New method interprets how Transformers fuse diverse data sources

Transformers have driven AI advances, but interpreting how they handle inputs from different sources remains a challenge. The authors categorize attention into homogenous (same source) and heterogeneous (different sources, like co-attention). Heterogeneous structures are key for multi-modal models and complex functions, yet their black-box nature poses problems for research and policy.

The paper proposes a generic interpretation method for such models, combining semantic and logical analysis. Experiments on representative models validate the approach. This work provides a foundation for understanding multi-modal AI, crucial for building trust and meeting regulatory requirements as these models become more prevalent.