Research map / Label-efficient and robust learning / OOD and anomaly detection
Neural Mean Discrepancy for Efficient Out-of-Distribution Detection
Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, H.T. Kung
Check 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 channel-wise neural mean discrepancies from a pretrained ConvNet with a logistic-regression detector instead of statistical kernel baselines | ConvNet (4 layers) · CIFAR-10 vs SVHN | AUROC: 84.9 -> 99.9 | Improved |
| Use channel-wise neural mean discrepancies from a pretrained ConvNet with a logistic-regression detector instead of statistical kernel baselines | ConvNet (4 layers) · SVHN vs CIFAR-10 | AUROC: 80.3 -> 99.8 | Improved |
| Use NMD-MLP instead of the strongest listed baseline | ResNet-34 OOD detection · CIFAR-10 ID; Texture OOD | AUROC: 97.5 -> 98.9 | Improved |
| Use NMD-MLP instead of the strongest listed baseline | ResNet-34 OOD detection · CIFAR-10 ID; CIFAR-100 OOD | AUROC: 87.2 -> 90.1 | Improved |
| Use NMD-MLP instead of the strongest listed baseline | ResNet-34 OOD detection · CIFAR-10 ID; LSUN-C OOD | AUROC: 98.8 -> 98.8 | No clear effect |
| Train the detector on CIFAR-100 and transfer it to unseen OOD data | NMD with ResNet-34 · CIFAR-10 ID; LSUN-C, SVHN, Texture, and ImageNet-C unseen OOD | AUROC | Improved |
| Replace activation mean discrepancy with activation variance discrepancy (ablation) | NMD detector · CIFAR-10 vs CIFAR-100 | AUROC: 95.4 -> 95.3 | No clear effect |
| Add activation variance discrepancy and concatenate it with NMD (ablation) | NMD detector · CIFAR-10 vs CIFAR-100 | AUROC: 95.4 -> 95.6 | Improved |
| Use an untrained rather than pretrained ResNet-34 to compute NMD (ablation) | NMD detector · OOD batch detection | AUROC | Hurt |
| Use only the first five layers' NMD values rather than more layers (ablation) | NMD detector · CIFAR-10 vs SVHN | AUROC | No clear effect |
| Use more NMD layers to capture semantic differences (ablation) | NMD detector · CIFAR-10 vs CIFAR-100 | AUROC | Improved |
| Use a higher-capacity MLP rather than LR with few training examples (ablation) | NMD detector · Few-shot OOD training | detection performance | Hurt |
Closest papers by meaning
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.91)
- Out-of-Distribution Detection based on In-Distribution Data Patterns Memorization with Modern Hopfield Energy (ICLR 2023 · similarity 0.91)
- A Statistical Framework for Efficient Out of Distribution Detection in Deep Neural Networks (ICLR 2022 · similarity 0.91)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.90)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.90)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.89)
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.
