Online Supervised Learning Based On Gradient Descent For Spiking Neural Networks

Understanding how the brain learns to compute functions reliably,efficiently is a fundamental challenge in neuroscience.It is widely believed that such learning is implemented by synaptic plasticity mechanisms that change synaptic weights as a function of joint pre-and postsynaptic activity.Spiking neural network is a kind of computational model which is closer to the biological nervous system.It has the ability to process spatio-temporal data and transmit biological information by sending spikes at accurate times by neurons.It overcomes the limitation of artificial neural network on computing power,and opens a new path to the development of models with memory ability and fast adaptation ability.From the results of neuroscience,the neurons in the brain can respond quickly to a single spike.The intensity of synaptic connection between neurons will immediately intensify or weaken when the spike is released,which means,the biological neural system in most cases uses the online way to change the synaptic weights.In order to get closer to real neurons and deal with numerous information,a more biologically interpretable online method has been noticed more by scholars.In this thesis,the different learning rules of spiking neural network are analyzed and compared.Its main idea and training characteristics are briefly elaborated,the application of gradient descent method in artificial neural network and different supervised learning algorithms extended to spiking neural network are introduced.Then,considering the characteristics of gradient descent method and online learning mode,this thesis proposes an online supervised learning algorithm based on gradient descent,which is suitable for the multi-spike training of spiking neural networks.The online training of the network model and the real-time adjustment of the synaptic weights are realized by minimizing the error function between the desired and actual output spike trains,and the online supervised learning is achieved with the spike as the processing object.The process of spiking neural networks online learning is different from the offline way in training the network.Once the output layer neurons with spike excitation,it will trigger the real-time adjustment of the network parameters.Namely in the course of training a sample,the network weights will be adjusted several times according to the spike of the output layer neuron during the training process.Finally,the proposed online supervised learning algorithm is applied to the learning task of spike sequence.Under different network parameters,the online and offline methods are compared.The results show that online learning is superior to offline way in learning cycle and learning accuracy.The online supervised learning algorithm of spiking neural network is further applied to the problem of nonlinear pattern classification,and use the One-pass method which the sample data is entered only once.The experimental results on the Iris data sets and Wisconsin breast cancer data sets show that the online supervised learning algorithm has high classification accuracy of the training set and the test set in once iterative learning process.The learning efficiency of pattern classification problem is improved.