Research map / Vision-language and multimodal / Human and embodied modeling
Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation 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 PACE, a variational Bayesian hierarchical model that explains ViT predictions with dataset-, image-, and patch-level concepts, evaluated on five desiderata against post-hoc baselines.
Problem
Post-hoc concept explanations for vision transformers lack faithfulness, stability, sparsity, multi-level structure, and parsimony, and existing methods are not evaluated comprehensively.
Approach
A variational Bayesian mixture-of-Gaussians model over ViT patch embeddings, with a Dirichlet image-level concept prior, a top-down generative process from dataset to image to patch concepts, a label-likelihood term for faithfulness, and a same-image contrastive term for stability.
Key results
- PACE averages 0.72 linear faithfulness, 0.11 stability (lower better), and 0.64 sparsity across four datasets, the best of all compared methods.
- On Color, PACE reaches faithfulness 1.00 and stability 0.20, with sparsity 0.97.
- PACE uses K=100 concepts versus 768 for SHAP, LIME, Saliency and AGI, matching CRAFT on parsimony.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PACE instead of Saliency | ViT-Base explanations · Color | faithfulness: 1.00 -> 1.00 | No clear effect |
| Use PACE instead of Saliency | ViT-Base explanations · Color | stability: 0.35 -> 0.20 | Helped |
| Use PACE instead of LIME | ViT-Base explanations · Color | sparsity: 0.59 -> 0.97 | Helped |
| Use PACE instead of Saliency | ViT-Base explanations · Flower | faithfulness: 0.57 -> 0.80 | Helped |
| Use PACE instead of Saliency | ViT-Base explanations · Flower | stability: 0.47 -> 0.12 | Helped |
| Use PACE instead of LIME | ViT-Base explanations · Flower | sparsity: 0.54 -> 0.48 | Hurt |
| Use PACE instead of Saliency | ViT-Base explanations · Cars | faithfulness: 0.50 -> 0.50 | No clear effect |
| Use PACE instead of Saliency | ViT-Base explanations · Cars | stability: 0.43 -> 0.05 | Helped |
| Use PACE instead of LIME | ViT-Base explanations · Cars | sparsity: 0.52 -> 0.49 | Hurt |
| Use PACE instead of Saliency | ViT-Base explanations · CUB | faithfulness: 0.49 -> 0.56 | Helped |
| Use PACE instead of Saliency | ViT-Base explanations · CUB | stability: 0.48 -> 0.05 | Helped |
| Use PACE instead of LIME | ViT-Base explanations · CUB | sparsity: 0.55 -> 0.63 | Helped |
Limitations
- Assumes a fixed number of concepts K, as with existing methods.
- Requires access to hidden states and attention weights inside the ViT, so it cannot explain output-only models.
Experimental setup
Datasets: Color (synthetic, constructed) Oxford 102 Flower Stanford Cars CUB-200-2011
Benchmarks: Color Flower Cars CUB
Metrics: linear faithfulness (logistic regression accuracy) stability (normalized L2 difference of concept activations) sparsity (fraction of near-zero concept entries) multi-level structure (qualitative) parsimony (number of concepts K)
Baselines: SHAP LIME Saliency AGI CRAFT
Builds on: CRAFT latent dirichlet allocation contrastive learning
Closest papers by meaning
- Language Model as Visual Explainer (NeurIPS 2024 · similarity 0.89)
- On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · similarity 0.89)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.88)
- Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers (NeurIPS 2025 · similarity 0.88)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.88)
- Sparse autoencoders reveal selective remapping of visual concepts during adaptation (ICLR 2025 · similarity 0.88)
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.