Research map / Label-efficient and robust learning / OOD and anomaly detection
Detecting Out-of-Distribution Through the Lens of Neural Collapse
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 NCI, a post-hoc OOD detector inspired by Neural Collapse that scores ID samples by the projection norm of centered features onto predicted-class weight vectors (pScore) plus an L1 feature-norm filter.
Problem
Post-hoc OOD detection that generalizes across ID datasets (CIFAR-10 vs ImageNet) and architectures while remaining computationally cheap.
Approach
Re-characterizes ID feature clustering through Neural Collapse: after centering by the training feature mean, ID features align with predicted-class weight vectors and lie farther from the origin (ETF structure). The detector combines an angle-based proximity score (pScore = cosine times weight-vector norm) with an L1 feature-norm filter.
Key results
- On CIFAR-10 ResNet-18 with SVHN, the method reaches AUROC 94.44 versus 91.27 for the softmax-confidence baseline by epoch 50 (introduction footnote).
- Average AUROC across all benchmarks is 88.57 for NCI versus 86.06 for KNN and 86.8 for Scale (Table 3).
- Inference latency of 0.54 ms on CIFAR-10 and 6.84 ms on ImageNet, comparable to MSP (0.53 and 6.85 ms) (Table 1b).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NCI instead of strong baselines | ResNet-18 and ResNet-50 OOD detection · CIFAR-10 and ImageNet OpenOOD benchmarks | Avg AUROC: 88.16 -> 89.51 | Helped |
| Use NCI instead of strong baselines | ViT B/16 and Swin v2 OOD detection · ImageNet OpenOOD benchmarks | Avg AUROC: 85.40 -> 87.31 | Helped |
| Add feature-norm filtering to NCI (ablation) | ResNet-50 OOD detection · ImageNet OpenOOD benchmark | Avg AUROC: 84.41 -> 88.56 | Helped |
| Add feature-norm filtering to NCI (ablation) | ViT B/16 OOD detection · ImageNet OpenOOD benchmark | Avg AUROC: 93.26 -> 93.64 | Helped |
| Add feature-norm filtering to NCI (ablation) | Swin v2 OOD detection · ImageNet OpenOOD benchmark | Avg AUROC: 78.87 -> 80.97 | Helped |
| Use L1 norm rather than Linf norm for filtering (ablation) | NCI with ResNet-50 · ImageNet OpenOOD benchmark, SSB-hard | AUROC: 66.81 -> 73.90 | Helped |
| Use L1 norm rather than L2 norm for filtering (ablation) | NCI with ResNet-50 · ImageNet OpenOOD benchmark, SSB-hard | AUROC: 69.12 -> 73.90 | Helped |
| Vary filtering strength within the tested range (ablation) | NCI with ResNet-50 · ImageNet ResNet-50 OpenOOD benchmark | Avg AUROC: 88.27; 88.55; 88.59; 88.50; 88.23 | No clear effect |
| Add L1-norm filtering to KNN (ablation) | KNN OOD detector · ImageNet OpenOOD benchmark | Avg AUROC: 82.43 -> 86.37 | Helped |
| Use NCI with L1 filtering instead of KNN + L1 | ImageNet OOD detection · ImageNet OpenOOD benchmark | Avg AUROC: 86.37 -> 88.56 | Helped |
| Use NCI instead of KNN | ImageNet OOD detection · ImageNet OpenOOD benchmark | per-image latency | Helped |
Surprising findings
- Strong baselines such as ASH and Scale suffer markedly on Swin v2 (e.g., ASH averaged AUROC near 57 on the ImageNet near-OOD sets), and NCI improves over them.
- Filtering with L1 norm applied to KNN gives a sizable gain, suggesting the feature-norm perspective helps clustering-based methods broadly.
Limitations
- Performance depends on the strength of ID clustering; weaker clustering (e.g., ImageNet ResNet-50) needs the norm-based filter.
- Full Neural Collapse does not occur in the off-the-shelf models studied, so the theory is motivational rather than exact.
- Filter strength must be chosen on a validation set of synthetic Gaussian noise, and the choice of norm affects results.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet
Benchmarks: OpenOOD CIFAR-10 benchmark OpenOOD ImageNet benchmark
Metrics: FPR95 AUROC latency
Baselines: MSP ODIN Energy Mahalanobis KNN ViM fDBD GradNorm NECO ReAct DICE ASH
Builds on: Neural Collapse KNN OOD detection OpenOOD benchmark ASH Scale
Closest papers by meaning
- NECO: NEural Collapse Based Out-of-distribution detection (ICLR 2024 · similarity 0.94)
- Pursuing Feature Separation based on Neural Collapse for Out-of-Distribution Detection (ICLR 2025 · similarity 0.92)
- MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse (NeurIPS 2026 · similarity 0.91)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.91)
- Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learning (ICML 2025 · similarity 0.91)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.90)
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.