Title: PrimeSVT: A Prioritized Compression Strategy for Automated, Memory-Aware Pruning in Spiking Vision Transformers

Abstract: The substantial model sizes of Spiking Vision Transformers (SViTs) continue to impede their deployment in embedded systems, underscoring the critical necessity for effective model compression. While current state-of-the-art techniques rely on unstructured pruning to reduce SViT footprint, this approach demands specialized hardware accelerators to exploit specific sparsity patterns for optimal efficiency. Furthermore, these methods typically require extensive manual effort to determine optimal pruning configurations for individual networks, rendering them inefficient and non-scalable.

To overcome these challenges, we introduce PrimeSVT, a novel framework designed for automated, memory-aware structured pruning of pre-trained SViT models. This approach maximizes inference efficiency while remaining compatible with standard computing architectures. PrimeSVT operates on a prioritized compression policy: it first ranks SViT layers by parameter count, then selects target layers based on their robustness to various pruning rates. The model is then compressed sequentially, layer by layer, from the largest to the smallest, while adhering to user-specified constraints regarding acceptable accuracy loss and memory reduction. Within each layer, PrimeSVT utilizes channel-wise filter pruning guided by L2-norm values to structurally eliminate insignificant weights.

Experimental evaluations demonstrate that PrimeSVT achieves a 26.68% reduction in memory usage through automated, single-shot pruning. Crucially, it maintains accuracy within a 3% margin of the original unpruned SViT (which achieved 73.3% accuracy), recording 70.3% without fine-tuning and 72.9% with fine-tuning. These results confirm that PrimeSVT facilitates the design automation of SViTs, thereby enabling their practical implementation in embedded environments.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC