The country allocated 390 million to support 13 transformative technology research and development, and 4 related to medicine

Recently, the Ministry of Science and Technology of the People's Republic of China issued the "Notice of the Ministry of Science and Technology on the publication of the 2017 National Project Application Guidelines for the Key Scientific Issues of Key Technologies for Transforming Technology in Key National R&D Programs" (hereinafter referred to as the Notice). The notice clearly states that the key scientific issues of transformative technology will be highlighted in 2017. The special project will deploy 13 research directions around the precise reconstruction of chemical bonds, superstructure materials, accurate mesoscopic measurement of organs, dexterous prosthetics, artificial intelligence, and new terahertz radiation sources. The total allocation of funds is about 390 million yuan.

At the same time, the "Notice" proposes that in the same direction, only one item is supported in principle. Only when the evaluation results of the project are similar, the technical routes are obviously different, and two items can be supported at the same time, and a dynamic adjustment mechanism is established. According to the mid-term evaluation results, Continue to support. Unless otherwise stated, there are no more than 5 topics under each project, and each project contains no more than 10 units. The project declaration needs to have relevant research foundation, and has been supported by the national science and technology plan and has good implementation effect and great application prospects. The person in charge of the project must have the experience of undertaking major national science and technology projects. The implementation period of the project is generally 5 years. The declared project requires a clear and visible 5-year overall goal and a 2-year target and assessment indicator (or research progress). The project implements “2+3” segmentation funding, and evaluates the completion of the target in about 2 years of project implementation, and determines the follow-up support mode of the project based on the assessment.

In addition, the Notice also stipulates detailed reporting qualifications. For example, government agencies may not take the lead or participate in the declaration. The person in charge of the project (subject) must have a senior professional title or doctoral degree. Foreign scientists employed in the Mainland units and Hong Kong, Macao and Taiwan. Regional scientists can be the key project (project) leaders. Full-time candidates must be provided with full-time employment certificates by the employers in the Mainland. Part-time candidates must be provided with valid certificates for employment by both the Mainland employers and overseas units. The paper project pre-declaration is submitted together.

Project (subject) responsible person limited to declare one project (subject); national key basic research development plan (973 plan, including major scientific research plan), national high-tech research and development plan (863 plan), national science and technology support plan, national international The person in charge of scientific and technological cooperation, national major scientific equipment and equipment development, public welfare industry research special projects, national science and technology major projects, and national key research and development programs, special projects (including tasks or topics) may not take the lead in reporting projects (projects). The person in charge of the research project (excluding the task or project leader) who is the key project of the national key R&D plan shall not participate in the application project (problem).

Among the key projects in this application, four of them are related to medicine. They are: accurate mesoscopic measurement of three-dimensional structure and function information of intact organs, accurate mesoscopic measurement of human organ chips, and new research for biomedical applications. Terahertz radiation sources, dexterous prosthetics of organisms and their neural information channel reconstruction.

1. Precise mesoscopic measurement of three-dimensional structure and function information of intact organs

Research content: To develop new principles and methods of precise mesoscopic measurement for biomedical frontier scientific problems, to break through the existing research methods, it is difficult to carry out high-resolution three-dimensional measurement bottlenecks in large-volume samples, and to realize multi-dimensional life science big data in important organs. High precision acquisition, reconstruction and visualization. Further, in an auxiliary coordinate or annotation having representative anatomical structures, tissue features, and physiological and pathological states, the structural and functional maps of different types of cells in the intact organ are visualized.

2. Precise mesoscopic measurement of human organ chips

Research content: Explore new principles and methods of mesoscopic measurement and characterization of biochemical characteristics of human organ chips, and establish high-resolution online with multi-parameter, multi-dimensional and multi-modality from multiple levels of molecules, cells, tissues, organs and even systems. Accurate detection means to realize real-time monitoring of micro-organs and objective evaluation of micro-structure bionic state, and to study the model characteristics of organ chips, verify their similarity with human tissues, and provide technical support for drug screening and disease treatment.

3. New terahertz radiation source for biomedical applications research

Research content: Facing biomedical applications such as terahertz wave biological effects and detection, exploring the physical mechanism of terahertz radiation from the interaction of free electrons with emerging materials and new structures, revealing the basic laws of transformative terahertz radiation, breaking through traditional terahertz radiation The technical bottleneck of the source produces wide-band tunable, high-power, continuous-wave miniaturization and coherent terahertz radiation with a certain diffraction-free length.

4. Physic prosthetic limbs and reconstruction of their neural information channels

Research content: Explore the scientific goal of "re-manual human function", explore the design, manufacture, neural interface coding and decoding algorithm and neural interface hardware system of operation-aware integrated biological body, and its information channel reconstruction method with neural system, and The fusion and interaction of neural intelligence and artificial intelligence. Focus on the design and manufacturing principle of flexible prosthetic mechanism based on soft materials, extraction and decoding of neural signals, neural coding rules of human motion information and a new generation of neural control models, neural afferent mechanisms of sensor signals and natural sensory function reconstruction methods of prostheses Achieve a closed-loop, two-way neural interface that completes the ability to stabilize and continuously learn to improve functionality.

In addition, there are two related to AI:

1. The next generation of deep learning theory and technology

Research content: for ubiquitous (such as mobile computing), high-risk (such as precision medicine), high reliability (such as intelligent transportation) and other application scenarios, breakthrough deep learning theory foundation, single model structure, resource consumption, data dependence Strong bottleneck. Study the basic theory of the next generation of deep learning; non-neural network, resource-saving deep learning model, method and efficient optimization technology; deep learning methods and techniques suitable for small sample/unsupervised samples, reinforcement/game learning.

Assessment indicators: Aiming at the complex characteristics of deep learning model, such as highly nonlinear, parameter space layering and huge, establish a set of theoretical frameworks to reveal the working mechanism of deep learning, form a set of tools and methods for deep learning model analysis; The new machine learning model, method and technology of neural network structure make breakthroughs in the interpretability, high scalability and easy configurability of deep learning models; propose various deep learning models and methods with low storage and computing resource consumption, and design quickly. Efficient, new gradient and non-gradient optimization techniques for non-convex deep learning training, greatly enhance the ability to deploy deep learning techniques; develop deep learning methods and techniques for small samples, unsupervised samples, weakly labeled samples, and non-single labeled samples. Reduce the deep dependence of deep learning on large-scale high-quality annotation data; develop multi-event-triggered deep learning models and techniques to adapt to the open environment of the information society and the rapid emergence of new phenomena; expand the field of deep learning applications, and propose to apply to online learning Intensive learning, game learning Of learning methods and techniques.

2. New principles, new structures and new methods of deep neural network processors

Research content: Deep neural networks have played a key supporting role in a variety of cloud and terminal applications. However, existing chips are far from being able to meet the speed and energy efficiency requirements of deep neural networks. It is necessary to explore the design principles, architecture, instruction sets and programming languages ​​of new processors that can efficiently process large-scale deep neural networks. The impact of process (≤16nm) and new devices on the design of deep neural network processors. Develop a new deep neural network processor chip to explore the extremely low power neural network architecture with fully asynchronous features.

Assessment indicators: Develop a deep neural network processor sample that can handle large-scale deep neural networks (including 100 million neurons and 1 billion synapses). The sample supports the domestic deep neural network instruction set, integrates hardware neurons/synapses as its computing part, supports time division multiplexing of hardware neurons, supports mainstream deep neural network programming frameworks such as Caffe, TensorFlow and MXNet, and can perform multi-layer sensing. Machine (MLP), Convolutional Neural Network (CNN), Long and Short Memory Network (LSTM), Recurrent Neural Network (RNN), Generated Confrontation Network (GAN), and Faster Region-Based Convolutional Neural Networks (Faster-RCNN) The use of mainstream deep neural networks, measured energy efficiency and performance is more than 20 times that of NVIDIA GPU product M40. Design a benchmark set of deep neural network processors that cover applications such as speech, images, and natural language understanding. Design efficient deep neural network processor cores and on-chip interconnect structures. Develop programming languages, compilers, and assemblers for deep neural network processors. Develop driver and system software for deep neural network processors. The application of the deep neural network processor in more than one million mobile terminals is completed, and the intelligent tasks that originally required cloud computing can be processed locally in the mobile terminal.