Design And Analysis Of Cortical-microcircuit-like Neuromorphic System

Hierarchical columnar microcircuits are the basic units of the cortex.Each columnar microcircuit is a complex neural network and completes the expression of cortical functional heterogeneity with similar structures.With the special organization and morphological features of modularity,the cortex reflects some characteristics of human brain such as multiple cell types,dynamic structures,local and global connections and so on,thus realizing the separation and integration of brain functions.If the neuromorphic system is built in a bio-realistic manner and with the similar network composition,the morphological structure and the function of cortical columnar microcircuits,its computing and application performance can be improved in many ways.However,current neuromorphic architectures often ignore the properties of the cortex,thus limiting the improvement of the computational performance of neuromorphic systems.To address the issues of scalability,simulation capability and application performance of neuromorphic systems,this thesis proposes a corticalmicrocircuit-like neuromorphic system based on the structural and functional characteristics of the cortex.The research is conducted on the structural design of the neuromorphic system and also the Spiking Neural Network(SNN)models and algorithms.The achieved results are summarized in the following:(1)A cortical microcircuit-like neuromorphic architecture is proposed by combining the structural characteristics of Field Programmable Gate Array(FPGA)with the morphological and functional characteristics of the cortex.The cortical columnar microcircuits are considered as "normalized" building blocks of the human brain in the architecture,and the neuromorphic system is constructed by combining multiple modules according to the real biological form,which improves the scalability and communication capability of the system.(2)Based on the structure and function of microcircuits,combinable neuromorphic models and algorithms are designed to support the simulation of multiple morphological microcircuits and the development of application functions.Based on the diversity of cortical neurons,on-chip configurable modular neuron models are designed and implemented to support the simulation of different morphological neurons,which can save up to 63.2% of resource consumption compared with the commonly used spiking neuron models.The limitation of the neuron model types that can be simulated by neuromorphic systems is also broken with these models.The evaluation of various spike encoding algorithms is completed,and the suitability scoring method for algorithm selection was proposed to improve the encoding efficiency of SNN.(3)A cortical-microcircuit-like neuromorphic system is built.The internal and external connections and topology of the system are designed based on the morphological characteristics of microcircuits.The system supports the implementation of neural network at the scale of three million neurons with cortical-like structure and high scalability.According to the system architecture,a three-dimensional multi-level cluster-mesh network-on-chip structure and routing algorithms are proposed,which can improve the injection rate of the saturation point and reduce the resource consumption,so as to archive lower the communication delay.(4)To improve the implementation efficiency of cortical microcircuit models on neuromorphic systems,a series of resource-optimized design methods such as nonlinear computational modules and an automatic fixed-point design algorithm with adjustable errors are proposed to address the problems of the nonlinear operations and fixed-point design of the models.The neuron and synapse models implemented by applying the proposed methods can save up to 78.9% of hardware resources consumption,and also simplify the model implementation process in the form of computational modules.(5)Based on the modular neuron models,learning algorithms and architecture of the proposed neuromorphic system,an SNN model of cerebellar microcircuits is constructed using FPGA.The constructed model follows the neuronal ratio,multitemporal scale plasticity and neuronal morphological diversity of biological neural networks,which successfully reproduces the spiking activities and various physiological phenomena.The model is mapped to the neuromorphic system in the form of microcircuits,and an adaptive control task of the robotic arm is accomplished.This thesis addresses several key issues in building high-performance neuromorphic systems,and investigates them in terms of models,algorithms,implementation optimization methods and system architectures based on the morphological properties of cortical microcircuits,which provides ideas for the implementation of neuromorphic computing and its applications.