Research map / Label-efficient and robust learning / Recognition and long-tail
Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 TCAV to compare image and caption concept importance | Image classifiers trained with noisy captions · Controlled cab and cucumber experiment | agreement with approximated ground truth | Helped |
| Communicate concept importance with saliency maps rather than TCAV | Image classifiers trained with noisy captions · Human evaluation of saliency maps | correct concept-importance judgments: 52% (random chance is 50%) | No clear effect |
| Apply statistical significance testing to filter CAVs | TCAV · Zebra class, mixed4a layer | spurious high TCAV signals retained | Helped |
| Use 30 examples per concept to learn CAVs | TCAV for the dumbbell class · ImageNet dumbbell class | identification of arms as important | Helped |
| Use TCAV to inspect concept importance | Image classifier · Adversarial zebra images | difference in TCAV-score distributions from regular zebras | Helped |
| Use TCAV to examine medical concepts | Diabetic retinopathy classifier · DR level 4 | alignment of concept importance with diagnostic relevance | Helped |
| Use TCAV to examine medical concepts | Diabetic retinopathy classifier · DR level 1 | alignment of concept importance with doctors' heuristics | Mixed |
Closest papers by meaning
- Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence (ICLR 2025 · similarity 0.88)
- FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks (ICML 2025 · similarity 0.87)
- LG-CAV: Train Any Concept Activation Vector with Language Guidance (NeurIPS 2024 · similarity 0.84)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.83)
- Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable? (NeurIPS 2024 · similarity 0.82)
- Manipulating Feature Visualizations with Gradient Slingshots (NeurIPS 2025 · similarity 0.82)
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.