Is multitask learning all you need in continual learning?

ICLR 2025 Conference Submission11449 Authors

Keywords: lifelong learning, multitask learning, continual learning

TL;DR: We challenge the assumption that multitask learning is optimal for continual learning and we show that it sometimes leads to suboptimal online performance

Abstract:

Continual Learning solutions often treat multitask learning as an upper-bound of what the learning process can achieve.
This is a natural assumption, given that this objective directly addresses the catastrophic forgetting problem, which has been a central focus in early works. However, depending on the nature of the distributional shift in the data, the multi-task solution is not always optimal for the broader continual learning problem. In this work, we draw on principles from online learning to formalize the limitations of multitask objectives, especially when viewed through the lens of cumulative loss, which also serves as an indicator of forward transfer.
We provide empirical evidence on when multi-task solutions are suboptimal, and argue that continual learning solutions should not and do not have to adhere to this assumption. Moreover, we argue for the utility of estimating the distributional drift as the data is being received and show preliminary results of how this could be exploited by a simple replay based method to move beyond the multitask solution.

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: 11449