Pruning One More Token is Enough: Leveraging Latency-Workload Non-Linearities for Vision Transformers on the Edge

TL;DR

This paper introduces a latency-aware token pruning method for vision transformers on edge devices, leveraging non-linear latency-workload relationships to improve efficiency without retraining.

Contribution

It identifies key factors affecting ViT latency, determines an optimal token pruning schedule based on non-linear relationships, and proposes a training-free pruning method that outperforms existing approaches.

Findings

Reduces latency by 9-26% compared to other methods.

Maintains high accuracy (78.6%-84.5%) at similar latency levels.

Outperforms Token Merging in efficiency and accuracy.

Abstract

This paper investigates how to efficiently deploy vision transformers on edge devices for small workloads. Recent methods reduce the latency of transformer neural networks by removing or merging tokens, with small accuracy degradation. However, these methods are not designed with edge device deployment in mind: they do not leverage information about the latency-workload trends to improve efficiency. We address this shortcoming in our work. First, we identify factors that affect ViT latency-workload relationships. Second, we determine token pruning schedule by leveraging non-linear latency-workload relationships. Third, we demonstrate a training-free, token pruning method utilizing this schedule. We show other methods may increase latency by 2-30%, while we reduce latency by 9-26%. For similar latency (within 5.2% or 7ms) across devices we achieve 78.6%-84.5% ImageNet1K accuracy, while the state-of-the-art, Token Merging, achieves 45.8%-85.4%.