Research map / Label-efficient and robust learning / OOD and anomaly detection
Dual Energy-Based Model with Open-World Uncertainty Estimation for 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 DEBO, a dual-classifier energy-based model whose 2K-output head separates ID classes from K negative OOD classes, trained with an unsupervised entropy term on auxiliary outliers to detect OOD samples without empirical energy thresholds.
Problem
Out-of-distribution detection in open-world settings, where classifiers are over-confident and energy-regularized methods (e.g., EBO) depend on empirically estimated energy thresholds that limit scalability and generalization.
Approach
A single classifier with 2K outputs (K positive ID labels plus K negative labels for the OOD samples nearest each class) is trained with a unified energy-based loss that approximates minimizing KL divergence for marginal and joint distributions, plus an entropy-minimization term on auxiliary OOD data; detection uses a hybrid score combining conditional confidence and negative marginal energy, remo
Key results
- On CIFAR-100, DEBO's average FPR95 is 33.73 versus 50.37 for the second-best method, EBO (w. Daux).
- On CIFAR-10, DEBO reaches FPR95 14.41 and AUROC 96.67 with ID-ACC 94.86, versus 16.99 and 93.74 for EBO (w. Daux).
- On ImageNet-200, DEBO reports AUROC 88.10 and FPR95 41.12, outperforming the other methods on three of four OOD indicators.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add C′′x to Ccse + Cen (ablation) | DEBO · CIFAR-10 | FPR95: 26.56 -> 14.41 | Helped |
| Add Cen to Ccse + C′′x (ablation) | DEBO · CIFAR-10 | AUROC: 96.46 -> 96.67 | Helped |
| Add C′′x to Ccse + Cen (ablation) | DEBO · CIFAR-100 | FPR95: 58.46 -> 33.73 | Helped |
| Add Cen to Ccse + C′′x (ablation) | DEBO · CIFAR-100 | AUROC: 86.28 -> 89.76 | Helped |
| Add C′′x to Ccse + Cen (ablation) | DEBO · ImageNet-200 | FPR95: 42.97 -> 41.12 | Helped |
| Add Cen to Ccse + C′′x (ablation) | DEBO · ImageNet-200 | AUROC: 87.55 -> 88.10 | Helped |
| Use Sdec instead of MCP score (ablation) | DEBO · CIFAR-10 | AUROC: 94.24 -> 96.67 | Helped |
| Use Sdec instead of Sjoint score (ablation) | DEBO · CIFAR-10 | AUROC: 96.23 -> 96.67 | Helped |
| Use Sdec instead of MCP score (ablation) | DEBO · CIFAR-100 | AUROC: 83.57 -> 89.76 | Helped |
| Use Sdec instead of Sjoint score (ablation) | DEBO · CIFAR-100 | AUROC: 85.67 -> 89.76 | Helped |
| Use Sdec instead of MCP score (ablation) | DEBO · ImageNet-200 | AUROC: 87.78 -> 88.10 | Helped |
| Use Sdec instead of Sjoint score (ablation) | DEBO · ImageNet-200 | AUROC: 87.73 -> 88.10 | Helped |
Surprising findings
- The ablation shows the score combining conditional confidence and negative marginal energy (Sdec) beats the joint-energy score and MCP baseline on all datasets.
Limitations
- Hyperparameters lambda1, lambda2 and gamma must be tuned using auxiliary outliers and validation data, and the search uses OOD test sets' AUROC.
- Results are limited to image classification with ResNet-18 on CIFAR and a 200-class ImageNet subset.
- ID-ACC on CIFAR-100 (76.26) is slightly below MSP/EBO (76.83), and on CIFAR-10 the ID accuracy is only comparable.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200 TIN-597 ImageNet-1K remaining 800 classes
Benchmarks: MNIST Textures SVHN Places365 LSUN-Crop LSUN-Resize iSUN SSB-hard NINCO iNaturalist OpenImage-O
Metrics: FPR95 AUROC AUPR-In AUPR-Out ID-ACC
Baselines: MSP EBO CIDER OE UDG EBO (w. Daux) MCD MIXOE
Builds on: Energy-based OOD detection (EBO) Joint Energy-based Models (JEM) Outlier Exposure CIDER Stochastic Gradient Langevin Dynamics
Closest papers by meaning
- Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.88)
- Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity (NeurIPS 2025 · similarity 0.88)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.87)
- EDGE: Unknown-aware Multi-label Learning by Energy Distribution Gap Expansion (AAAI 2025 · similarity 0.87)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.87)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · 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.