Research map / Learning theory and algorithms / Prediction and decision losses
Window-Based Distribution Shift Detection for Deep Neural Networks
Guy Bar-Shalom, Yonatan Geifman, Ran El-Yaniv
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 coverage-based window detector instead of KS-Softmax | window shift detector · ResNet50, window 10 | AUROC: 61 -> 78 | Improved |
| use coverage-based detector instead of KS-Softmax (ViT-T, small window) | window shift detector · ViT-T, window 10 | AUROC: 62 -> 89 | Improved |
| use coverage-based detector instead of KS-Softmax at window 1000 | window shift detector · ResNet50, window 1000 | AUROC: 100 -> 100 | No clear effect |
| use coverage-based detector instead of KS-Softmax on MobileNetV3-S at window 100 | window shift detector · MobileNetV3-S, window 100 | Detection Error: 10 -> 11 | Hurt |
| use single-instance baselines (SR, Entropy) for population-level detection | single-instance t-test · ResNet50, window 1000 | AUROC: 88 -> 100 | Hurt |
| use coverage detector instead of KS-Softmax under FGSM shift | window shift detector · ResNet50 FGSM eps=1e-4, window 50 | AUROC: 64 -> 88 | Improved |
| use coverage detector instead of MMD-Embeddings on ImageNet-O, window 10 | window shift detector · ResNet50 ImageNet-O, window 10 | AUROC: 96 -> 61 | Hurt |
| use coverage detector instead of KS-Softmax under Zoom out 90% shift | window shift detector · ResNet50 Zoom 90%, window 50 | AUROC: 61 -> 81 | Improved |
| use coverage-based detection instead of KS-Softmax for run-time | window shift detector · detection-training set 1,000,000 | detection time (s): 95.3 -> 4.66 \cdot 10^{-4} | Improved |
| use a single target coverage C_target=1 instead of multiple coverages (ablation) | coverage detector · ResNet50 | average AUROC | Hurt |
| increase number of target coverages C_target (1 vs 10 vs 100) (ablation) | coverage detector · ResNet50 | average AUROC | Improved |
| vary confidence level delta over 0.1, 0.01, 0.001, 0.0001 (ablation) | coverage detector · ResNet50 | average AUROC | No clear effect |
Closest papers by meaning
- A Statistical Framework for Efficient Out of Distribution Detection in Deep Neural Networks (ICLR 2022 · similarity 0.88)
- A Fine-Grained Analysis on Distribution Shift (ICLR 2022 · similarity 0.85)
- Towards In-Distribution Compatible Out-of-Distribution Detection (AAAI 2023 · similarity 0.85)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.84)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.84)
- Learning Neural Networks with Distribution Shift: Efficiently Certifiable Guarantees (ICLR 2025 · 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.
