Divide and Orthogonalize: Efficient Continual Learning with Local Model Space Projection
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: Continual learning, Low-rank approximation, Optimization
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose a local model space projection (LMSP) based efficient continual learning framework.
Abstract: Continual learning (CL) attracts more and more research interests recently since it enables a learning model's ability to continuously learn new tasks without forgetting the previously learned knowledge. However, existing CL methods require either an extensive amount of resources for computing gradient projections or memorizing lots of old tasks as the candidates for related old tasks selection. Thus, a low-complexity CL approach is necessary for the model deployment on huge data. In this paper, we propose a local model space projection (LMSP) based efficient continual learning framework, which helps to not only reduce the complexity of computation, but also extend to several local model tasks to increase the candidate pool with strong correlations. We also theoretically show that the proposed LMSP approach enables backward knowledge transfer, which is a highly desirable feature in CL. Extensive experiments on several public datasets demonstrate the efficiency of our approach.
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: 4563
Loading