Research map / Label-efficient and robust learning / OOD and anomaly detection
NECO: NEural Collapse Based Out-of-distribution detection
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 NECO, a post-hoc OOD detector that scores the relative norm of a feature projected onto the top principal components of ID features (leveraging neural collapse and a newly observed ID/OOD orthogonality property), multiplied by MaxLogit.
Problem
Post-hoc out-of-distribution detection for overconfident deep classifiers; prior feature-subspace methods (ViM, NuSA) do not exploit neural collapse structure.
Approach
Uses neural collapse properties (NC1, NC2) plus a newly observed ID/OOD orthogonality (NC5) so that OOD features project near zero onto the principal subspace spanned by ID class means; the score is the normalized norm of the principal-subspace projection, multiplied by MaxLogit for calibration.
Key results
- On ImageNet-1K with ViT-B/16, NECO reaches average AUROC 94.05 and FPR 27.51, beating the second-best method ViM by about 1.95 FPR points.
- On CIFAR-10/100 with ViT, NECO averages AUROC 99.44 (CIFAR-100 ID) and 95.41 (CIFAR-10 ID), the best average among compared methods.
- With SwinV2-B/16 on ImageNet-1K, NECO has the lowest FPR on every dataset, reducing average FPR by 7.47 points versus the second-best Softmax score.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NECO instead of ViM | ViT-B/16 · ImageNet-1K OOD average | AUROC: 93.56 -> 94.05 | Helped |
| Use NECO instead of ViM | ViT-B/16 · ImageNet-1K OOD average | FPR: 29.46 -> 27.51 | Helped |
| Use NECO instead of Energy+ReAct | SwinV2 · ImageNet-1K OOD average | AUROC: 81.44 -> 80.56 | Hurt |
| Use NECO instead of Softmax score | SwinV2 · ImageNet-1K OOD average | FPR: 66.66 -> 59.19 | Helped |
| Use NECO instead of Energy+ReAct | ViT · CIFAR-10 ID, CIFAR-100 and SVHN OOD average | AUROC: 99.30 -> 99.44 | Helped |
| Use NECO instead of ViM | ViT · CIFAR-10 ID, CIFAR-100 and SVHN OOD average | FPR: 2.84 -> 2.47 | Helped |
| Use NECO instead of ViM | ViT · CIFAR-100 ID, CIFAR-10 and SVHN OOD average | AUROC: 95.17 -> 95.41 | Helped |
| Use NECO instead of Energy+ReAct | ViT · CIFAR-100 ID, CIFAR-10 and SVHN OOD average | FPR: 20.91 -> 19.43 | Helped |
| Use NECO instead of NuSA | ResNet-18 · CIFAR-10 ID, CIFAR-100 and SVHN OOD average | AUROC: 90.71 -> 91.27 | Helped |
| Use NECO instead of NuSA | ResNet-18 · CIFAR-10 ID, CIFAR-100 and SVHN OOD average | FPR: 40.67 -> 39.12 | Helped |
| Use NECO instead of NuSA | ResNet-18 · CIFAR-100 ID, CIFAR-10 and SVHN OOD average | AUROC: 79.38 -> 79.42 | Helped |
| Use NECO instead of NuSA | ResNet-18 · CIFAR-100 ID, CIFAR-10 and SVHN OOD average | FPR: 71.97 -> 70.03 | Helped |
Surprising findings
- Using only the NECO score without MaxLogit multiplication performs noticeably worse on ViT and DeiT, which the authors attribute to limited feature dimension preventing full collapse to the Simplex ETF.
- NECO underperforms on DeiT, which the authors attribute to DeiT's distillation training procedure hindering neural collapse.
Limitations
- Performance degrades on DeiT and models whose feature dimension is small relative to the number of classes, requiring MaxLogit calibration.
- Relies on over-parameterized models trained into the terminal phase of training to exhibit neural collapse.
- Hyperparameter choice of principal subspace dimension d is tuned per ID/OOD pair using OOD results in the main tables, which may be optimistic.
Experimental setup
Datasets: ImageNet-1K ImageNet-21K CIFAR-10 CIFAR-100 SVHN Textures Places365 iNaturalist ImageNet-O SUN
Benchmarks: ImageNet-1K OOD suite (Textures, ImageNet-O, iNaturalist, SUN, Places365) CIFAR-10 vs CIFAR-100/SVHN CIFAR-100 vs CIFAR-10/SVHN
Metrics: AUROC FPR@95TPR
Baselines: Softmax score MaxLogit Energy Energy+ReAct ViM Residual GradNorm Mahalanobis KL-Matching ASH NuSA
Builds on: ViM NuSA neural collapse MaxLogit
Closest papers by meaning
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.94)
- MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse (NeurIPS 2026 · similarity 0.90)
- Pursuing Feature Separation based on Neural Collapse for Out-of-Distribution Detection (ICLR 2025 · similarity 0.90)
- Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learning (ICML 2025 · similarity 0.90)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.89)
- Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity (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.