Research map / Label-efficient and robust learning / OOD and anomaly detection
WeiPer: OOD Detection using Weight Perturbations of Class Projections
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 WeiPer, which perturbs the final-layer class projection vectors to build a richer logit representation, and a KL-divergence score (WeiPer+KLD) comparing histogram densities to the ID mean, achieving near-OOD state of the art on OpenOOD.
Problem
Post-hoc out-of-distribution detection for image classifiers, especially near-OOD samples that lie close to the training distribution, where existing logit- and penultimate-feature methods degrade.
Approach
Add normalized random perturbations to the rows of the final FC weight matrix to create a cone of projections around class vectors, use these to build a perturbed logit space, and score samples by KL divergence between their activation histograms and the mean ID histogram in the penultimate and perturbed spaces, combined with MSP on the perturbed logits.
Key results
- WeiPer+KLD reaches 80.05 AUROC on ImageNet-1K near-OOD with ResNet50, outperforming ASH (78.17) by about 1.88 points.
- WeiPer+KLD obtains the best near-OOD performance across all benchmarks in the OpenOOD relative score (0.988 near, first place), with a notable dent on ViT-B/16 ImageNet far-OOD.
- WeiPer improves MSP by roughly 1-4.1 AUROC points across OpenOOD benchmarks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add WeiPer weight perturbations to MSP | MSP · OpenOOD benchmarks | AUROC: 1-4.1% AUROC | Helped |
| add WeiPer weight perturbations with KLD scoring | OOD detection · near OOD benchmarks | AUROC: 80.05 vs 78.17 | Helped |
| add the MSP term to WeiPer+KLD (ablation) | WeiPer+KLD · CIFAR10 | AUROC: 90.53 -> 90.54 | No clear effect |
| use class-correlated perturbed projections instead of weight-independent Gaussian projections (ablation) | WeiPer+KLD · CIFAR10 near and far | AUROC: 69.62 -> 90.54; 75.83 -> 93.12 | Helped |
| increase the number of random perturbations r (ablation) | WeiPer · CIFAR10 and ImageNet near OOD | AUROC | Helped |
| set perturbation strength δ above zero instead of using no perturbations (ablation) | WeiPer+KLD, WeiPer+MSP, and WeiPer+ReAct · near OOD tasks | AUROC | Helped |
| use δ=2 for WeiPer+KLD (ablation) | WeiPer+KLD · near OOD tasks | AUROC: δ=2 | Helped |
| use WeiPer+KLD with ViT-B/16 | ImageNet-1K OOD detection · near and far OOD | AUROC: 75.00 vs 80.09; 90.32 vs 93.16 | Hurt |
Surprising findings
- Random projections independent of the classifier weights (RP) detect almost no OOD samples (e.g. 69.62 AUROC on CIFAR10 near), so the class-direction perturbations are essential.
- WeiPer+KLD performs poorly on ViT-B/16 ImageNet, where the penultimate layer has fewer features than classes, with a large far-OOD drop.
Limitations
- Six hyperparameters for WeiPer+KLD, more than competitors, tuned on OpenOOD validation sets.
- Performance degrades on ViT-B/16 ImageNet, especially on far-OOD, attributed to narrow penultimate layers.
- Memory cost grows with the number of perturbations r.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet-1K MNIST SVHN Textures Places365 TinyImageNet iNaturalist OpenImage-O SSB-hard NINCO
Benchmarks: OpenOOD CIFAR10 OpenOOD CIFAR100 OpenOOD ImageNet-1K
Metrics: AUROC FPR95
Baselines: MSP ReAct KNN ASH RMDS VIM GEN NAC
Builds on: MSP ReAct OpenOOD Cramer-Wold theorem
Closest papers by meaning
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.88)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.88)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.87)
- Revisiting Logit Distributions for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.87)
- CASE: Exploiting Intra-class Compactness and Inter-class Separability of Feature Embeddings for Out-of-Distribution Detection (AAAI 2024 · similarity 0.87)
- Position: Supervised Classifiers Answer the Wrong Questions for OOD Detection (ICML 2025 · similarity 0.87)
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.