Research map / Label-efficient and robust learning / Recognition and long-tail
Post hoc Explanations may be Ineffective for Detecting Unknown Spurious Correlation
Julius Adebayo, Michael Muelly, Harold Abelson, Been Kim
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 changed | Where | Measured | Effect |
|---|---|---|---|
| Use feature attribution to inspect a known, visible Tag artifact on the aligned class | Feature attribution methods · Tag, across Bone Age, Knee, and Dog Breeds | K-SSD | Improved |
| Attempt to detect an unknown spurious signal using explanations on normal inputs only | Feature attribution, concept activation, and training point ranking · Unknown spurious signals | CCM | No clear effect |
| Apply feature attribution methods to the non-visible background blur | Feature attribution methods · Blur | spurious signal detection reliability | Hurt |
| Use Guided Backprop rather than other tested attribution methods for known spurious signals | Feature attribution methods · Tag and Stripe | K-SSD | Mixed |
| Present feature attributions for inputs containing a spurious artifact even when the model is normal | Feature attribution methods · Tag and Stripe | FAM-GT | Hurt |
| Explicitly test the spurious concept using TCAV | TCAV · Known visible Tag and Stripe signals | K-SSD | Improved |
| Use TCAV to test the non-visible Blur concept | TCAV · Blur | K-SSD | Hurt |
| Rely on rankings of known non-spurious concepts to reveal an unknown artifact | TCAV · Unknown spurious signals | CCM | No clear effect |
| Use training point ranking to identify reliance without knowing which artifact to inspect | Influence-function training point ranking · Spurious signals unknown at test time | spurious reliance detection | No clear effect |
| Give participants explanations without prior knowledge of possible artifacts | SmoothGrad, TCAV, and Influence · Blinded user study | rejecting defective models | No clear effect |
| Give participants prior knowledge of possible artifacts and TCAV explanations | TCAV · Not-blinded user study | rejecting defective models | Improved |
Closest papers by meaning
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.88)
- Severing Spurious Correlations with Data Pruning (ICLR 2025 · similarity 0.86)
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.86)
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.86)
- Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · similarity 0.86)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.85)
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.
