Research map / Label-efficient and robust learning / Recognition and long-tail
FACE: Faithful Automatic Concept Extraction
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 FACE, an NMF-based concept extraction method that adds a KL-divergence term between classifier-head predictions on original and reconstructed activations to make discovered concepts more faithful to the model.
Problem
Automatic concept discovery methods (ACE, ICE, CRAFT) factorize encoder activations without accounting for the downstream classifier, so extracted concepts may not reflect the model's true decision-making, compromising explanation faithfulness.
Approach
Augment NMF reconstruction loss with a KL divergence between softmax predictions of the frozen classifier head on original vs. reconstructed activations, so classifier supervision shapes concept learning; optimized with projected gradient descent and NNDSVD initialization.
Key results
- On ResNet-34 ImageNet, FACE reaches C-Del 0.891 vs. 0.752 for CRAFT and 0.484 for ICE, and C-Gini 0.895 vs. 0.835 (CRAFT).
- FACE attains the lowest KL divergence among methods on all datasets (e.g., ResNet-34 ImageNet 0.220 vs. 0.240 CRAFT and 0.359 ICE), at the cost of higher MSE.
- FACE retains 100% top-1 accuracy on reconstructed activations across classes, whereas CRAFT drops to about 40% on 'Train'.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FACE instead of ICE or CRAFT | concept explanations with ResNet-34 or MobileNetV2 · ImageNet and COCO | C-Ins, C-Del, and C-Gini | Helped |
| Use FACE instead of ICE or CRAFT | concept explanations with ResNet-34 or MobileNetV2 · CelebA | C-Ins, C-Del, and C-Gini | Helped |
| Add a small KL regularization penalty (ablation) | FACE with ResNet-34 · ImageNet and COCO | C-Ins and C-Del | Helped |
| Increase KL regularization beyond a moderate threshold (λ ≥ 10³) (ablation) | FACE with ResNet-34 · ImageNet and COCO | C-Ins and C-Del | Hurt |
| Use stronger KL regularization, up to around λ = 10⁵ (ablation) | FACE with ResNet-34 · CelebA | C-Ins and C-Del | Helped |
| Add mild KL regularization (λ ∈ [10⁻⁵, 10¹]) (ablation) | FACE with MobileNetV2 · ImageNet and COCO | C-Ins and C-Del | Helped |
| Use excessively large KL regularization (ablation) | FACE with MobileNetV2 · ImageNet and COCO | C-Ins and C-Del | Hurt |
| Increase KL regularization up to λ = 10⁵ (ablation) | FACE with MobileNetV2 · CelebA | C-Ins and C-Del | Helped |
| Increase NMF decomposition rank from 5 to 25 (ablation) | FACE with ResNet-34 · ImageNet, COCO, and CelebA | C-Ins, C-Del, and C-Gini | Helped |
| Increase NMF decomposition rank beyond 25 (ablation) | FACE with ResNet-34 · ImageNet and COCO | C-Ins and C-Del | No clear effect |
| Increase NMF decomposition rank from 5 to 25 (ablation) | FACE with MobileNetV2 · ImageNet, COCO, and CelebA | C-Ins, C-Del, and C-Gini | Helped |
| Increase NMF decomposition rank beyond 25 (ablation) | FACE with MobileNetV2 · ImageNet, COCO, and CelebA | C-Ins and C-Del | No clear effect |
Surprising findings
- FACE has higher reconstruction MSE than CRAFT and ICE on most settings, showing that low reconstruction error does not imply faithful predictions.
- Large KL regularization (lambda >= 10^3) degrades faithfulness on ImageNet and COCO but not on CelebA, which benefits from stronger regularization.
Limitations
- Class-specific, global concept explanations rather than per-instance explanations.
- Requires hyperparameter tuning (lambda) per dataset; no human evaluation of interpretability.
- Applicable only to CNN architectures; not directly applicable to transformers such as ViT.
Experimental setup
Datasets: ImageNet COCO CelebA
Benchmarks: ImageNet COCO CelebA
Metrics: MSE KL divergence Concept Insertion (C-Ins) Concept Deletion (C-Del) Gini index (C-Gini) top-1 accuracy on reconstructed activations
Baselines: ICE CRAFT ACE
Builds on: CRAFT ICE NMF Sobol indices NNDSVD
Closest papers by meaning
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.86)
- ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models (CVPR 2025 · similarity 0.86)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.85)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.85)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.85)
- Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models (CVPR 2024 · similarity 0.85)
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.