AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Measuring the (Un)Faithfulness of Concept-Based Explanations

Shubham Kumar, Narendra Ahuja

CVPR 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Self-explaining and visual explanation networks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
randomize concept importances in the explanation (ablation)SURF surrogate · Caltech-101 ResNet-50SURF MAE: 0.00 -> 2.70Hurt
randomize concept importances in the explanation (ablation)ICE-Eval · Caltech-101 ResNet-50Top-1 agreement: 100 -> 100No clear effect
randomize concept importances in the explanation (ablation)C-SHAP-Eval (CEL) · Caltech-101 ResNet-50SURF MAE: 1.97 -> 1.97No clear effect
randomize both CAVs and importances (ablation)SURF surrogate · Caltech-101 ResNet-50Top-1 agreement: 100 -> 1.3Hurt
randomize both CAVs and importances (ablation)C-SHAP-Eval (CEL) · Caltech-101 ResNet-50Top-1 agreement: 9.02 -> 97.6Mixed
use HU-MCD instead of MCD for object classificationU-CBEM · ResNet-50 object classificationSURF MAE: 2.60 -> 1.97Improved
use SAE instead of other U-CBEMsU-CBEM · ResNet-50 object classificationSURF EMD: 0.384 -> 0.195Improved
increase the number of conceptsCDISCO, CRAFT, C-SHAP, ICE · object classificationSURF MAE/EMDNo clear effect
increase the number of conceptsMCD and HU-MCD · object classificationSURF MAE/EMDImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy