MMPareto: Innocent Uni-modal Assistance for Enhanced Multi-modal Learning
Keywords: multi-modal learning, imbalanced multi-modal learning, gradient integration
Abstract: Multi-modal learning methods with targeted uni-modal constraints have exhibited their superior efficacy in alleviating the imbalanced multi-modal learning problem, where most models cannot jointly utilize all modalities well, limiting their performance. However, in this paper, we first identify that there are gradient conflict between multi-modal and uni-modal learning objectives, potentially misleading the optimization of shared uni-modal encoders. The necessity of diminishing conflict during gradient integration naturally accords with the idea of Pareto methods, which could provide a gradient that benefits all objectives. Unfortunately, conventional Pareto method surprisingly fails in the context of multi-modal scenarios. We further theoretically analyze this counterintuitive phenomenon and attribute it to the priority of Pareto method for multi-modal gradient with small magnitude, weakening model generalization. To this end, we propose MMPareto algorithm, which could ensure a direction that is common to all learning objectives while preserving magnitude with guarantees for generalization, providing innocent uni-modal assistance for primary multi-modal learning. Finally, empirical results across multiple dataset with different modalities indicate our superior method performance. The proposed method is also expected to facilitate multi-task cases with a clear discrepancy in task difficulty, demonstrating its scalability.
Primary Area: representation learning for computer vision, audio, language, and other modalities
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/2024/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: 5275
Loading