Improving Performance of Parsimonious Spiking Neural Networks Used for Few-Shot Image Recognition by Embedding the Function of Astrocyte in Its Structure

Spiking neural networks are known as a powerful framework, capable of solving tasks that conventional neural networks are capable of solving, while their dynamics are closer to biological neurons. In other words, spiking neural networks imitate the temporal dynamics of biological neurons, although the role of astrocytes has not often been considered in most models. Inspired by the influence of astrocytes and the concept of the tripartite synapse, we constructed a spiking neural network for learning and incorporated astrocytes to provide feedback to neurons for regulating synaptic strength. Two types of neuron-astrocyte connections were implemented. Using a few-shot learning framework for handwritten digit classification, we demonstrated that neuron-astrocyte interactions improved both classification accuracy and convergence speed of neurons to labels, even with limited training data. Since we employed a few-shot learning approach, we adopted a ten-way ten-shot setup with five query samples per class. Additionally, the network consisted of 80 excitatory and 80 inhibitory neurons. Using this configuration, we achieved an accuracy of 71.95% for the neuronal network, 72.3% for the first configuration where each astrocyte was placed at the synapse between an excitatory and inhibitory neuron, and 72.95% for the second configuration where each astrocyte was placed between two pairs of excitatory and inhibitory neurons. These results highlight the potential of biologically inspired mechanisms to enhance the efficiency and generalization of neuromorphic systems.