Diverse via bounded Agreement: Geometric Regularization for Multimodal Fusion
The results suggest that explicitly regulating representation geometry is an effective complement to optimization balancing, and provide evidence that geometry-aware regularization can improve multimodal learning across diverse architectures and domains.
Zixuan Xia, Hao Wang, Peng-Cheng Weng et al.
· 1 citation