Research map / Label-efficient and robust learning / Recognition and long-tail
Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SALF-CBM, a label-free, spatially-aware concept bottleneck that turns backbone features into CLIP-supervised concept maps, giving both concept and heatmap explanations with classification accuracy competitive with the backbone.
Problem
Existing concept bottleneck models give only global concept explanations without localizing concepts in the image, and their interpretable bottleneck often lowers task accuracy; attribution heatmaps lack semantic grounding.
Approach
A 1x1-conv spatial concept bottleneck over backbone feature maps is trained to match CLIP local image-concept similarities computed via red-circle visual prompting on a grid, yielding per-location concept maps without human labels; a sparse linear layer over pooled concept maps makes the prediction.
Key results
- With a sparse final layer, SALF-CBM reaches 46.73% on Places365 and 75.32% on ImageNet, above the sparse standard backbone (38.46%, 74.35%) and LF-CBM.
- On zero-shot segmentation over ImageNet-segmentation, SALF-CBM achieves 76.94% pixel accuracy and 58.30% mIoU, best of the compared heatmap methods, with 85.31% mAP (second to FullGrad's 88.35%).
- In a user study, SALF-CBM concept neurons score higher than baseline ResNet-50 neurons on semantic consistency and concept accuracy across neuron interpretability groups.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SALF-CBM instead of LF-CBM with a sparse final layer | ResNet-18 · CUB-200 | accuracy: 74.31% -> 74.35% | No clear effect |
| Use SALF-CBM instead of LF-CBM with a sparse final layer | ResNet-50 · Places365 | accuracy: 43.68% -> 46.73% | Helped |
| Use SALF-CBM instead of LF-CBM with a sparse final layer | ResNet-50 · ImageNet | accuracy: 71.95% -> 75.32% | Helped |
| Use SALF-CBM instead of the standard model with a sparse final layer | ResNet-18 · CUB-200 | accuracy: 75.96% -> 74.35% | Hurt |
| Use SALF-CBM instead of the standard model with a sparse final layer | ResNet-50 · Places365 | accuracy: 38.46% -> 46.73% | Helped |
| Use SALF-CBM instead of the standard model with a sparse final layer | ResNet-50 · ImageNet | accuracy: 74.35% -> 75.32% | Helped |
| Use SALF-CBM instead of the standard model with a non-sparse final layer | ResNet-18 · CUB-200 | accuracy: 76.70% -> 76.21% | Hurt |
| Use SALF-CBM instead of the standard model with a non-sparse final layer | ResNet-50 · Places365 | accuracy: 48.56% -> 49.38% | Helped |
| Use SALF-CBM instead of the standard model with a non-sparse final layer | ResNet-50 · ImageNet | accuracy: 76.13% -> 76.26% | Helped |
| Use SALF-CBM heatmaps instead of FullGrad heatmaps | ResNet-50 · ImageNet-segmentation | pixel accuracy: 73.04% -> 76.94% | Helped |
| Use SALF-CBM heatmaps instead of FullGrad heatmaps | ResNet-50 · ImageNet-segmentation | mIoU: 55.78% -> 58.30% | Helped |
| Use SALF-CBM heatmaps instead of FullGrad heatmaps | ResNet-50 · ImageNet-segmentation | mAP: 88.35% -> 85.31% | Hurt |
Surprising findings
- The non-sparse SALF-CBM slightly exceeds the non-sparse backbone on ImageNet and Places365 despite using interpretable concepts.
Limitations
- Concept lists and hyperparameters (grid size, circle radius) are tuned per dataset
- Heatmap resolution depends on the visual prompting grid density
- Relies on GPT-generated concept lists and CLIP's local similarity quality
Experimental setup
Datasets: ImageNet Places365 CUB-200 ImageNet-segmentation
Benchmarks: CUB-200 Places365 ImageNet ImageNet-segmentation
Metrics: classification accuracy pixel accuracy mIoU mAP user study scores (semantic consistency, concept accuracy)
Baselines: Standard backbone (sparse and non-sparse) P-CBM LF-CBM LRP Integrated gradients GradCAM GradCAM++ ScoreCAM FullGrad
Builds on: Label-Free CBM Post-Hoc CBM CLIP visual prompting GLM-SAGA
Closest papers by meaning
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.88)
- V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer (AAAI 2025 · similarity 0.87)
- Explanation Bottleneck Models (AAAI 2025 · similarity 0.87)
- Interpretable Generative Models through Post-hoc Concept Bottlenecks (CVPR 2025 · similarity 0.87)
- Object-Centric Concept-Bottlenecks (NeurIPS 2025 · similarity 0.87)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 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.