Frequency-Aware Token Reduction for Efficient Vision Transformer

TL;DR

This paper introduces a frequency-aware token reduction method for Vision Transformers that enhances computational efficiency and performance by selectively preserving high-frequency tokens and aggregating low-frequency tokens, addressing issues like rank collapsing and over-smoothing.

Contribution

It proposes a novel frequency-aware token reduction strategy that considers frequency characteristics to improve efficiency and accuracy in Vision Transformers, surpassing existing methods.

Findings

Significantly improves accuracy while reducing computational cost.

Effectively mitigates rank collapsing and over-smoothing phenomena.

Provides analysis of previous methods' implicit frequency characteristics.

Abstract

Vision Transformers have demonstrated exceptional performance across various computer vision tasks, yet their quadratic computational complexity concerning token length remains a significant challenge. To address this, token reduction methods have been widely explored. However, existing approaches often overlook the frequency characteristics of self-attention, such as rank collapsing and over-smoothing phenomenon. In this paper, we propose a frequency-aware token reduction strategy that improves computational efficiency while preserving performance by mitigating rank collapsing. Our method partitions tokens into high-frequency tokens and low-frequency tokens. high-frequency tokens are selectively preserved, while low-frequency tokens are aggregated into a compact direct current token to retain essential low-frequency components. Through extensive experiments and analysis, we demonstrate that our approach significantly improves accuracy while reducing computational overhead and mitigating rank collapsing and over smoothing. Furthermore, we analyze the previous methods, shedding light on their implicit frequency characteristics and limitations.