Research On The Algorithms And Applications Of The Spiking Neural Networks

In recent years,as the development of the artificial intelligence,more and more intelligent devices and service make our life convenient.As one of the important research fields in the artificial intelligence,neuromorphic computing attempts to make the computer possess the intelligence like the human being by simulating the human brain cognitive principles and neuron's operating mechanisms.According to the research results of the neuroscientists,the information is processed and transmitted by spikes in the biological neuron system.To simulate this mechanism,the spiking neural network is proposed.It is more biological plausible and computational more powerful than traditional neural networks.Many achievements have been made in both theory and application fields based on the spiking neural network.The research on the spiking learning algorithm and its applications can promote both the improvement of the artificial intelligence and neurosynaptic chip,which has been regarded as a promising method to overcome the disadvantage of the traditional von Neumann architecture.Supervised learning is an important principle both in the neuromorphic computing and artificial intelligence.It trains neural networks by target driven mechanism,and is important to the knowledge accumulation process of the cognitive learning and pattern recognition.Then,we study the supervised learning algorithms of the spiking neural network and their various applications on pattern recognition in this thesis,containing the following aspects:(1)We propose a new supervised learning algorithm for the single layer spiking neural networks,and apply it to the classification tasks.Inspired by the selective attention mechanism of the primate visual system,our algorithm selects only the target spike time as attention areas,and ignores voltage states of the untarget ones.This reduces the processing information in the training process.Besides,our algorithm employs a cost function based on the voltage difference instead of the traditional precise firing time distance.Adopting these strategies,this algorithm solves the inefficient training problem in the single layer spiking neural networks,and its training efficiency is at least four times faster than traditional methods.(2)We propose a new supervised learning algorithm for the multi-layer spiking neural networks,and apply it to the classification tasks.In this algorithm,the computational error is propagated to previous layers by the presynaptic spike jitter instead of traditional back propagation methods.Besides,the normalized parameter in the error function,and the selective attention mechanism applied in this algorithm make it avoid computing redundancy information.The mathematical principle of both the feedforward calculation and the feedback weight modification are analyzed in this thesis.The proposed algorithm solves the inefficient training problem in the multi-layer spiking neural networks.(3)We propose an efficient recognition method based on the spiking neural networks to identify the user's validity for the smart device.In this method,the dynamic parameter adjustment mechanism is applied to the Perceptron-Based Spiking Neuron Learning Rule(PBSNLR),which makes its learning parameter relative to the firing time difference and the voltage difference.This mechanism solves the inefficient and parameter sensitive problem of the PBSNLR.After which,this improved method is applied to solve the user authentication problem in the smart device.(4)We propose an efficient recognition system for the human action recognition in the videos.In this system,the particle swarm optimization method is applied to the traditional remote supervised method(ReSuMe)of spiking neural networks.By which,the remote supervised method of spiking neural networks can adapt learning parameters automatically according to different input data.After which,a recognition system based on this improved algorithm is devised to the human action application in videos.