Research map / Label-efficient and robust learning / OOD and anomaly detection
Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss
SoonCheol Noh, DongEon Jeong, Jee-Hyong Lee
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 |
|---|---|---|---|
| train with cosine-based softmax loss instead of standard softmax so feature norm separates InD and OOD (ablation) | DenseNet3 on CIFAR-10 · CIFAR-10 vs SVHN feature norm | feature norm separation | Improved |
| use normalized penultimate feature f(x)/|f(x)| for Mahalanobis distance (ablation) | COD · SVHN vs TINc, iSUN OOD | MD score overlap | Improved |
| use cosine-based softmax loss instead of standard loss (ablation) | ResNet34 and DenseNet3 · CIFAR-10, CIFAR-100, SVHN | classification accuracy: ResNet34 CIFAR-10 95.25 -> 95.34; DenseNet3 CIFAR-10 95.14 -> 95.08 | No clear effect |
| replace GRAM with COD for OOD scoring | DenseNet3 CIFAR-10 · LSUN-crop OOD | TNR@TPR95: 88.4 -> 99.6 | Improved |
| replace GRAM with COD for OOD scoring | DenseNet3 · per-image OOD scoring time | seconds per image: 0.35 -> 0.018 | Improved |
| replace GRAM with COD on some OOD datasets | DenseNet3 CIFAR-10/100 · Individual OOD sets such as TINr | AUROC: TINr CIFAR-10 99.7 (GRAM) vs 99.5 (COD) | Mixed |
| add input processing with perturbation to OOD scoring (ablation) | COD, ODIN, MAHA · DenseNet3 CIFAR-100, SVHN/LSUNc/iSUN OOD | AUROC | Mixed |
| fine-tune a standard-softmax model with cosine loss for 20 epochs (ablation) | DenseNet3 · CIFAR-10, CIFAR-100, SVHN | OOD AUROC and accuracy: accuracy 95.1 -> 94.1 (CIFAR-10); 96.4 -> 96.1 (SVHN) | No clear effect |
| use only the last layer MD (layer selection) for COD on ImageNet-30 (ablation) | COD ResNet18 · ImageNet-30 | AUROC: comparable to CSI | Improved |
Closest papers by meaning
- Out-of-Distribution Detection with Deep Nearest Neighbors (ICML 2022 · similarity 0.91)
- Out-of-Distribution Detection based on In-Distribution Data Patterns Memorization with Modern Hopfield Energy (ICLR 2023 · similarity 0.91)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.90)
- CASE: Exploiting Intra-class Compactness and Inter-class Separability of Feature Embeddings for Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse (NeurIPS 2026 · similarity 0.89)
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · 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.
