Research On Efficient And Robust Multi-spike Learning And Its Application

Due to the development of deep learning,artificial neural networks have prospered in many fields.However,compared with the high efficiency of the brain,the current deep learning methods consume a lot of computing resources,rely heavily on data,and lack biological plausiblity.Therefore,the spiking neural network that simulates the basic mechanism of the brain and provides more powerful computing power and more biological plausible has attracted more and more attention.Spiking neural network uses spikes for information transmission and processing.spikes are also considered a way of processing information with lower energy consumption.However,how to efficiently process and learn spikes is still a challenging task.At present,some multi-spike learning algorithms have been proposed to process and learn the information carried in the spike train.Moreover,most of existing learning algorithms are inefficient with complex neuron dynamics and learning processes being involved,which limits the application capabilities of spiking neural networks.This thesis focuses on the subject of multi-spike learning algorithm,and proposes a simplified neuron model and two efficient learning algorithms for the above difficulties.The main contributions of this thesis are as follows:1.A simplified spike neuron model is proposed by impulse function to simulate the effects of synaptic input and firing output on membrane potential,which can effectively process spike signals.Considering the highly complex nonlinear dynamics of other neuron models,the model is more concise and efficient.In addition,an event-driven simulation method was introduced,in which calculations are driven by spikes,thereby improving the efficiency of processing and learning.2.Based on the simplified neuron model,two improved efficient and robust multispike learning algorithms are proposed,namely,the efficient multi-spike learning algorithm EML and the learning algorithm EMLC,which only depends on the current response state of the neuron.Combining the characteristics of the simplified neuron model,Some tedious calculations and derivation processes can be ignored without affecting the results,making the calculations more efficient,and can further provide references for the application and development of neuromorphic systems.3.The performance of the proposed learning algorithm has been evaluated on some typical tasks,including efficient processing of spike pattern,multi-class classification experiments,robustness experiments,etc.,and compared with other baseline methods,showing advanced performance of the proposed algorithm.At the same time,some explorations are made on the EML algorithm,such as derivative evaluation,evaluation of the multi-classification ability under different spike coding schemes and the detection ability of features from background activity.