Causal Concept Explanations for Deep Neural Models
CCW-Net trains neural models to make concept explanations causally account for the outputs they explain.
A concept explanation can appear correct without being necessary or sufficient for a model’s output, the authors warn. Their method uses interventional mediation analysis as a training objective, targeting necessity, sufficiency and independence of concept contributions. They also use that analysis to quantify how much of a model’s output is causally attributable to each concept explanation. The paper evaluates CCW-Net across aircraft control, driving and image classification and is accepted at NeurIPS 2026.
ArXiv · AI/CL/LG's note
A concept explanation can appear correct without being necessary or sufficient for a model’s output, the authors warn. Their method uses interventional mediation analysis as a training objective, targeting necessity, sufficiency and independence of concept contributions. They also use that analysis to quantify how much of a model’s output is causally attributable to each concept explanation. The paper evaluates CCW-Net across aircraft control, driving and image classification and is accepted at NeurIPS 2026.
ArXiv · AI/CL/LG's note
score 5