Artificial neural network with dynamic synapse model

Abstract

The purpose of this study is to develop and investigate a new short-term memory model based on an artificial neural network without short-term memory effect and a dynamic short-term memory model with astrocytic modulation. Methods. The artificial neural network is represented by a classical convolutional neural network that does not have short-term memory. Short-term memory is modeled in our hybrid model using the Tsodyks-Markram model, which is a system of third-order ordinary differential equations. Astrocyte dynamics is modeled by a mean field model of gliotransmitter concentration. Results. A new hybrid short-term memory model was developed and investigated using a convolutional neural network and a dynamic synapse model for an image recognition problem. Graphs of dependence of accuracy and error on the number of epochs for the presented model are given. The sensitivity metric of image recognition d-prime has been introduced. The developed model was compared with the recurrent neural network and the configuration of the new model without taking into account astrocytic modulation. A comparative table has been constructed showing the best recognition accuracy for the introduced model. Conclusion. As a result of the study, the possibility of combining an artificial neural network and a dynamic model that expands its functionality is shown. Comparison of the proposed model with short-term memory using a convolutional neural network and a dynamic synapse model with astrocytic modulation with a recurrent network showed the effectiveness of the proposed approach in simulating short-term memory.

Keywords

• short-term memory
• machine learning
• convolutional neural network
• astrocytic regulation