Interior interpretability with attention rollout: contraction and propagation profiles in Transformers
The paper treats attention rollout as an internal propagation diagnostic, not as proof of causal attribution.
Biccari, Huang, and Zuazua define “interior interpretability” for tabular Transformers by studying how attention-mediated operators compose through layers. They use Doeblin-Dobrushin contraction theory to show when a rollout operator is close to a rank-one stochastic matrix, giving structure to its propagation profile. In metabolomic age-prediction Transformers, measured rollout contraction grows stronger with depth, and trained models differ from randomly initialized ones. The authors say their experiments do not establish the predictive relevance of individual rollout-ranked variables. ArXiv · AI/CL/LG's note
Biccari, Huang, and Zuazua define “interior interpretability” for tabular Transformers by studying how attention-mediated operators compose through layers. They use Doeblin-Dobrushin contraction theory to show when a rollout operator is close to a rank-one stochastic matrix, giving structure to its propagation profile. In metabolomic age-prediction Transformers, measured rollout contraction grows stronger with depth, and trained models differ from randomly initialized ones. The authors say their experiments do not establish the predictive relevance of individual rollout-ranked variables. ArXiv · AI/CL/LG's note
score 4