Research On The Membrane Potential Driven Learning Method For Spiking Neural Networks

The Spiking neural networks known as the "third-generation neural network" use spike sequences to simulate the information transfer process of biological neurons.Neurons combine temporal and spatial information on the basis of spike distribution and have stronger biomimetic properties.Traditional artificial neurons use rate coding,while the new generation of spiking neurons encodes external stimuli with time as a feature,carrying more information,which provides a powerful tool for analyzing the information processing process in the biological brain.However,it is still in its initial development stage.The improvement of the whole system and the efficient learning methods still need to be excavated.Therefore,this paper is very meaningful for the research of the supervised learning mechanism based on membrane potential driving in Spiking Neural Networks.The Spiking neural networks currently have two classic models of I&F and SRM in the research field.This paper focuses on the membrane potential driven supervised learning methods on the most commonly used LIF neuron models under the I&F model.This paper presents an efficient learning method for multi-spike output,and proposes a fast supervised learning algorithm based on the membrane potential of the EPs to further improve the learning efficiency of multi-spike output scenarios.In order to prove the general adaptation of the EPs idea,it is used to improve the classical PBSNLR algorithm.The main research contents are as follows:1.A multi-spike learning method based on gradient descent is proposed.The classic Tempotron learning algorithm is introduced,and the defect that it can only be applied to single spike models is pointed out.A multi-spike learning method based on gradient descent is proposed on the LIF multi-spike neuron model and successfully applied to multi-spike fire training.2.A method for calculating the membrane potential EPs(Extreme Points)of the membrane potential is proposed.Existing EMPD algorithms learn by monitoring stagnation points and breakpoints,but EMPD is only suitable for simple SRM models,with low universality and many monitoring points.This paper proposes a method based on the local membrane potential extremes to calculate the EPs on the most universal LIF multi-spike neuron model.3.EP-Tempotron and EP-PBSNLR algorithms based on membrane potential extreme points training are proposed.Using EPs to improve the multi-spike learning method based on gradient descent,EP-Tempotron was proposed to further improve the learning efficiency in multi-spike scenarios.In order to prove the universality of EPs idea,an offline learning algorithm PBSNLR based on spike perceptron neuron SPN was introduced.Combining EPs training mechanism,an EP-PBSNLR algorithm was proposed to train accurate feature vectors,which improved the unstable problem of PBSNLR at low selection rate.4.The learning performance of membrane potential driven and spike-driven algorithm is compared.And then the EP-Tempotron and EP-PBSNLR algorithms were successfully applied to XOR logic problems and optical character recognition problems.The practicability of the two algorithms was verified and compared.Finally,the results were analyzed in detail.