Research map / Label-efficient and robust learning / OOD and anomaly detection
Your classifier is secretly an energy based model and you should treat it like one
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the classifier as a joint energy-based model using the log-sum-exp of its logits | WideResNet-28-10 · CIFAR100 | calibration | Helped |
| train the classifier as a joint energy-based model | WideResNet-28-10 · CIFAR100 | accuracy: 74.2% -> 72.2% | Hurt |
| use the log-density gradient norm as an approximate-mass OOD score instead of likelihood | JEM · CIFAR10 vs CIFAR100 | AUROC: .67 -> .82 | Helped |
| add EBM training to standard classifier training | CIFAR10 classifier · CIFAR10 adversarial robustness | robustness | Helped |
| increase the number of input-seeded SGLD refinement steps (ablation) | JEM · CIFAR10 adversarial robustness | robustness | Helped |
| use semi-supervised EBM training on the unlabeled remainder of the training set (ablation) | JEM · CIFAR10 with 4,000 labeled examples | generalization | No clear effect |
| sample with log-sum-exp energy instead of class-conditional energy (ablation) | JEM · CIFAR10 | accuracy: 91.2% -> 92.9% | Helped |
| factor the objective as p(x|y) + p(y) instead of p(x) + p(y|x) (ablation) | JEM · CIFAR10 | Inception Score: 8.76 -> 6.36 | Hurt |
| factor the objective as p(x|y) + p(y) instead of p(x) + p(y|x) (ablation) | JEM · CIFAR10 | FID: 38.4 -> 61.8 | Hurt |
| use 10 rather than 5 samples in the expectation-over-transformations attack (ablation) | JEM with one refinement step · CIFAR10 adversarial robustness evaluation | robustness | No clear effect |
Closest papers by meaning
- Learning normalized image densities via dual score matching (NeurIPS 2025 · similarity 0.86)
- Dual Energy-Based Model with Open-World Uncertainty Estimation for Out-of-distribution Detection (CVPR 2025 · similarity 0.85)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.84)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.84)
- NRGBoost: Energy-Based Generative Boosted Trees (ICLR 2025 · similarity 0.84)
- Energy-Based Modelling for Discrete and Mixed Data via Heat Equations on Structured Spaces (NeurIPS 2024 · similarity 0.84)
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.