Online Weight Approximation for Continual Learning

Alessandro Betti, Gabriele Ciravegna, Marco Gori, Kevin Mottin, Frederic Precioso

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.

Keywords: online function approximation, continual learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Continual Learning primarily focuses on studying learning scenarios that challenge a learner’s capacity to adapt to new problems, while reducing the loss of previously acquired knowledge. This work addresses challenges arising when training a deep neural network across numerous tasks. We propose an Online Weight Approximation scheme to model the dynamics of the weights of such a model across different tasks. We show that this represents a viable approach for tackling the problem of catastrophic forgetting both in domain-incremental and class-incremental learning problems, provided that the task identities can be estimated. Empirical experiments under several configurations demonstrate the effectiveness and superiority of this approach also when compared with a powerful replay strategy.

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: 6374