Study On The Supervised Learning And Image Recognition In Recurrent Spiking Neural Networks

In recent years, with the rapid development of spiking neural networks and its study on learning algorithms. More and more result showed that the working principle of the spike neurons by accepting, firing spike to transmit information and to realize works functions closer to the real biological neurons. The spiking neural network base on the precise timing spike has become the effective tools for information processing in the nervous system. Base on the characteristics of precise timing spike sequence encoding information, the purpose of supervised learning of the spiking neural network is to make the neural network can output spike sequence by adjusting the synaptic weights, and the spike sequence can expressing particular information. Base on spike precise timing characteristics of the spiking neural network has a more powerful storage and computer capacity, it can simulate all kinds of neuronal information and arbitrary continuous function, and it is very suitable for the brain information processing problems.First, most of the learning algorithms of the spiking neural network to learn from in the error back propagation algorithm of artificial neural network. On this basis, given the gradient descent learning rule, and supervised learning algorithm is proposed based on local recurrent spiking neural network with feedback. For precise timing spike encoding, first define a multi-spike error function. Base on gradient descent construction neurons synaptic weights learning rules between output layer and hidden layer, hidden layer and the recursive layer, to achieve the automatic adjustment synaptic weights of recurrent spiking neural network. At present most of the learning algorithm based on gradient descent only learning single spike, and analyzes the reasons of them. Construct a spiking response neuron model suitable to this algorithm, the output layer neurons can fire multiple spikes, and so can be used for classification problems. Improve the recurrent spiking neural network in the application of the ability to solve complex problems.Then, by simulating a series of spike sequence learning, to verified the proposed supervised learning of recurrent spiking neural network by a variety of condition of spike sequence learning ability. In the process of spike sequence learning, through the single spike sequence of learning, to verify the input spike sequence for a given, the recurrent spiking neural network can train out an desired spike sequence. Then, through the multi-spike sequence learning demonstrated the ability to learn a random input spike sequence model to map multiple target output spike sequence. The results show that the algorithm can achieve spike sequence complex spatial-temporal model of learning.Finally, in the learning process of multitasking spike sequence shows the algorithm to different input modes can be done in parallel multi-mode learning ability. The experimental results show that supervised learning algorithm of recurrent spiking neural network can effectively complete different spike sequence learning task, and can achieve learning of spike sequence of complex spatial-temporal model. At the same time, the algorithm is applied to the part image classification problem of the Label Me image datasets, and verified the algorithm of image classification problem solving ability.