Echo State Networks with Artificial Astrocytes and Hebbian Connections

Abstract

For the last few decades, the neuroscientific research has highlighted the importance of astrocytes, a type of glial cells, in the information processing capabilities. By dynamic bidirectional communication with neurons, astrocytes regulate their excitability through a variety of mechanisms. Traditional artificial neural networks (ANNs) are connectionist models that describe how information passes throughout layer of neurons abstracting from low-level mechanisms. However, very little research has addressed artificial astrocytes and their incorporation into ANNs. In this paper, we present an echo state network (ESN) extended with astrocytes which influence the neurons by fixed or Hebbian connections. By systematic analysis we investigate their role on five classification tasks and show that they can outperform the standard ESN without astrocytes. Although the model with fixed astrocytic weights yields from none to little improvement, the model with Hebbian weights from astrocytes to neurons is significantly superior.