Conflict-Aware Representation Fusion: A Theoretical and Practical Framework for Learning from Disagreeing Modalities

El-Din et al. (2024) address modality ambiguity via Dempster–Shafer theory, but their framework largely resolves conflict in a late-fusion, feature-aggregation sense. Building on the "generate theories from conflict" heuristic, this research would design a new learning architecture that explicitly models when and how different modalities disagree—e.g., when image and text representations send conflicting signals about a class label. Instead of simply aggregating, the model would learn "conflict signatures" and adaptively allocate trust to each modality based on context, possibly using meta-learning or adversarial training. Theoretical work would formalize the value of conflict and its resolution in representation learning, extending current Dempster-Shafer approaches. Practically, this could yield models that are both more interpretable (by surfacing when/why conflicts occur) and more robust in multimodal, real-world settings.

References:

1. An adaptive and late multifusion framework in contextual representation based on evidential deep learning and Dempster-Shafer theory. Doaa Mohey El-Din, A. Hassanien, Ehab E. Hassanien (2024). Knowledge and Information Systems.