A method (ARM) that corrects modality imbalance in multimodal learning by valuing each modality's contribution with mutual information rather than only boosting weak modalities. A mutual-information-based valuation metric (MIV) uses mutual information and conditional mutual information between each unimodal feature and the fused feature to estimate each modality's marginal contribution and the joint lower-bound contribution per sample. These values drive dynamic feature-fusion weighting, a balanced min-max loss that maximizes joint contribution while narrowing per-modality gaps, and dynamic resampling of low-contribution samples, reinforcing weak modalities without degrading dominant ones (avoiding 'modality forgetting').
Proposes ARM (Asymmetric Reinforcing against Multi-modal representation bias), a method for imbalanced multimodal learning where a dominant modality suppresses weaker ones. Introduces a mutual-inform…