The China Brain Project covers both basic research on neural mechanisms underlying cognition and translational research for the diagnosis and intervention of brain diseases as well as for brain-inspired intelligence technology. We discuss some emerging themes, with emphasis on unique aspects.

 Non-human Primate Research

Being phylogenetically most proximal to humans, NHPs are excellent animal models for studying human cognitive functions and for exploring pathogenesis mechanisms and therapeutic approaches for brain diseases. There is increasing interest in the use of NHPs among Chinese neuroscientists, and new NHP research facilities are being established in many research institutions (Cyranoski, 2016). For example, the Kunming Institute of Zoology of Chinese Academy of Sciences (CAS), which already has a large colony of macaque monkeys for research purposes, is in the process of further expansion into the National Primate Resource Center, potentially the largest NHP research resource in China. Yunnan Key Laboratory of Primate Biomedical Research in Kunming now houses a collection of gene-edited monkeys that may serve as models of Duchenne muscular dystrophy, autism, and PD. A NHP facility maintained by Chinese Academy of Medical Sciences in Kunming for vaccine development is now shifting its interest toward developing NHP models for brain diseases. Finally, the Institute of Neuroscience of Chinese Academy of Sciences in Shanghai has established the largest NHP research facility in East China for both macaque and marmoset monkeys, together with ten research laboratories working on reproductive biology, gene-edited models, systems neurophysiology, and cognitive behaviors of NHPs.

Surgical intervention and chemical induction have been used in the past to generate NHP models for brain disorders such as drug addiction, spinal cord injury, epilepsy, and PD and AD (Zhang et al., 2014). In recent years, transgenic and gene-editing approaches have begun to be used for developing NHP models for those diseases with prominent genetic causes, including Huntington’s disease, PD, Duchenne muscular dystrophy, and autism spectrum disorders (Chen et al., 2016, Liu et al., 2016a). Virus-mediated gene delivery was the main approach for generating transgenic monkeys expressing a-Synuclein andMeCP2. New gene-editing technologies have been used to generate gene-edited NHP models (Chen et al., 2016): TALEN technique was used to knock in human-derived mutation in MeCP2 

for modeling Rett syndrome caused by MeCP2 deficiency, and CRISPR-Cas9 approach was first achieved in deleting a non-neuronal gene, Dystrophin, in macaques, and the same method also achieved double-knockout of Ppar-γ and Rag1 in one-cell-stage macaque embryos. Notably, the p53 biallelic genes were knocked out simultaneously, resulting in the generation of homozygous mutant monkeys (Chen et al., 2016). The long duration of sexual maturation and gestation in macaques imposes a significant barrier to the development of genetically modified NHP models that require germline transmission, although the duration for generating F1 in macaques (5–6 years) could now be shortened to 2.5 years with testis xenografting technique that accelerated sperm maturation (Liu et al., 2016a, Liu et al., 2016b).

Rich resources of NHPs and strong interest in using NHPs do not imply that ethical standards for NHP experimentation could be more relaxed in China (Zhang et al., 2014). The China Brain Project aims to establish nationwide ethical regulations for NHP experimentation that are compatible with international standards and to promote public dissemination of the awareness that NHP research is indispensable for developing effective therapies for human diseases, particularly brain disorders, and for advancing our knowledge of the evolution and function of the human brain. Given the declining NHP research in the Europe and the U.S., national brain projects in Asian countries also shoulder the responsibility of sustaining the tradition of NHP research in neuroscience and training the new generation of primate neurobiologists.

Brain-Inspired Computation

Neuroscience has focused on detailed studies of neural coding, dynamics, and circuits, while machine learning tends to pursue brute-force optimization of a cost function, often using simple and relatively uniform initial architecture (Marblestone et al., 2016). Recent progress in AI and deep learning in particular has shown its capability for handling cognitive tasks in restricted specific fields. Despite the fact that AI systems (such as AlphaGo) outperform human beings in certain tasks (Silver et al., 2016), they still suffer from the lack of generalizability and the ability to transfer learned knowledge from one task (domain) to another. Also, labor-intensive labeled data are needed to tune the huge number of parameters of such deep learning models. Another key problem is the high computational (energy) cost and high throughput data for training and running these AI systems. The human brain is currently the only truly general intelligent system in nature, capable of coping with different cognitive functions with extremely low energy consumption. Learning from information processing mechanisms of the brain is clearly a promising way forward in building stronger and more general machine intelligence.

Although we are far from completely understanding how the brain really works, current findings from neuroscience could potentially impact AI research from several perspectives. From the structure perspective, the morphology of different types of neurons, the stabilization and pruning of connections during development and learning, the layered architecture of the neocortex, feedforward and feedback connections within and among brain regions, and the motifs of brain building blocks at multiple levels offer new insights into the architectural design of artificial neural networks. From the mechanism perspective, spike information encoding and decoding, different types of spiking neurons with distinct functions, multiple synaptic types and plasticity mechanisms, rules for conversion from short- to long-term memories, and integration of information processing at different levels (neurons, micro-circuits, brain regions) bring potential operational principles for designing efficient computational models and algorithms for general AI. From the behavior perspective, observations and analysis on how different cognitive functions are coordinated and integrated by the brain will bring inspiration and evaluation criteria for intelligent systems that are brain-like in their cognitive performance.

The brain was shaped by evolution into a highly energy efficient system. Its structure and underlying mechanisms may provide inspiration for the design of future computing infrastructures. Unlike traditional computation, neural systems process information in a way of total binding of computing and storage. Recent efforts in designing neuromorphic chips have focused on creating brain-inspired chips that are highly energy efficient (Tuma et al., 2016), by implementing a few microscopic-level principles of neural circuits, such as nonlinear neuronal properties of integrate-and-fire, spike timing-dependent plasticity (STDP), and integrated computation and storage. Higher-level architecture could also be considered in the future as building blocks for chips to simulate organized structures of cortical columns, brain regions, and neural pathways that connect multiple brain regions, in order to achieve efficiency and high throughput in information processing.

The China Brain Project aims at better understanding of mechanisms and principles of the brain at multiple levels and is expected to promote deep and close collaboration between neuroscientists and AI researchers. Cognitive computational models and brain-inspired chips will be the primary focus of the intelligence wing. At the level of computational models, artificial neural network algorithms with more biological plausible learning mechanism will be explored. At the network architectural level, typical human cognitive behavior will be modeled through introduction of brain-like domains and sub-domains within the network that are coordinated, integrated, and modifiable through learning. The goal is to simulate in principle the mechanisms and architecture of the brain at multiple levels to meet the grand challenge to making a general AI that is capable of multitasking, learning, and self-adapting.

Machines with Human Intelligence

Achievements of AI in past decades, including recent deep learning models, have been inspired in part by neuroscience. Recent development mainly depends on single optimization principles of objectives, such as minimizing classification errors. This led to the formation of rich internal representations and powerful algorithmic capabilities in multilayer and recurrent networks (LeCun et al., 2015). In the past five years, deep learning has enjoyed great success in solving a variety of problems such as speech recognition, image recognition and classification, and natural language processing. In speech recognition, an accuracy as high as 95% has been reported by IBM and Microsoft in human telephone call conversation tasks, greatly exceeding the level that had plateaued for a decade. In computer vision, the deep learning network also surpasses human performance in ImageNet classification challenge in locating and recognizing hundreds kinds of objects. In natural langue processing, an LSTM-based sequence-to-sequence model for machine translation almost reaches the human interpreter level. Machines could even annotate an image using natural language, after being trained with millions of image-text pairs collected from the internet. In all the above examples, structured architectures are used, including dedicated systems for attention, recursion, and various forms of short- and long-term memory storage.

However, models driven by massive training data meet great challenges for more open and ill-defined tasks like natural language understanding, human dialog system, general visual information retrieval, and robotic adaptation to complex environment. Aligning to brain mechanisms, AI systems are expected to exhibit stronger intelligence with less training data or even with unsupervised learning. Furthermore, they are expected to process and integrate multimodal information and handle multiple tasks in parallel. We have witnessed many new types of networks with more dedicated object functions, for example varied across layers and over time, to handle these challenges. The new advance in adversarial networks, where the cost function is provided by another network, allows gradient-based training of generative models (Goodfellow et al., 2014). Such a heterogeneously optimized system, enabled by a series of interacting cost functions, makes the learning data-efficient and precisely targeted to the need of the intelligence, and it is one of the future directions for machine intelligence.

Another essential issue in developing machine intelligence is to build an AI platform that could interact with human and local environment efficiently, in which both human and machine are in the loop of problem solving. Cognitive robotics could serve as an integrating platform of this kind to integrate many efforts of brain-inspired research. Traditional robotics research focuses on control theory and mathematical optimizations. The models work well in structured environment and restricted tasks (e.g., robotic arm in factories), but cannot navigate correctly even in not-so-complex environments. Substantial improvement has been made in robotics for the coordination and integration of multi-sensory inputs and execution mechanisms with more information and flexibility. But for cognitive robotics, much more could be learned from the network structure, operating principles, and circuit mechanisms of sensorimotor transformation in the brain, including multi-sensory integration, decision-making, motor planning, and motor coordination and execution, in a manner that the operation could be learned and underlying circuits self-modified by experience.

The China Brain Project will focus its efforts on developing cognitive robotics as a platform for integrating brain-inspired computational models and devices. The goal is to build intelligent robots that are highly interactive with humans and properly reactive in uncertain environments, with the skills for solving various problems that can grow through interactive learning, and the ability to transfer and generalize knowledge acquired from different tasks—even to share learned knowledge with other robots. The interface between human and machine is essential; the robot needs not only to understand what the human means and to respond smartly, but also to learn to understand the human intention and the way the human makes decisions. Thus, a useful milestone for cognitive robotics is to build a robot that acquires behaviorally equivalent capability of empathy and theory of mind, a cognitive hallmark of humans and a few primate species.

Concluding Remarks

Future breakthroughs in basic and applied neuroscience depend upon not only fundamental discoveries and technological development in individual laboratories, but also collaborative efforts by large teams of researchers from diverse disciplines. As exemplified by recent advances in the frontier of physics and astronomy, the key to success often lies in the effective organization of the team work, which calls for consensus among the participants for a framework of equitable sharing of duties and credits. This is particularly important for team work that requires individual scientists to devote their major research efforts and resources into the project. Furthermore, the weights our research institutions place on collaborative work versus independent accomplishment in the evaluation of a scientist’s achievements are becoming increasingly critical, especially for young scientists in the process of establishing their own research careers. In biological sciences, our institutions have yet to adopt a system of evaluation, e.g., for tenure review, that is conducive to team work.

Complete understanding of the structure and function of the human brain is an attractive but remote goal of neuroscience. However, the limited understanding of the brain that neuroscience has achieved is already useful for addressing some urgent problems our society is facing. For example, identification of early molecular or functional markers for AD could be accomplished prior to our full understanding of the pathogenesis of AD. The China Brain Project aspires to achieve a balance between basic and applied neuroscience, in which some research scientists are capable of pursuing their interest in exploring the secrets of the brain, while others may apply what we know already for preventing and curing brain disorders and for developing brain-inspired intelligence technology.

• 上一条：2016医学科技管理创新发展专题研讨会会议通知
• 下一条：中共中央　国务院关于完善产权保护制度依法保护产权的意见