BioMotion-SNN: Spiking neural network modeling for visual motion processing.

Abstract

Neuroscience remains to be challenged by the decoding of the neural processes underlying biological motion perception. To address this, we propose BioMotion-SNN, a spiking neural network (SNN) framework inspired by the MT-MST pathways, designed to model the dynamic interactions between these brain regions. Grounded in biological experimental phenomena, BioMotion-SNN processes event-driven stimuli in a manner closely resembling real sensory inputs, setting it apart from conventional models reliant on static datasets and predefined labels. The framework incorporates contrastive self-supervised learning with a motion-perception contrastive loss function to enhance feature representation, while L₁-norm-based synaptic pruning mimics sparse biological connectivity by reducing redundant connections. Leveraging real electrophysiological data augmented through controlled transformations, BioMotion-SNN reduces the need for extensive biological data collection, enriches dataset diversity, and bridges the gap between experimental neuroscience and computational modeling. Achieving a classification accuracy of 93.00 %, the framework effectively captures complex motion patterns and establishes a novel paradigm for integrating computational modeling with empirical neuroscience. Our data/codes are available at https://github.com/BrainCogLab/biomotion_snn.