Research map / Label-efficient and robust learning / Recognition and long-tail
Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification
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 LaBo rather than a linear probe | CLIP image classification, 11-dataset mean · 11-dataset mean, 1 shot | accuracy: 51.69 -> 63.35 | Helped |
| Use LaBo rather than a linear probe | CLIP image classification, 11-dataset mean · 11-dataset mean, full training data | accuracy: 87.38 -> 85.72 | Hurt |
| Use LaBo rather than PCBM without an end-to-end residual | CIFAR-10 and CIFAR-100 · CIFAR-10 and CIFAR-100 | accuracy: 84.5 -> 87.9; 56.0 -> 69.1 | Helped |
| Use 50 concepts per class rather than 1 (ablation) | LaBo · 11-dataset mean, 1 shot | accuracy: 41.89 -> 63.03 | Helped |
| Jointly select for coverage and discriminability rather than random selection (ablation) | LaBo concept selection · 11-dataset mean, 1 shot | accuracy: 59.24 -> 63.03 | Helped |
| Jointly select for coverage and discriminability rather than class similarity alone (ablation) | LaBo concept selection · 11-dataset mean, 1 shot | accuracy: 54.59 -> 63.03 | Helped |
| Use softmax rather than no activation on concept weights (ablation) | LaBo · 11-dataset mean, 1 shot | accuracy: 52.66 -> 63.03 | Helped |
| Use softmax rather than ReLU on concept weights (ablation) | LaBo · 11-dataset mean, 1 shot | accuracy: 50.40 -> 63.03 | Helped |
| Use larger GPT-3 Davinci rather than Curie to generate concepts (ablation) | LaBo · ImageNet, 1 shot | accuracy: 45.75 -> 51.09 | Helped |
| Use LaBo rather than CompDL | CUB · CUB, 1 shot | accuracy: 13.6 -> 35.1 | Helped |
| Remove language-model weight initialization and use random initialization (ablation) | LaBo · Few-shot classification | accuracy | Hurt |
| Add class names to LaBo concepts (ablation) | LaBo with CLIP ViT-B/32, 16 shots · Aircraft | accuracy: 37.29 -> 37.53 | Helped |
Closest papers by meaning
- VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · similarity 0.92)
- V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer (AAAI 2025 · similarity 0.91)
- Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations (NeurIPS 2026 · similarity 0.90)
- DCBM: Data-Efficient Visual Concept Bottleneck Models (ICML 2025 · similarity 0.90)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.90)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.89)
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.