Towards Energy Efficient Spiking Neural Networks: a Homeostatic Metabolic Pruning Framework
Abstract
Sparse structural learning is crucial for deploying Spiking Neural Networks (SNNs) on resource-constrained neuromorphic hardware. However, most existing sparse structural learning methods for SNNs, largely inherited from Artificial Neural Networks (ANNs), focus on static magnitude to reduce parameters while overlooking spiking activity. As a result, the remaining neurons may fire excessively, undermining energy savings and fundamentally misaligned with the intrinsic spatiotemporal dynamics of SNNs. Inspired by GABA conductance-based synapses, we propose a bio-inspired inhibitory mechanism for SNNs that leverages accumulated firing traces to selectively prune hyperactive redundancy. We further propose a Homeostatic Finetuning strategy that recalibrates BatchNorm biases and reweights gradients based on historical neuronal activity, to compensate for the distribution shift introduced by removing inhibitory connections. Our method naturally evolves a distinct inverted pyramid topology, which preserves information-dense convolutional layers while aggressively compressing parameter-heavy fully connected layers, leading to a better storage–efficiency trade-off than conventional magnitude-based pruning. Experimental results on CIFAR-10, DVS-CIFAR10, and ImageNet datasets demonstrate that the proposed topology achieves competitive accuracy and energy trade-offs among existing methods.