Research map / Vision-language and multimodal / Human and embodied modeling
Concepts from Representations: Post-hoc Concept Bottleneck Models via Sparse Decomposition of Visual Representations
Shizhan Gong, Xiaofan Zhang, Qi Dou
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 PCBM-ReD interpretable bottleneck instead of end-to-end linear probe | CLIP ViT-L/14 · 11 classification datasets average | test accuracy (%): 87.38 -> 86.97 | No clear effect |
| PCBM-ReD instead of LaBo | CLIP ViT-L/14 · 11 classification datasets average | test accuracy (%): 85.72 -> 86.97 | Improved |
| PCBM-ReD instead of LaBo in few-shot | PCBM-ReD · few-shot average over 11 datasets (1-16 shots) | test accuracy (%): +5.01% on average | Improved |
| zero-shot with PCBM-ReD decomposition instead of original CLIP representation | CLIP ViT-L/14 · zero-shot average over 11 datasets | test accuracy (%): 69.69 -> 69.73 | No clear effect |
| add reconstruction-guided concept selection instead of random or k-means (ablation) | PCBM-ReD · CIFAR10/CIFAR100/CUB average | test accuracy | Improved |
| sparse representation decomposition instead of CLIP-similarity concept scores (ablation) | PCBM-ReD · CIFAR10/CIFAR100/CUB average | test accuracy | Improved |
| data-driven concept creation instead of LLM-generated or WordNet concepts (ablation) | PCBM-ReD · CIFAR10/CIFAR100/CUB average | test accuracy | Improved |
| extract concepts from a mismatched image encoder (ablation) | PCBM-ReD · CIFAR10/CIFAR100/CUB average | test accuracy | Hurt |
| increase bottleneck size (ablation) | PCBM-ReD · CIFAR10/CIFAR100/CUB average | test accuracy: saturates near 300 concepts | Improved |
| add LLM-based concept rating (ablation) | PCBM-ReD · human evaluation, 5 datasets | explanation rating (1-5) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VLG-CBM — VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DCBM: Data-Efficient Visual Concept Bottleneck Models (ICML 2025 · similarity 0.89)
- V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer (AAAI 2025 · similarity 0.89)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.88)
- Attribute-formed Class-specific Concept Space: Endowing Language Bottleneck Model with Better Interpretability and Scalability (CVPR 2025 · similarity 0.88)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.88)
- Object-Centric Concept-Bottlenecks (NeurIPS 2025 · similarity 0.87)
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.
