Research Of Spiking Supervised Algorithm And Application

As the third generation of neural network models,Spiking Neural Networks(SNNs)has attracted considerable interest in recent years.Inspired by discovering in neuroscience,SNNs considers the time of spike firing and has a more biological plasible neuron structure.Both of two characters make SNNs achieve stronger computational power compared to traditional neural networks.SNN is an important field of neuromorphic computing.For this reason,the study on SNNs will promote the improvement of the artificial intelligence.Meanwhile neurosynaptic chip,which is considered as non-von Neumann architecture and a promising innovation of computer science,is also based on SNNs.For the moment,the research on SNNs mainly focused on coding method,neuron model and learning algorithm.Among these tasks,supervised learning is a particularly important and common issue in SNNs and mechine learning at the same time.It aims to learn a pattern which could obtain desired result with corresponding input.And this thesis discussed supervised spiking learning algorithm in a primary way.More concretely,we proposed a new spiking neuron model--probablistic spike response model(PSRM).The model based on a conclusion which is deduced by representing the spike event train with an undirect graph model and assuming an energy function related to membrane potential.We also proposed a supervised learning algorithm for PSRM.Because of the continuity of PSRM,our algorithm is more scalable and can overcome issues commonly encountered by other traditional learning algorithms,such as over-adjusted problems and silent neurons problems.Besides,we use the proposed method to the multilabel classification problem.We designed a PSRM-based multi-layer supervised learning algorithm,and a label dependency analysis algorithm to optimize the results.The experiment proves the validity and rationality of the proposed algorithm.Aiming at the disadvantages of the proposed multi-label classification algorithm on the image,this thesis also proposed a specially method for image.The method change the first step in PSRM-based multilayer supervised learning algorithm to two procedures.Firstly,detecting the object windows in the image.Then solveing image multiple classifications.Finally,the effectiveness of the algorithm is demonstrated in experiment.