6533b871fe1ef96bd12d15e9

RESEARCH PRODUCT

Modeling of neuron-astrocyte interaction : application to signal and image processing

subject

description

The introduction of the tripartite synapse and the discovery of calcium wave propagation motivated our research to explore the potential of astrocytes as active components in brain circuits. For decades, astrocytes have been considered passive cells whose primary function is metabolic and structural support to neurons; however, recent physiological measurements suggest that astrocytes modulate neural communication, strengthen synaptic efficacy, enhance synchronization, and promote homeostasis. Inspired by these biological functions, this research aimed to implement astrocytes in artificial spiking networks for deep learning applications. First, we modeled the biological interaction between neurons and astrocytes ‒ from the tripartite connection to neuron-astrocyte networks. The results suggest that astrocytic connectivity and heterogeneity determine whether astrocytes would improve or impair neural activities. Then, we designed a spiking neuron-astrocyte network architecture for image recognition using simplified biologically inspired models. We trained the network to recognize features and classify handwritten digits using spike-timing-dependent plasticity and an unsupervised learning algorithm. Here, the astrocyte-mediated networks displayed advantages over neuron networks alone, such as faster learning, higher accuracy, and improved bias-variance tradeoff. One of the challenges in the study is the extended duration needed for training. Therefore, we proposed a dendritic abstraction supporting dendrite-specific computations for faster learning. We analyzed the signal propagation along a pyramidal neuron dendritic tree and determined that a single neuron performs more complex computations previously attributed only to neural networks by following a multilayer-multiplexer scheme. We proposed that dendritic abstractions connected in this scheme could promote faster synaptic updates independent of backpropagating signals from the soma. This research is one of the first attempts to implement astrocytes as computational elements in artificial networks.