Research map / Probabilistic and scientific modeling / Variational and generative sampling
Why Normalizing Flows Fail to Detect Out-of-Distribution Data
Polina Kirichenko, Pavel Izmailov, Andrew G Wilson
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 RealNVP and Glow on raw image pixels using maximum likelihood | RealNVP and Glow · CIFAR-10, CelebA, and SVHN OOD pairs | AUROC: 56.0; 6.0; 98.4; 67.7; 6.3; 99.0 | Mixed |
| Train the flow on EfficientNet semantic image embeddings rather than raw pixels | RealNVP · CelebA, CIFAR-10, and SVHN embeddings | AUROC: 99.99; 99.99; 99.99; 73.31; 100.0; 99.98 | Improved |
| Use a cycle-mask instead of checkerboard or horizontal masks (ablation) | RealNVP · FashionMNIST (OOD: MNIST) | relative likelihood ranking | Improved |
| Use a cycle-mask instead of checkerboard or horizontal masks (ablation) | RealNVP · CelebA (OOD: SVHN) | relative likelihood ranking | Mixed |
| Add a low-dimensional bottleneck to the st-networks (ablation) | RealNVP · FashionMNIST (OOD: MNIST) | relative likelihood ranking | Improved |
| Add a low-dimensional bottleneck to the st-networks (ablation) | RealNVP · CelebA (OOD: SVHN) | relative likelihood ranking | Mixed |
| Train with a maximum-likelihood objective that also minimizes likelihood on a specified OOD dataset | RealNVP · CIFAR-10 vs CelebA, CIFAR-10 vs SVHN, and FashionMNIST vs MNIST | likelihood separation | Improved |
| Minimize likelihood on one selected OOD dataset | RealNVP · Other unseen OOD datasets | likelihood separation | No clear effect |
Closest papers by meaning
- Understanding Likelihood of Normalizing Flow and Image Complexity through the Lens of Out-of-Distribution Detection (AAAI 2024 · similarity 0.93)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.88)
- Robustly Train Normalizing Flows via KL Divergence Regularization (AAAI 2024 · similarity 0.87)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.86)
- Out-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothness (ICLR 2024 · similarity 0.86)
- Mahalanobis++: Improving OOD Detection via Feature Normalization (ICML 2025 · similarity 0.86)
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.
