Feature-level Interaction Explanations in Multimodal Transformers

Researchers introduce FL-I2MoE, a new Mixture-of-Experts layer for multimodal Transformers that explicitly identifies synergistic and redundant cross-modal feature interactions. The method provides more interpretable explanations for how different data modalities contribute to AI decision-making compared to existing approaches.

• →FL-I2MoE separates unique, synergistic, and redundant evidence at the feature level in multimodal AI systems.
• →The method uses Shapley Interaction Index to score synergistic feature pairs and redundancy-gap scores for substitutable pairs.
• →Testing across three benchmarks shows FL-I2MoE produces more concentrated importance patterns than dense Transformers.
• →Pair-level masking experiments confirm that identified interactions are causally relevant to model performance.
• →The research addresses a key limitation in current multimodal explainable AI methods that focus on individual modalities rather than cross-modal interactions.