Research map / Label-efficient and robust learning / OOD and anomaly detection
Energy-based Hopfield Boosting 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 Hopfield Boosting, an outlier-exposure OOD detection method that samples auxiliary outliers near the ID-OOD decision boundary via a modern Hopfield energy-based weighting and trains with an energy-based loss.
Problem
Improving out-of-distribution detection with outlier exposure by finding informative auxiliary outliers that lie near the in-distribution decision boundary.
Approach
Uses a modern Hopfield energy-based boosting energy E_b over ID and auxiliary memories to assign AdaBoost-style sampling weights to auxiliary outliers near the decision boundary, and trains with an energy-based OOD loss plus cross-entropy, with a projection head.
Key results
- Mean FPR95 on CIFAR-10 improves from 2.28 (POEM) to 0.92 for Hopfield Boosting.
- Mean FPR95 on CIFAR-100 improves from 11.76 to 7.94.
- On ImageNet-1K, mean FPR95 improves from 50.74 (POEM) to 36.60, with mean AUROC 91.00.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Hopfield Boosting instead of POEM | OOD detection with ResNet-18 · CIFAR-10 mean OOD | FPR95: 2.28 to 0.92 | Helped |
| use Hopfield Boosting instead of POEM | OOD detection with ResNet-50 · ImageNet-1K mean OOD | FPR95: 50.74 to 36.60 | Helped |
| use Hopfield Boosting instead of the previous best method | OOD detection with ResNet-18 · CIFAR-100 mean OOD | FPR95: 11.76 to 7.94 | Helped |
| add weighted sampling instead of random sampling (ablation) | Hopfield Boosting · CIFAR-10 mean OOD | FPR95: 1.46 to 0.92 | Helped |
| add the projection head while using random sampling (ablation) | Hopfield Boosting without weighted sampling · CIFAR-10 mean OOD | FPR95: 2.06 to 1.46 | Helped |
| add the OOD loss (ablation) | Hopfield Boosting without weighted sampling or projection head · CIFAR-10 mean OOD | FPR95: 50.40 to 2.06 | Helped |
Surprising findings
- All methods, including Hopfield Boosting, achieve perfect FPR95 on LSUN-Resize and iSUN, suggesting these benchmarks may be misleading due to resizing artifacts.
- Hopfield Boosting, like EBO-OE, misclassifies a substantial number of outliers from datasets designed to expose OOD weaknesses as inliers.
Limitations
- Reliance on auxiliary outlier data; selection of AUX data determines decision boundary and may be prohibitive when outliers are scarce.
- Evaluation benchmarks (e.g., LSUN-Resize, iSUN) may be unreliable due to image artifacts, so performance may not reflect general OOD ability.
- Sample weights are computed only for about 40% of ImageNet-21K and with 50,000-pattern memories, which may be noisy.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-RC ImageNet-1K ImageNet-21K MNIST SVHN Textures iSUN Places 365 LSUN SUN
Benchmarks: CIFAR-10 OOD suite (SVHN, LSUN-Crop, LSUN-Resize, Textures, iSUN, Places 365) CIFAR-100 ImageNet-1K OOD suite (Textures, SUN, Places 365, iNaturalist)
Metrics: FPR95 AUROC accuracy
Baselines: DOS DOE DivOE DAL MixOE POEM EBO-OE MSP-OE Hopfield Energy (HE) Simplified Hopfield Energy (SHE)
Builds on: AdaBoost outlier exposure modern Hopfield networks EBO-OE POEM
Closest papers by meaning
- Out-of-Distribution Detection based on In-Distribution Data Patterns Memorization with Modern Hopfield Energy (ICLR 2023 · similarity 0.91)
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · similarity 0.89)
- Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity (NeurIPS 2025 · similarity 0.88)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.87)
- DOS: Diverse Outlier Sampling for Out-of-Distribution Detection (ICLR 2024 · similarity 0.87)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · 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.