Research Of The Critical Problems In Multi-level Brain-inspired Computing

Brain-inspired computing refers to novel computing paradigms that are inspired by the neural information processing in human brain,which contains the aspects of hardware implementations,computing architectures and models.On one hand,brain-inspired computing is meaningful for understanding the neural information processing principles.On the other hand,it is helpful for the supercomputing systems with stronger computing power and lower power consumption.Neuroscience researches have revealed different levels of neural information processing at different scales of the human brain have various types of effects on cognitive behaviors.The current studies still stay in the status of spiking neural networks and deep learning,while few researches of brain-inspired computing are based on different levels of working mechanisms in human brain.In this dissertation,models,architectures and implementation methods are studied for the cross-level brain-inspired computing based on the working mechanisms of human brain.By describing the computational characteristics and digital implementation methodology of neurons,neural networks and neural nuclei,we gradually realize multi-level brain-inspired computing and large-scale brain-inspired computing system to study the cognition functions with self-learning capability across different brain areas,including object recognition,motor control,multimodal learning,decision making and the mechanisms underlying the movement disorders.The research contents include:(1)A brain-inspired computing framework is proposed,which integrates different levels of neural information processing mechanisms.It uses the digital neuromorphic technology to link the levels of neuron and nuclei.A brain-inspired computing system is realized with self-learning capability,high computational power,and brain-inspired multi-level mechanisms.(2)A brain-inspired computing methodology at the neuron level is presented.A cost-efficient multi-compartment neuron model CMN is presented,which reproduces the neural dynamic behaviors.A high-performance implementation technique of the CMN model is presented,which can reduce hardware resource cost and increase computational speed by 35.14%.A novel scalable computing architecture and routing algorithm are proposed,which contains biological plausibility,high computational capability and self-learning ability.(3)A brain-inspired computing methodology based on the mechanisms at the network level is presented.An implementation method of different kinds of synaptic plasticity algorithms is designed.The central pattern generator is realized and the primary bipedal gait control of the robot is realized.Neural network model and hardware architecture are proposed,which realizes the large-scale brain-inspired computing with the neural information mechanisms at the network level.(4)A brain-inspired computing methodology based on the mechanisms at the nucleus level is presented.A cost-efficient multi-nucleus collaborative neural network model is proposed,and novel router architecture MSIP and routing algorithm IMP using synaptic events are proposed,which can support brain-inspired multi-nucleus collaborative mechanisms.(5)A novel scalable hierarchical heterogeneous multi-core non-von Neumann architecture is presented according to the proposed brain-inspired computing methodologies.Large-scale brain-inspired computing system Bi Co SS is implemented.The system power density is 2.8k times larger than that of GPU.Kinds of applications are explored,which contains the object recognition with unsupervised learning,decision making with reinforcement learning,motor control with cerebellar supervised learning,multimodal learning based on hippocampus-prefrontal cortex mechanism and the mechanism investigation of the movement disorders.It reveals that Bi Co SS has the advantage of the fusion of the multiple kinds of cognition with self-learning mechanism.