Research On Reservoir Learning Algorithms Based On Temporal Encoding

In the process of modeling the biological nervous system,it is found that the communication between the biological neurons is accomplished by nerve impulse.As the most suitable network model in biological modeling,spiking neural network is better for complex space-time information.Spiking neuron apply the temporal encoding rule code neural information,the temporal encoding mechanism more in line with the simulation of the biological nervous system,contains more effective information.Reservoir only trained the output layer synaptic weights,although it solves some shortcomings of the traditional spiking recurrent neural network,and has been well applied in solving large-scale problems and dealing with long-term problems,usually neglects the dynamic characteristics of the reservoir.The hidden layer of the reservoir structure is a large and irregular intelligent biological model,and the learning of the hidden layer is more consistent with the overall modeling of the biological nervous system.First,this paper analyzes the network structure and network training rules of the reservoir model,as well as online and offline learning algorithms.In the temporal encoding reservoir model,the Remote Supervised Method(ReSuMe)learning algorithm is used to adjust the output layer synaptic weights.The input layer synaptic weights and the hidden synaptic weights are generated in the process of network structure initialization and no longer change.Although the learning performance is higher than that of the traditional network,the learning accuracy of the spike train is improved,but the dynamic performance of the reservoir is often neglected.Second,the historical activity of synaptic neurons before and after the actual learning will have a certain impact on synaptic strength and synaptic effect.The correlation between the neuron spike train of the presynaptic neurons can not only cause long term potentiation and long term depression,and the synaptic enhancement or inhibition will retrograde and propagate rapidly the neurons of the dorsal neurons.Based on the bi-directional effect of electrical activity on the synapses this paper proposes a bidirectional modification learning algorithm to study the synaptic weights of reservoir layer based on time-coding reservoir structure.Finally,a spike train pattern learning analysis was performed.In the temporal encoding reservoir model,the learning of the hidden layer synaptic weights is based on the bidirectional modification learning algorithm.The output layer adopts the ReSuMelearning algorithm,which is based on the learning rate and the reverse propagation rate and reservoir size on the network structure.At the end of the experiment,a comparison of the learning performance of different network learning rule with hidden learning algorithm and no hidden learning algorithm is carried out.The experimental results show that the learning function of the hidden synapses based on the bidirectional modification can improve the learning performance of the network structure.