The Brain Serotonergic System Offers Insights into Novel Dropout Algorithms for Image Classification Tasks
Deep Neural Networks (DNNs) with a large number of trainable parameters are highly effective in classification tasks in machine learning. Nevertheless, overfitting is a persistend issue with these networks. Random dropout has become a standard regularization strategy in deep learning, but these algorithms typically disregard the geometry of artificial neurons and their relative spatial location. We hypothesize that the brain serotonergic fibers (axons) may suggest dropout-like mechanisms in biological neural networks due to their pervasive existence, stochastic structure, and potential to produce new paths throughout the course of an individuals’ lifetime. Therefore, mathematical models of the structure and dynamics of serotonergic fibers may contribute to the development of dropout algorithms in DNNs. In this study, we describe an experimental method uniquely tags individual serotonergic fibers in the adult mouse brain. Since the trajectories of fibers can be described as pathways of anomalous diffusion processes, we then examined the performance of a dropout method based on superdiffusive fractional Brownian motion (FBM) in Convolutional Neural Networks (CNNs). The findings indicate that serotonergic fibers have the capacity to execute a dropout-like process in brain tissue, hence promoting neuroplasticity.