Search papers, labs, and topics across Lattice.
This paper introduces Decomposition-based Multimodal Interaction Learning (DMIL), a novel framework that systematically analyzes and learns from dynamic, sample-specific interactions in multimodal data. By employing a variational decomposition architecture, DMIL isolates redundant, unique, and synergistic information across modalities, addressing the limitations of conventional modality ensemble and joint learning approaches. Extensive experiments show that DMIL significantly outperforms existing methods by effectively adapting to the unique interaction patterns present in diverse multimodal tasks.
Conventional multimodal learning fails to capture dynamic interactions, but DMIL adapts to sample-specific nuances, leading to superior performance across tasks.
Multimodal learning hinges on capturing redundant, unique, and synergistic information across modalities, which collectively constitute multimodal interactions. A critical yet underexplored challenge is that these implicit interactions vary dynamically across samples. In this work, we present the first systematic, information-theoretic analysis highlighting why learning these dynamic, sample-specific interactions is critical for effective multimodal learning. Our analysis further reveals deficits in conventional paradigms at learning these distinct interaction types: modality ensemble approaches struggle to capture synergy, while joint learning paradigms often under-utilize redundant information. This highlights the need for an approach that can adaptively learn from different interaction types on a per-sample basis. To this end, we propose Decomposition-based Multimodal Interaction Learning (DMIL), a novel paradigm that explicitly models and learns from sample-specific interactions. First, we design a variational decomposition architecture to isolate the constituent interaction components. Second, we employ a new learning strategy that leverages these explicit interaction components in a fine-tuning process to achieve comprehensive interaction learning. Extensive experiments across diverse tasks and architectures demonstrate that DMIL consistently achieves superior performance by adapting to holistic sample-specific interactions. Our framework is flexible and broadly applicable, establishing an interaction-centric paradigm for multimodal learning. The code is available at https://github.com/GeWu-Lab/DMIL.