HyperAdapter: Generating Adapters for Pre-Trained Model-Based Continual Learning

Qizhe Zhang, Ruichuan An, Bocheng Zou, Zhi Zhang, Shanghang Zhang

Keywords: hypernetworks, adapter tuning, class-incremental learning

TL;DR: We propose a novel hypernetwork-based framework to generate task-oriented adapters for pre-trained model-based continual learning.

Abstract: Humans excel at leveraging past experiences to learn new skills, while artificial neural networks suffer from the phenomenon of catastrophic forgetting during sequential learning. Efforts have been made to alleviate forgetting by introducing a rehearsal buffer into the model, but this way is impractical in real-world scenarios with data privacy. Recently, pre-trained model-based continual learning methods have provided new insights into addressing this issue by effectively utilizing the powerful representational capabilities of pre-trained models to avoid catastrophic forgetting without a rehearsal buffer. In this work, we propose a novel pre-trained model-based continual learning framework, HyperAdapter, which utilizes a hypernetwork to generate adapters based on the current input, adapting the pre-trained model to the corresponding task. This paradigm requires fewer additional parameters as the number of tasks increases, which is a critical advantage for scaling to long sequences continual learning. Unlike methods that partition task-related knowledge into relatively independent subspaces, it promotes positive knowledge transfer across tasks. Comprehensive experiments across various datasets demonstrate that HyperAdapter consistently outperforms all existing methods and even exceeds the upper bounds of multi-task learning, establishing a new state-of-the-art for pre-trained model-based continual learning. Our code will be released.

Primary Area: transfer learning, meta learning, and lifelong learning

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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.

Submission Number: 9828