Dyn-Adapter: Towards Disentangled Representation for Efficient Visual Recognition

TL;DR

Dyn-Adapter introduces a dynamic, disentangled feature extraction method that significantly reduces FLOPs during inference while maintaining or improving accuracy in visual recognition tasks.

Contribution

It proposes a novel dynamic architecture with multi-level feature disentanglement and bidirectional sparsity to enhance PETL efficiency and effectiveness.

Findings

Reduces FLOPs by 50% during inference.

Maintains or improves recognition accuracy.

Demonstrates effectiveness across diverse datasets and backbones.

Abstract

Parameter-efficient transfer learning (PETL) is a promising task, aiming to adapt the large-scale pre-trained model to downstream tasks with a relatively modest cost. However, current PETL methods struggle in compressing computational complexity and bear a heavy inference burden due to the complete forward process. This paper presents an efficient visual recognition paradigm, called Dynamic Adapter (Dyn-Adapter), that boosts PETL efficiency by subtly disentangling features in multiple levels. Our approach is simple: first, we devise a dynamic architecture with balanced early heads for multi-level feature extraction, along with adaptive training strategy. Second, we introduce a bidirectional sparsity strategy driven by the pursuit of powerful generalization ability. These qualities enable us to fine-tune efficiently and effectively: we reduce FLOPs during inference by 50%, while maintaining or even yielding higher recognition accuracy. Extensive experiments on diverse datasets and pretrained backbones demonstrate the potential of Dyn-Adapter serving as a general efficiency booster for PETL in vision recognition tasks.