Research map / Label-efficient and robust learning / Recognition and long-tail
Overlooked Factors in Concept-Based Explanations: Dataset Choice, Concept Learnability, and Human Capability
Vikram V. Ramaswamy, Sunnie S. Y. Kim, Ruth Fong, Olga Russakovsky
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 |
|---|---|---|---|
| Switch the probe dataset between ADE20k and Pascal | Baseline, NetDissect, and TCAV explanations for a Places365 scene classifier · ADE20k vs. Pascal probe datasets | explanation similarity: 0.078 average cosine similarity | Hurt |
| Use the Pascal probe dataset rather than ADE20k | NetDissect neuron labels · 123 labeled neurons | label agreement: 69/123 | Hurt |
| Use Broden concepts as prediction targets instead of Places365 classes | Linear concept predictors · Broden concepts vs. Places365 classes | median normalized AP: 37.5% -> 7.6% | Hurt |
| Use CUB concepts as prediction targets instead of CUB classes | Linear concept predictors · CUB concepts vs. CUB classes | median normalized AP: 65.9% -> 2.3% | Hurt |
| Show 32 rather than 8 concepts | Baseline explanations in the human study · Human concept recognition | mean recall: 71.7% ± 27.7% -> 56.8% ± 24.9% | Hurt |
| Show more concepts, from 8 to 16 or 32 | Baseline explanations in the human study · Human model-output prediction | mean accuracy: 64.8% ± 23.9% (8 concepts), 63.2% ± 26.9% (16 concepts), 63.6% ± 22.2% (32 concepts) | Mixed |
| Use example-based explanations instead of 8-concept explanations | Human model-output prediction · Human model-output prediction | mean accuracy: 64.8% ± 23.9% vs. 60.0% ± 30.2% | No clear effect |
| Limit explanations to 8, 16, or 32 concepts rather than 4 or 64 | Concept explanations in the preference study · Human preference for explanation complexity | participants preferring the selected concept counts: 82% | Improved |
Closest papers by meaning
- Causal Concept Explanations for Deep Neural Models (NeurIPS 2026 · similarity 0.90)
- IDEA: Unwrapping Visual Black-box Models by Interaction Decomposition (NeurIPS 2026 · similarity 0.90)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.90)
- Understanding Inter-Concept Relationships in Concept-Based Models (ICML 2024 · similarity 0.90)
- Attention-based Interpretability with Concept Transformers (ICLR 2022 · similarity 0.90)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.88)
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.
