AQ4SViT: An Automated Quantization Framework with Search Gating Policy for Compressing Spiking Vision Transformers
Researchers have developed AQ4SViT, an automated framework designed to compress Spiking Vision Transformers (SViTs) for use in resource-constrained embedded AI systems. This new framework addresses the scalability issues of manual quantization by employing a search gating policy that leverages membrane potential drift as a performance proxy. AQ4SViT offers two search variants: Greedy search, which is faster but may find local optima, and Beam search, which is slower but aims for global optima. Experiments show significant memory savings, with AQ4SViT-Greedy achieving up to 82.5% reduction and AQ4SViT-Beam reaching up to 90%, all while maintaining high accuracy. AI
IMPACT This framework could enable the deployment of more efficient AI models on edge devices, expanding the capabilities of embedded AI systems.