Attention Meets Post-hoc Interpretability: A Mathematical Perspective
This work addresses the debate on using attention as explanations in AI, providing a mathematical perspective that is incremental but clarifies differences between explanation types.
The paper mathematically analyzes a simple attention-based architecture to compare attention weights with post-hoc interpretability methods, finding that post-hoc methods capture more useful insights despite their limitations.
Attention-based architectures, in particular transformers, are at the heart of a technological revolution. Interestingly, in addition to helping obtain state-of-the-art results on a wide range of applications, the attention mechanism intrinsically provides meaningful insights on the internal behavior of the model. Can these insights be used as explanations? Debate rages on. In this paper, we mathematically study a simple attention-based architecture and pinpoint the differences between post-hoc and attention-based explanations. We show that they provide quite different results, and that, despite their limitations, post-hoc methods are capable of capturing more useful insights than merely examining the attention weights.