AQ4SViT framework shrinks Spiking Vision Transformers by 90%
Automated quantization cuts memory while preserving accuracy within 1.5%
AQ4SViT is an automated quantization framework for compressing Spiking Vision Transformers (SViTs), using a search gating policy that leverages membrane potential drift as a performance proxy to quickly evaluate and select promising quantization candidates. It offers two search algorithm variants: Greedy search (up to 6.6x faster search time and up to 82.5% memory saving vs. state-of-the-art) and Beam search (up to 90% memory saving but 4.5x longer search time). Both maintain high accuracy within 1.5% of the original non-quantized models on ImageNet, advancing SViT deployment on embedded AI systems.
- AQ4SViT-Greedy accelerates quantization search by 6.6x while cutting memory by 82.5%
- AQ4SViT-Beam reduces memory footprint up to 90% at the cost of 4.5x longer search
- Both variants maintain accuracy within 1.5% of the original model on ImageNet
Why It Matters
Automated compression enables low-power spiking vision transformers to run on drones and IoT devices.