Feature-level Interaction Explanations in Multimodal Transformers

2026-03-04

arXiv.org

AI Summary

This work addresses the challenge of uncovering how synergistic and redundant interactions between cross-modal feature pairs jointly support decision-making in multimodal Transformers—a relationship inadequately captured by existing methods. The authors propose a novel approach that introduces a structured mixture-of-experts layer atop frozen token/patch sequences from pretrained encoders, enabling the first explicit modeling and quantification of multimodal synergy and redundancy at the feature level. They devise a pairwise interaction probing mechanism based on Shapley interaction indices and redundancy gap scores, complemented by attribution analysis, top-K% masking, and Monte Carlo probing for interpretability. Evaluated on MMIMDb, ENRICO, and MMHS150K benchmarks, the method yields more focused and interaction-specific importance patterns, with pairwise masking experiments confirming the causal relevance of the identified interactions.

Abstract

Multimodal Transformers often produce predictions without clarifying how different modalities jointly support a decision. Most existing multimodal explainable AI (MXAI) methods extend unimodal saliency to multimodal backbones, highlighting important tokens or patches within each modality, but they rarely pinpoint which cross-modal feature pairs provide complementary evidence (synergy) or serve as reliable backups (redundancy). We present Feature-level I2MoE (FL-I2MoE), a structured Mixture-of-Experts layer that operates directly on token/patch sequences from frozen pretrained encoders and explicitly separates unique, synergistic, and redundant evidence at the feature level. We further develop an expert-wise explanation pipeline that combines attribution with top-K% masking to assess faithfulness, and we introduce Monte Carlo interaction probes to quantify pairwise behavior: the Shapley Interaction Index (SII) to score synergistic pairs and a redundancy-gap score to capture substitutable (redundant) pairs. Across three benchmarks (MMIMDb, ENRICO, and MMHS150K), FL-I2MoE yields more interactionspecific and concentrated importance patterns than a dense Transformer with the same encoders. Finally, pair-level masking shows that removing pairs ranked by SII or redundancy-gap degrades performance more than masking randomly chosen pairs under the same budget, supporting that the identified interactions are causally relevant.

Problem

• Multimodal Transformers
• feature-level interaction
• synergy
• redundancy
• explainable AI

Innovation

• Feature-level Interaction
• Mixture-of-Experts
• Shapley Interaction Index
• Multimodal Explainability
• Redundancy-gap Score