Advancing Few-shot Continual Learning via Selective Knowledge Transfer

Cuong N. Nguyen, Quang Pham, Cuong V. Nguyen, Xiaoli Li, Savitha Ramasamy

Keywords: continual learning, transfer learning

Abstract:

Continual learning with large language models (LLMs) is a promising and challenging research that greatly impacts many applications. Existing solutions treat previous tasks equally, making them vulnerable to task interference, lacking scalability with a large number of tasks, and oblivious to the intrinsic relationships among tasks. This work presents selective knowledge transfer (SKT), a novel
and principled framework for continual learning with LLMs. SKT aims to maximize positive knowledge transfer while systematically minimizing the effects of irrelevant information from dissimilar tasks. To this end, SKT first assesses the degree of interference between the current and previous tasks and then selectively aggregates the tasks that maximize knowledge transfer for continual
training. In addition, we integrate SKT into the current state-of-the-art continual language learning algorithm, Progressive Prompts, to introduce Log-evidence Progressive Prompts (LePP), which facilitate knowledge transfer between tasks. Comprehensive evaluations on challenging few-shot continual learning benchmarks demonstrate that LePP can surpass existing baselines for continual learning with LLMs with minimal overhead. Our extensive ablation studies reveal that SKT can discover useful task correlations without any prior knowledge, many of which align with human evaluations. Code will be published upon acceptance.

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.

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: 2362