mlsanityData to 7 Oct 2026

Do Deep Generative Models Know What They Don't Know?

arXiv 1810.09136 ↗First posted 22 Oct 2018stat.ML cs.LGcited by 443

Not on the map. Why?

Cited by 46 in Diffusion models, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

A neural network deployed in the wild may be asked to make predictions for inputs that were drawn from a different distribution than that of the training data. A plethora of work has demonstrated that it is easy to find or synthesize inputs for which a neural network is highly confident yet wrong. Generative models are widely viewed to be robust to such mistaken confidence as modeling the density of the input features can be used to detect novel, out-of-distribution inputs. In this paper we challenge this assumption. We find that the density learned by flow-based models, VAEs, and PixelCNNs cannot distinguish images of common objects such as dogs, trucks, and horses (i.e. CIFAR-10) from those of house numbers (i.e. SVHN), assigning a higher likelihood to the latter when the model is trained on the former. Moreover, we find evidence of this phenomenon when pairing several popular image data sets: FashionMNIST vs MNIST, CelebA vs SVHN, ImageNet vs CIFAR-10 / CIFAR-100 / SVHN. To investigate this curious behavior, we focus analysis on flow-based generative models in particular since they are trained and evaluated via the exact marginal likelihood. We find such behavior persists even when we restrict the flows to constant-volume transformations. These transformations admit some theoretical analysis, and we show that the difference in likelihoods can be explained by the location and variances of the data and the model curvature. Our results caution against using the density estimates from deep generative models to identify inputs similar to the training distribution until their behavior for out-of-distribution inputs is better understood.

Key facts. Cited by 443 papers in the map (data through 7 Oct 2026). First posted on arXiv 22 Oct 2018.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 88; 2022: 112; 2023: 79; 2024: 59; 2025: 52; 2026: 53.

443 papers.

Sort:

By area:

  1. 2022DreamFusion: Text-to-3D using 2D Diffusioncited by 3,006
  2. 2022Language Models (Mostly) Know What They Knowcited by 1,783
  3. 2023ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillationcited by 1,207
  4. 2021Generalized Out-of-Distribution Detection: A Surveycited by 796
  5. 2021ReAct: Out-of-distribution Detection With Rectified Activationscited by 486
  6. 2021On the Importance of Gradients for Detecting Distributional Shifts in the Wildcited by 338
  7. 2021Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecastingcited by 332
  8. 2021CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flowscited by 320
  9. 2021SSD: A Unified Framework for Self-Supervised Outlier Detectioncited by 264
  10. 2021MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Spacecited by 233
  11. 2022Delving into Out-of-Distribution Detection with Vision-Language Representationscited by 215
  12. 2023OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detectioncited by 178
38 more
  1. 2021A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challengescited by 154
  2. 2021Student-Teacher Feature Pyramid Matching for Anomaly Detectioncited by 145
  3. 2021OpenGAN: Open-Set Recognition via Open Data Generationcited by 135
  4. 2022How to Exploit Hyperspherical Embeddings for Out-of-Distribution Detection?cited by 128
  5. 2023Score-Based Diffusion Models as Principled Priors for Inverse Imagingcited by 126
  6. 2021Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localizationcited by 116
  7. 2024D-Flow: Differentiating through Flows for Controlled Generationcited by 115
  8. 2022Is Out-of-Distribution Detection Learnable?cited by 109
  9. 2023LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learningcited by 107
  10. 2022Denoising diffusion models for out-of-distribution detectioncited by 102
  11. 2021Can multi-label classification networks know what they don't know?cited by 100
  12. 2021MOOD: Multi-level Out-of-distribution Detectioncited by 94
  13. 2021Understanding Failures in Out-of-Distribution Detection with Deep Generative Modelscited by 89
  14. 2021On Feature Collapse and Deep Kernel Learning for Single Forward Pass Uncertaintycited by 87
  15. 2022GMMSeg: Gaussian Mixture based Generative Semantic Segmentation Modelscited by 87
  16. 2021Types of Out-of-Distribution Texts and How to Detect Themcited by 80
  17. 2021Graph Posterior Network: Bayesian Predictive Uncertainty for Node Classificationcited by 79
  18. 2021Provable Guarantees for Understanding Out-of-distribution Detectioncited by 78
  19. 2022Generating High Fidelity Data from Low-density Regions using Diffusion Modelscited by 78
  20. 2023Video Prediction Models as Rewards for Reinforcement Learningcited by 76
  21. 2021Prior and Posterior Networks: A Survey on Evidential Deep Learning Methods For Uncertainty Estimationcited by 71
  22. 2022Unknown-Aware Object Detection: Learning What You Don't Know from Videos in the Wildcited by 70
  23. 2021On the Impact of Spurious Correlation for Out-of-distribution Detectioncited by 68
  24. 2021Revisiting Mahalanobis Distance for Transformer-Based Out-of-Domain Detectioncited by 67
  25. 2022RbA: Segmenting Unknown Regions Rejected by Allcited by 67
  26. 2021On the Practicality of Deterministic Epistemic Uncertaintycited by 62
  27. 2021Continual Learning via Local Module Compositioncited by 61
  28. 2021A Hierarchical Transformation-Discriminating Generative Model for Few Shot Anomaly Detectioncited by 60
  29. 2021On Memorization in Probabilistic Deep Generative Modelscited by 60
  30. 2023Prediction-Oriented Bayesian Active Learningcited by 60
  31. 2021Hierarchical VAEs Know What They Don't Knowcited by 58
  32. 2022Your ViT is Secretly a Hybrid Discriminative-Generative Diffusion Modelcited by 58
  33. 2023Unsupervised Out-of-Distribution Detection with Diffusion Inpaintingcited by 57
  34. 2021Score-Based Generative Classifierscited by 53
  35. 2022RankFeat: Rank-1 Feature Removal for Out-of-distribution Detectioncited by 52
  36. 2021Rectangular Flows for Manifold Learningcited by 49
  37. 2022Latent Outlier Exposure for Anomaly Detection with Contaminated Datacited by 48
  38. 2022Sample-dependent Adaptive Temperature Scaling for Improved Calibrationcited by 48

50 most cited of 443 All 443 (JSON)