studying a reinforcement learning technique for the spiking neural network

The article describes a reinforcement learning technique for the spiking neural network. Spiking artificial neural networks, also called neural networks of the third-generation are a special class of the artificial neural networks, in which the signal is a set of impulses (spikes) of the same amplitude and duration. There is a need to use special techniques to learn these networks. Development and research of such techniques is a relevant task now. The article discusses a reinforcement learning technique for the neural network using a particular type of connections between neurons, called a hedonistic synapse. The hedonistic synapse partly reflects the properties of its real biological counterpart. It alters its activity depending on the intrinsic activity in previous times and on the reinforcement in the system. The study was conducted to define an impact of such synapse parameters on the learning efficiency of the spiking neural network. The article shows a mathematical model of the hedonic synapse and gives a description of its parameters and variables. It describes the role of these variables and parameters in reinforcement learning the neural network. To assess the learning quality are used such criteria as an error index and AUC. The article describes their application to have a generalized assessment of learning.

It also describes a generation technique and gives examples of test data used in the study. A technique to calculate the reinforcement for the network is described depending on the output signals. The impact of the variables and parameters of the hedonic synapse model on the learning quality for the neural network is analysed. The article reviews the features of the Bayesian optimization technique and describes its using to optimize quality of classification. It provides an indepth analysis of the optimization results and comes to the conclusion, as a result of the study, that such a learning technique can be used to classify the multivariate time series. The article defines potential trends for further research in the field concerned that is an optimization of the neural network structure and its applications.