Research map / Label-efficient and robust learning / OOD and anomaly detection
Understanding Anomaly Detection with Deep Invertible Networks through Hierarchies of Distributions and Features
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 |
|---|---|---|---|
| Replace raw full-image likelihood with the likelihood ratio against Tiny-Glow | Glow anomaly detection · CIFAR10 in-distribution; SVHN, CIFAR100, LSUN outliers | AUROC: 8.8 -> 93.9; 51.7 -> 66.8; 69.3 -> 89.2 | Helped |
| Use PNG instead of Tiny-Glow as the general reference | Glow likelihood-ratio anomaly detection · SVHN in-distribution; CIFAR10, CIFAR100, LSUN outliers | AUROC: 98.3 -> 74.4; 97.9 -> 79.5; 99.6 -> 96.8 | Hurt |
| Use a matching model family for the general and in-distribution likelihoods | Likelihood-ratio anomaly detection · CIFAR10 and CIFAR100 outliers | AUROC: CIFAR10: 66.8 vs. 53.4; CIFAR100: 52.8 vs. 54.2 | Mixed |
| Add the proposed outlier loss during in-distribution model training (ablation) | Glow anomaly detection · CIFAR10 in-distribution; SVHN, CIFAR100, LSUN outliers | AUROC: 93.9 -> 98.6; 66.8 -> 84.5; 89.2 -> 94.1 | Helped |
| Replace the proposed outlier loss with margin loss (ablation) | Glow anomaly detection · CIFAR10 in-distribution; SVHN, CIFAR100, LSUN outliers | AUROC: 93.9 -> 96.5; 66.8 -> 71.1; 89.2 -> 75.7 | Mixed |
| Also train Tiny-Glow with in-distribution samples as outliers (ablation) | Glow likelihood-ratio anomaly detection · CIFAR10 in-distribution; SVHN, CIFAR100, LSUN outliers | AUROC: 98.6 -> 99.0; 84.5 -> 86.8; 94.1 -> 95.8 | Helped |
| Use only the final 4x4 scale contribution instead of full Glow likelihood | Glow anomaly detection · CIFAR10 in-distribution; SVHN, CIFAR100, LSUN outliers | AUROC: 8.8 -> 92.9; 51.7 -> 60.0; 69.3 -> 82.8 | Helped |
| Replace convolutional Glow blocks with fully connected blocks (ablation) | Glow · Fashion-MNIST vs. MNIST | AUROC: 15% -> 81% | Helped |
| Use a BRATS-trained reference instead of raw likelihood | Glow anomaly detection · BRATS MRI modalities | AUROC: 53.3% -> 78.3% | Helped |
| Use a class-conditional model and treat other classes as outliers | Glow anomaly detection · CIFAR10 and CIFAR100; mean across outlier datasets | AUROC: 84.7 -> 87.0; 77.0 -> 80.8 | Helped |
Closest papers by meaning
- Deep Anomaly Detection with Outlier Exposure (ICLR 2019 · similarity 0.89)
- Further Analysis of Outlier Detection with Deep Generative Models (NeurIPS 2020 · similarity 0.89)
- ODIM: Outlier Detection via Likelihood of Under-Fitted Generative Models (ICML 2024 · similarity 0.88)
- ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.88)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.87)
- RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 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.