliability of the training of large model safety ethics value alignment algorithms.
(5) Research on automatic evaluation methods of generative models
In order to solve the problem of difficulty in the evaluation of model-generated content, a multi-dimensional automatic evaluation method of generated content was studied, and the evaluation criteria and corresponding evaluation data sets were constructed in the dimensions of factual accuracy, usefulness, logic, safety and harmlessness, instruction compliance ability, and long document processing ability.
(6) Research on industry and professional models of generative AI
In order to solve the application challenges of generative AI technology in the fields of finance, education, medical care, and science, we study domain pre-training and domain instruction fine-tuning methods, build augmentation models that integrate domain knowledge, and explore breakthrough generative AI applications.
3. Funding Scheme
There are 6 projects to be funded, with a funding intensity of 500,000 yuan. The funding period is 1 year, and the research period should be "March 1, 2024 - February 28, 2025" in the application form.
4. Application requirements and precautions
(1) Qualifications for application
1.Experience in undertaking basic research projects;
2.Have senior professional and technical positions (professional titles).
Postdoctoral researchers, those who are studying for a graduate degree, and those who do not have a work unit or whose unit is not a supporting unit are not allowed to apply as applicants.
(2) Provisions on limited applications
1.This special project is not included in the scope of the total number of applications and projects undertaken.
2.Applicants can only apply for one research project in the same year.
(3) Precautions for application
1.Paperless application for special projects
The application application is submitted from January 18, 2024 to January 22, 2024 at 4 p.m.
2.Note to Applicants:
1) Before filling in the application form, the applicant should carefully read the relevant content of this special guide and the "2023 National Natural Science ** Project Guide", and the application projects that do not meet the project guidelines and related requirements will not be accepted.
2) This special project aims to focus on the core scientific problems and concentrate the domestic superior research team to conduct research, and become a special project group. Applicants should formulate the project name, scientific objectives, research content, key scientific issues, technical routes and corresponding research funds according to the specific scientific problems to be solved by the project and the research direction to be funded announced in the project guidelines.
3) If the applicant logs in to the scientific network information system and does not have a system account, please apply to the management contact person of the supporting unit to open an account), and write the application in accordance with the writing outline and related requirements.
4) Select "Special Project" as the funding category in the application, select "Research Project" as "Research Project" for the subcategory description, and select "Comprehensive Research Project of the Ministry of Science" for the note description(The application **1 shall be in accordance with the requirements of the research direction to be funded, and the application under F02 and F06 of the Ministry of Information Science shall be selected**.) Applications for projects that are inaccurate or unselected above will not be accepted).
5) Please write the application form according to the "Outline of Special Project-Research Project Application Writing".Please indicate "Basic Research in Generative Artificial Intelligence: xxx (fill in one of the 6 research directions to be funded)" at the beginning of the body of the application.
The application should highlight the limited goals and key breakthroughs, and clarify the contribution to the realization of the overall scientific objectives of the project and the solution of core scientific problems.
If the applicant has undertaken other science and technology projects related to this project, the difference and connection between the application project and other related projects should be discussed in the "Research Basis and Working Conditions" section of the main body of the application.
6) The applicant shall strictly follow the relevant provisions of the "Measures for the Management of Funds of National Natural Science ** Funded Projects" and the specific requirements of the "Instructions for the Preparation of the Budget Table of the National Natural Science ** Project", and carefully prepare the "National Natural Science ** Project Budget Table" in accordance with the basic principles of "target relevance, policy consistency and economic rationality".
3.Precautions for relying on units
1) The relying unit shall review the authenticity, completeness and compliance of the application materials submitted by the applicant;Review the relevance of the applicant's objectives, policy alignment, and economic reasonableness in preparing the budget.
2) The electronic application form and attachment materials of the unit should be confirmed and submitted through the information system one by one before the specified project application deadline, and there is no need to submit a paper application. After the project is approved, the paper signature and stamp page of the application form shall be bound at the end of the "Funded Project Plan" and submitted together. The information signed and stamped should be strictly consistent with the electronic application.
3) If the supporting unit has not uploaded the "2024 National Natural Science ** Project Application Commitment Letter" (hereinafter referred to as the "Letter of Commitment") in 2024, it shall upload the electronic scanned copy to the information system after the "Letter of Commitment" is signed by the legal representative and stamped with the official seal of the supporting unit (only need to be uploaded once this year). The relying unit can submit the application only after completing the above-mentioned commitment procedures.
4) Within 24 hours after the deadline for project application, the supporting unit shall submit the project application list of the unit through the information system**. After the list is submitted, the Natural Science Committee can receive the project application materials.
4.This special consultation method
The Second Division of Information Science of the National Natural Science Commission.
Contact: Wu Guozheng, Wang Zhiheng, Xie Guo. Contact**:
(4) Other matters needing attention
1.In order to achieve the overall scientific objectives of the project, the person in charge of the funded project should pay attention to the mutual support relationship with other projects of the project during the implementation of the project.
2.In order to strengthen the academic exchanges between projects, the project will set up a special project management coordination group, and will organize academic seminars in related fields from time to time. The person in charge of the funded project must participate in the above-mentioned academic exchange activities and carry out academic exchanges in earnest.
National Natural Science Commission.