Relaxing Representation Alignment with Knowledge Preservation for Multi-Modal Continual Learning
Keywords: Continual learning, Multi-modal
Abstract: In continual learning, developing robust representations that adapt to new distributions or classes while retaining prior knowledge is crucial. While most traditional approaches focus on single-modality data, multi-modal learning offers significant advantages by leveraging diverse sensory inputs, akin to human perception. However, transitioning to multi-modal continual learning introduces additional challenges as the model needs to effectively combine new information from different modalities while avoiding catastrophic forgetting. In this work, we propose a relaxed cross-modality representation alignment loss and utilize a dual-learner framework to preserve the relation between previously learned representations. We validate our framework using several multi-modal datasets that encompass various types of input modalities. Results show that we consistently outperform baseline continual learning methods in both class incremental and domain incremental learning scenarios. Further analysis highlights the effectiveness of our solution in preserving prior knowledge while incorporating new information.
Supplementary Material: zip
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: 9009
Loading