Training-Free Approach of Convolutional Neural Networks with Astrocyte-Inspired Architectures

Abstract

This work introduces the Artificial Neuron-Astrocyte Network (ANAN), a novel approach that incorporates artificial astrocytes into pre-trained Convolutional Neural Networks (CNN) to enhance performance without requiring additional training. By dynamically modulating synaptic weights based on neuronal activity, astrocytes allow the network to adapt to input data efficiently. The proposed approach only requires optimizing four parameters, instead of the millions that are typically required in CNN fine-tuning. This offers a resource-saving alternative to traditional fine-tuning methods. Experimental results demonstrate statistically significant improvements in performance employing four different datasets, including one balanced and one imbalanced biomedical dataset, as well as two balanced ones encompassing natural images. Results in different application domains highlight the potential of astrocytes to optimize network performance without the need of going through traditional training cycles.