Tiny Adapters for Vision Transformers

Imad Eddine MAROUF, Enzo Tartaglione, Stéphane Lathuilière

Keywords: Vision Transformers, Parameter Efficient Training, Adapters

TL;DR: Tiny Adapters for Vision Transformers

Abstract:

Vision Transformers (ViTs) have become one of the dominant architectures in computer vision and pretrained ViT models are commonly adapted to new tasks via fine-tuning of its parameters. Recent works in NLP proposed a variety of parameter-efficient transfer learning methods such as adapters to avoid the prohibitive storage cost of fine-tuning.

In this work, we start from the observation that adapters perform poorly when the dimension of adapters is small and we propose a training algorithm that addresses this issue. We start from large adapters which can be trained easily and iteratively reduce the size of every adapter. We introduce a scoring function that can compare neuron importance across layers and consequently allow automatic estimation of the hidden dimension of every adapter. Our method outperforms existing approaches in terms of the trade-off between accuracy and trained parameters across domain adaptation benchmarks. We will release our code publicly upon acceptance.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning