Feature-level Interaction Explanations in Multimodal Transformers

Multimodal AI systems often hide how different data types work together: Current explainability methods for AI systems that combine multiple types of information (like text and images) typically only highlight important parts within each type separately, rather than revealing how they interact to reach decisions—a gap this research addresses.

The research identifies three distinct types of cross-modal relationships: The paper distinguishes between synergistic pairs (where two modalities provide complementary evidence together), redundant pairs (where one modality can substitute for another), and unique contributions from single modalities, enabling more nuanced understanding of model reasoning.

New metrics quantify how modalities interact: The Shapley Interaction Index (SII) and redundancy-gap score provide measurable ways to identify which specific feature pairs across modalities are most important for predictions, moving beyond general importance rankings to causal relevance.

The identified interactions actually matter for predictions: When researchers removed feature pairs ranked as most important by their method, model performance degraded more than when removing randomly selected pairs of equal size, demonstrating that the identified interactions are genuinely used by the model rather than spurious correlations.

Better explanations could improve AI trustworthiness in multimodal applications: By clarifying exactly how different data sources (text, images, audio) jointly support decisions in real-world systems like content moderation or recommendation engines, this approach makes multimodal AI systems more interpretable and potentially easier to audit for bias or errors.