Research On FMRI Classification Based On Spiking Deep Belief Networks

Motivated by recent advances in neurosciences, neural information is encoded in the brain through the precisely timed spike trains, not only through the neural firing rate. Encoding and processing information with precisely timed spike trains, spiking neural network is a new generation of computational model of neural network. Compared with previous traditional artificial neural network based on firing rate of spikes, spiking neural network not only has better biological interpretation, but also has a powerful computing potential. There are many researchers to study unremittingly to tap the enormous potential of spiking neural networks since spiking neural network was proposed. Simulating multilayer structure of brain, deep learning of neural networks extract features from lower layers to higher layers to improve process ability of complex spatial- temporal information. Therefore, combining spiking neural network and deep learning effectively to construct algorithm of deep learning in spiking neural network, it will become an important research direction in the field of neural networks to solve complex problems.Firstly, encoding strategies of neural information were analyzed in this article. According to different methods of information encoding, neural encoding strategies can be divided into spiking frequency and precise timing. Spiking frequency encoding encodes and processes information on the basis of frequency of spikes in neurons, but precise timing encoding uses the time of spikes to encode the neural information. Besides an encoding method of functional Magnetic Resonance Imaging(fMRI) was constructed according to the domains of brain characteristic of fMRI.Secondly, according the structural analysis of spiking deep belief network, the unsupervised pre-training algorithm and supervised training algorithm are proposed combining the rule of Triplet-based Spike Timing Dependent Plasticity(TSTDP). Applying TSTDP into Contrastive Divergence(CD) algorithm, a better unsupervised algorithm of spiking deep belief network is obtained. Combining TSTDP rule and a supervised algorithm named Remote Supervised Method(ReSuMe), a new supervised learning algorithm of spiking deep belief network is obtained. The improved supervised algorithm has both higher accuracy and less iteration.Finally, the constructed spiking deep belief network and algorithms are applied into actual problem of classification of fMRI. Selecting the standard data set Attention Deficit & Hyperactive Disorder(ADHD) as the set of classification, it is used to verify performance and problem solving ability of this network. Experimental results shows that the spiking deep belief network in this article has better classification results than some other methods.