Research And Application Of Deep Pulse-coupled Neural Networks

Deep learning has achieved great success in many fields,but artificial general intelligence is still a distant goal.Researchers generally believe that neuromorphic computing is an effective way to achieve this goal.Brain-like neural networks simulate the information processing mechanisms of the brain and are a hot research topic in neuromorphic computing.Learnable brain-like neural networks,represented by spiking neural networks(SNNs),perform tasks such as feature extraction and pattern recognition by adjusting synaptic weights.However,due to their use of simple LIF(Leaky Integrate-and-Fire)neurons which lack complex spatiotemporal dynamics,their expressive power is limited.Brain-like neural networks for low-level vision,represented by pulse-coupled neural networks(PCNNs),can simulate more complex neural activities in the visual cortex and thus more biologically plausible.However,their complex spatiotemporal dynamics make it difficult to design effective learning algorithms for them,limiting their application to complex tasks.In this thesis,we combine SNN learning algorithms with PCNN models to propose deep pulse-coupled neural networks(DPCNNs)and explore applications and properties of DPCNNs.The main contributions and innovations of this thesis are as follows:(1)Based on the working principles of SNN learning algorithms,this thesis classifies existing methods hierarchically.Based on this classification,we summarize the development and current status of SNN learning algorithms and summarize their advantages and limitations.Finally,we summarize and compare the performance of SNN learning algorithms on several public datasets.(2)This thesis constructs DPCNNs through convolution and combines them with spatiotemporal backpropagation learning algorithms to perform end-to-end image classification tasks.To improve the model performance,we introduce inter-channel coupling,allowing neurons in different channels to be coupled.To solve the problem of DPCNNs’ difficulty in convergence,we propose receptive field and time dependent batch normalization(RFTD-BN).We evaluate the performance of DPCNNs on four datasets: MNIST,N-MNIST,Fashion-MNIST,and CIFAR-10.The results show that the performance of DPCNNs are better than that of SNNs based on LIF.In addition,the VGG9-based DPCNN achieves 94.2% accuracy on CIFAR-10,which is the SOTA performance of this architecture.(3)This thesis further explores the characteristics of DPCNNs.The results show that inter-channel coupling can achieve communication and coordination between neurons,enhance the network’s expressive power,and thus effectively improve the model’s performance.Compared to deepening the network,introducing inter-channel coupling can achieve comparable or greater performance improvements while maintaining the same number of network layers.Compared to widening the network(increasing the number of convolution channel),introducing inter-channel coupling can achieve similar or even better performance using fewer neurons and synapses.Comparative experiments of RFTD-BN show that it can accelerate the convergence of DPCNNs and improve performance.Additionally,we investigate the impact of hyperparameters on DPCNNs and the robustness of DPCNNs.