MorphSNN: Adaptive Graph Diffusion and Structural Plasticity for Spiking Neural Networks

A research team led by Yongsheng Huang has introduced MorphSNN, a groundbreaking framework designed to overcome a major bottleneck in Spiking Neural Networks (SNNs). While individual SNN neurons mimic biological dynamics, their overall architecture has remained stuck in static, hierarchical patterns, limiting their adaptability. MorphSNN directly addresses this by incorporating two core biological principles: a Graph Diffusion (GD) mechanism that enables efficient, undirected signal propagation alongside traditional feedforward paths, and a Spatio-Temporal Structural Plasticity (STSP) mechanism that allows the network's physical connections to rewire and reorganize dynamically in response to specific inputs.

This 'self-evolving' topology is a game-changer. Experiments show MorphSNN achieves state-of-the-art performance, notably reaching 83.35% accuracy on the challenging neuromorphic vision dataset N-Caltech101 using only 5 timesteps, indicating high efficiency. Beyond raw accuracy, the dynamic architecture itself acts as an intrinsic 'distribution fingerprint.' This means MorphSNN can perform superior Out-of-Distribution (OOD) detection—identifying data that is fundamentally different from its training set—without any auxiliary models or additional training, a significant step toward more robust and autonomous AI systems. The code is publicly available, inviting further development in neuromorphic computing.