Towards The Computational Model And Key Technologies On Heterogeneous Brain-Inspired Computing Platform

Nowadays,with the explosive development of brain-inspired computing platforms inspired by the information processing paradigm of the brain,the AI era is coming.There majorly exist two kinds of implementations: deep learning processor and neuromorphic chip.The former imitates the hierarchical processing and learning feature of the brain,and optimizes the von Neumann processor-based architecture to provide efficient deep learning engines;however,its performance on spatio-temporal pattern representation and processing generalization is still far less satisfactory.The latter imitates the spatio-tempral correlation of the brain,and uses distributed parallelism and integrated memory/computation to emulate the brain and provides efficient platform for artificial general intelligence;however,its fundamental mechanism and application are unclear and in the infant stage.To this end,this paper proposes heterogeneous brain-inspired computing platform that deeply integrates the deep learning,neural dynamics and neuromorphic architecture,which promises the exploration of ideal brain-inspired computing architecture based on cross modeling and simultaneously supports the AI application and neuroscience research.The main contents and results of this dissertation are summarized as follows.1.For the deep neural networks,an online learning algorithm with discrete state transition(DST)in weight space is proposed to break the memory constraints on braininspired computing platforms.It greatly saves the memory cost in the training phase by eliminating the storage of high-precision hidden weights,and enables the flexible modification of the number of discrete states according to the memory capacity of different hardware platforms to make full use of the hardware resources.This provides guidance rule for the design of future brain-inspired computing system with on-chip learning ability.2.For the dynamic neural networks,this paper introduces the computing and learning framework of spiking neural networks(SNNs),and investigates the leaky integrate and fire(LIF)continuous dynamic networks,including :(a)design a hardware implementation of the recurrent LIF networks from neuroscience for pattern learning;(b)propose a hierarchical chunking of sequential memory(HCSM)model and the coding method for its parameters,demonstrate the dynamic memory trace and reveal the relationship between the chunking mechanism and synaptic plasticity;(c)re-design the continuous attractor neural networks(CANNs)for target tracking on brain-inspired computing platform,and analyze its sensitivity and stability.3.The ideal brain-inspired computing platform based on memristive networks is designed.In particular,the circuit simulation system is designed,the recurrent LIF networks for pattern learning and sequential memory are implemented,and the influence of the memristor dynamic range and variation are quantitatively analyzed.Furthermore,an estimation methodology for the energy consumption of memristor modulation is proposed,and a low-power-first modulation strategy enables the reduction of energy consumption when programming the memristive networks.4.Tianjic series chips for heterogeneous brain-inspired computing based on current digital circuits are designed,as well as the software and hardware platforms.Under the Tianjic architecture,the data integralization and structure mapping methods for various neural networks are proposed.TianjicⅠis aimed at the functional integration of deep neural networks,neural dynamics and neuromorphic architecture.An intelligent unmanned bycicle based on deep neural networks and dynamic networks for balance control,object detection and real-time tracking is designed based on the TianjicⅠsystem.TianjicⅡ is aimed at improving the chip performance by systematically optimizing the network scale,model complexity and scalability,and running speed.A heterogeneous communication interface between deep neural networks and dynamic networks is proposed,which promises the cross modeling of these two kinds of networks to explore future brain-inspired computing architecture.By designing a row-wise pipeline mapping method and a ‘float convolution & ternary fullconnection’ training scheme,TianjicⅡ is able to support the real-time running of large scale neural networks with high performance.