Token-Label Alignment for Vision Transformers

TL;DR

This paper identifies a token fluctuation issue in vision transformers during data mixing and proposes a token-label alignment method to improve training accuracy and model performance across various vision tasks.

Contribution

The paper introduces TL-Align, a novel token-label alignment technique that maintains accurate token-label correspondence in vision transformers during data mixing.

Findings

Improves ViT performance on image classification tasks

Enhances semantic segmentation and object detection accuracy

Achieves better transfer learning results

Abstract

Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphoria16/TL-Align.