ICON Decomposition: Multivariate Concept-Level Explanations of Deep Representations for Model Auditing
ICON tries to separate real model reliance from concept correlations inside a network layer.
The paper says existing concept audits can overread shortcuts because they test concepts one at a time. ICON decomposition measures each concept’s share of representation variance after accounting for the other concepts and the outcome. The authors report stronger recovery of known concept importance on synthetic data than seven baselines. In skin-lesion and brain-imaging models, they say it identifies the concepts the model actually relies on and leaves a measured remainder for variance not explained by the supplied concepts. ArXiv · AI/CL/LG's note
The paper says existing concept audits can overread shortcuts because they test concepts one at a time. ICON decomposition measures each concept’s share of representation variance after accounting for the other concepts and the outcome. The authors report stronger recovery of known concept importance on synthetic data than seven baselines. In skin-lesion and brain-imaging models, they say it identifies the concepts the model actually relies on and leaves a measured remainder for variance not explained by the supplied concepts. ArXiv · AI/CL/LG's note
score 4