Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
REPEAT: Improving Uncertainty Estimation in Representation Learning Explainability
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 REPEAT, an R-XAI method that treats each pixel's importance as a Bernoulli variable obtained by repeatedly thresholding stochastic RELAX importance maps, yielding importance and certainty estimates.
Problem
Existing representation learning XAI (R-XAI) uncertainty estimates measure variability in importance scores rather than whether a pixel is certainly important or not.
Approach
Model each pixel as a Bernoulli random variable (important vs unimportant) by thresholding repeated stochastic R-XAI importance maps with histogram-based (mean) thresholding; importance is the estimated probability and uncertainty is p(1-p).
Key results
- REPEAT is the only method passing the eMPRT sanity check, scoring 0.48 on ResNet50 and 0.13 on ViT on PASCAL-VOC versus -0.08 and -0.07 for RELAX.
- On EuroSAT OOD detection with ResNet50 on PASCAL-VOC, REPEAT reaches AUROC 1.000 versus 0.000 for RELAX.
- REPEAT yields lowest complexity across settings, e.g. 9.97 vs 10.82 for RELAX on PASCAL-VOC with ResNet50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use REPEAT uncertainty estimates instead of RELAX | R-XAI uncertainty estimation · PASCAL-VOC and MS-COCO, ResNet50 and ViT | eMPRT: -0.08 -> 0.48; -0.10 -> 0.13; -0.08 -> 0.44; -0.08 -> 0.13 | Helped |
| use REPEAT uncertainty estimates instead of Integrated Gradients uncertainty estimates | OOD detection · PASCAL-VOC and MS-COCO vs EuroSAT, ResNet50 | AUROC: 0.998 -> 1.000; 0.999 -> 1.000 | Helped |
| use REPEAT uncertainty estimates instead of baseline uncertainty estimates | OOD detection · PASCAL-VOC and MS-COCO vs EuroSAT, ViT | AUROC: 0.002 -> 0.973; 0.002 -> 0.939 | Helped |
| use REPEAT uncertainty estimates instead of the strongest baseline estimates | uncertainty estimation · PASCAL-VOC and MS-COCO, ResNet50 and ViT | uncertainty complexity: 10.41 -> 9.97; 10.46 -> 9.99; 10.35 -> 9.91; 10.40 -> 9.94 | Helped |
| use Kernel SHAP instead of RELAX as the base stochastic R-XAI method (ablation) | REPEAT · PASCAL-VOC and MS-COCO vs EuroSAT, ResNet50 and ViT | AUROC: 0.999 vs 1.000; 0.999 vs 0.973; 1.000 vs 1.000; 1.000 vs 0.939 | No clear effect |
| use Kernel SHAP instead of RELAX as the base stochastic R-XAI method (ablation) | REPEAT · PASCAL-VOC and MS-COCO, ResNet50 and ViT | uncertainty complexity: 9.97 -> 10.40; 9.99 -> 10.36; 9.91 -> 10.28; 9.94 -> 10.28 | Hurt |
Surprising findings
- RELAX's uncertainty gives lower values for OOD data than in-distribution data, the opposite of the desired behavior.
Limitations
- Uncertainty for baselines relies on test-time augmentation with Dropout on inputs, which is an ad hoc substitute.
- Relies on stochasticity of the base R-XAI method; repeated thresholding cost scales with number of realizations K=10.
Experimental setup
Datasets: PASCAL VOC MS-COCO EuroSAT FashionMNIST ImageNet
Benchmarks: PASCAL VOC MS-COCO
Metrics: eMPRT sanity check score AUROC for OOD detection entropy-based complexity
Baselines: RELAX Saliency Guided Backpropagation Integrated Gradients Kernel SHAP
Builds on: RELAX Label-Free Feature Importance Kernel SHAP Test-time augmentation eMPRT
Closest papers by meaning
- Benchmarking Uncertainty Disentanglement: Specialized Uncertainties for Specialized Tasks (NeurIPS 2024 · similarity 0.85)
- Using Stratified Sampling to Improve LIME Image Explanations (AAAI 2024 · similarity 0.84)
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.84)
- Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? (NeurIPS 2024 · similarity 0.84)
- SLICE: Stabilized LIME for Consistent Explanations for Image Classification (CVPR 2024 · similarity 0.84)
- Maxitive Donsker-Varadhan Formulation for Possibilistic Variational Inference (NeurIPS 2026 · similarity 0.83)
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.