Research map / Vision-language and multimodal / CLIP and zero-shot
Concept-Guided Fine-Tuning: Steering ViTs away from Spurious Correlations to Improve Robustness
Yehonatan Elisha, Oren Barkan, Noam Koenigstein
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 |
|---|---|---|---|
| concept-guided relevance fine-tuning (CFT) | ViT-B · ImageNet-A | Top-1: 13.26 -> 27.76 | Improved |
| CFT instead of RRDA baseline | RRDA · ImageNet-A | Top-1: 25.12 -> 27.76 | Improved |
| use concept masks instead of ground-truth object segmentation masks (ablation) | CFT · ImageNet-A | Top-1: 24.12 -> 26.05 | Improved |
| add classification-consistency loss (ablation) | CFT · ImageNet-A | Top-1: 17.96 -> 24.11 | Improved |
| add non-concept suppression loss (ablation) | CFT · ImageNet-A | Top-1: 19.14 -> 24.11 | Improved |
| add concept relevance loss (ablation) | CFT · ImageNet-A | Top-1: 24.10 -> 24.11 | No clear effect |
| use classification-consistency loss instead of ground-truth CE (ablation) | CFT · ImageNet-A | Top-1: 21.35 -> 24.11 | Improved |
| use AttnLRP instead of GradCAM relevance (ablation) | CFT · ImageNet-A | Top-1: 25.88 -> 27.82 | Improved |
| fine-tune on half of classes, evaluate held-out classes (ablation) | CFT · ImageNet-A held-out classes | Top-1: 17.20 -> 25.43 | Improved |
Closest papers by meaning
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.89)
- Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers (NeurIPS 2025 · similarity 0.88)
- Proxy-FDA: Proxy-based Feature Distribution Alignment for Fine-tuning Vision Foundation Models without Forgetting (ICML 2025 · similarity 0.87)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.86)
- ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features (ICML 2025 · similarity 0.86)
- Explaining Domain Shifts in Language: Concept Erasing for Interpretable Image Classification (CVPR 2025 · similarity 0.86)
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.
