Enhancing the astrocyte spiking deep neural networks and calcium based neural network model for classification

Abstract

The ongoing challenge in machine learning is to advance general and biological inspired artificial models of neural networks which are compatible with the spatial and temporal constraints of the brain. For instance, a new term has recently been emerged to describe the communication between two neurons and a single astrocyte, the other type of cells in the brain, called tripartite synapse. The communication between astrocytes and astrocytes in a network called astrocytic syncytium is explored. The calcium dynamics in the brain is also considered as a significant player in brain information processing. Therefore, the study proposes to generalize mathematical models for tripartite synapse and astrocytic syncytium to advance new Tripartite Synapse Model (TSM), Artificial Astrocytic Syncytium (AAS) model and Calcium Based Artificial Neural Network (caANN) to mimic the calcium dynamics in the brain. Moreover, the study utilizes the proposed models of TSM and AAS within the architecture of the deep neural networks such as convolutional neural networks and deep belief neural networks. The simulation results of incorporating the real astrocyte in spiking response model (RSM) and Recurrent-Simple Neural Network (RSNN) has shown that astrocyte increases the postsynaptic potential and in turn, improves the RSM. Besides, the simulations of TSM in two neuron models, leaky integrate and fire (LIF) and Izhikevich model, has shown that using TSM has changed the spiking behavior (rate and firing pattern) of these models. The simulation related to the AAS by probability distribution and K-L divergence comparison has shown that the gap junction channels can be opened by the higher probability astrocytes. Finally, the result of the simulations of TSM in CNN and DBN showed that incorporating astrocytes’ dynamics, properties and roles in deep learning networks compete and sometimes outperform the standard architectures of deep neural networks in terms of training and validation accuracy.