Boomda: Balanced Multi-objective Optimization for Multimodal Domain Adaptation

In the paper 'Boomda: Balanced Multi-objective Optimization for Multimodal Domain Adaptation', published on arXiv, the authors tackle the costly issue of manual annotation in multimodal learning by exploring unsupervised domain adaptation. They focus on heterogeneous multimodal domain adaptation, where different modalities experience varying domain shifts. By employing the information bottleneck method, they learn independent representations for each modality and utilize correlation alignment to match source and target domains. The problem is framed as a multi-objective task to ensure balanced domain alignment across modalities, simplifying it to a quadratic programming problem. This leads to an efficient algorithm that demonstrates effectiveness through extensive empirical results, marking a significant advancement in multimodal learning methodologies.