Research On Design Of Brain-Inspired Visual Neural Network And Its FPGA Implementation

Perceiving and processing of visual information has long been one of the hot issues in brain-inspired computing.On the one hand,it will help to reveal the efficient information processing mechanism of visual system in human brain.On the other hand,it will help people solve the problems of information understanding and processing in brain-inspired computing.However,the research on visual information processing combined with brain-inspired computing has not yet been perfect.In this thesis,we discussed biological vision theory and the computational model for simulating visual perception mechanism combined with spiking neural network.A brain-inspired neural network model was proposed for simulating the function of ventral visual pathway.Moreover,the high-efficiency hardware implementation of the model based on FPGA platform was studied.The main contributions and research results are listed as follows:First of all,the characteristics of the basic framework of spiking neural network model were introduced,including single neuron models,synapse models and the types of spike train encoding,etc.Based on the simulation and analysis of a typical spiking neural network and its online learning experiments,we found that neuron models and synapse models with different complexity,different spike train encoding types and synaptic plasticity learning rules may have an important impact on the performance of the spiking neural network.Secondly,the structure and function of the human brain visual pathway was introduced.Then we explored the structure of the ventral visual pathway which was closely related to object recognition,hierarchical biological vision theory and visual information processing mechanism.Combined with bio-plausible neural mechanism unsupervised STDP learning rules and spiking neural network,we designed a multilayer brain-inspired neural network model with the rules of biological cognitive behavior to simulate the feature extraction,feature learning and cognitive decision in ventral visual pathway.In addition,we realized efficient learning and recognition of handwritten numbers based on the proposed model.Finally,we designed and constructed a neural network hardware simulation platform which used FPGA chip as the core of hardware implementation.And the braininspired neural network visual pathway model was implemented based on the platform.The construction of each part in the framework was given.Then,the efficient hardware implementation of the self-regulation LIF neuron and event-driven triplet-based STDP learning rule were introduced in detail.A series of efficient hardware implementation methods greatly reduced the resource consumption.And the results and functions were verified.Meanwhile,the theoretical analysis of the corresponding hardware resource consumption was given with different network scales,which provided a possible way for realizing a larger plasticity neural network in future work.