Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Amortized Conditional Normalized Maximum Likelihood: Reliable Out of Distribution Uncertainty Estimation
Aurick Zhou, Sergey Levine
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 |
|---|---|---|---|
| Use ACNML rather than Bayesian model averaging with the same posterior | Bayes-by-Backprop on rotated MNIST · Rotated MNIST, severe rotations | ECE | Improved |
| Use ACNML rather than Bayesian model averaging with the same posterior | Bayes-by-Backprop on rotated MNIST · MNIST in-distribution and lowest rotation levels | ECE | Hurt |
| Use ACNML rather than MAP and Bayesian baselines | VGG16 on CIFAR10-C · CIFAR10-C, more severe corruptions | ECE | Improved |
| Use ACNML rather than MAP and Bayesian baselines | WideResNet28x10 on CIFAR10-C · CIFAR10-C, heavier corruptions | ECE | Improved |
| Use ACNML rather than the corresponding Bayesian methods | VGG16 and WideResNet28x10 on CIFAR10-C · CIFAR10-C, more severe corruptions | Brier score | Improved |
| Use ACNML rather than the corresponding Bayesian methods | VGG16 and WideResNet28x10 on CIFAR10-C · CIFAR10-C, all corruption severities | accuracy | No clear effect |
| Use ACNML rather than naive CNML | MNIST MLP · MNIST inference | inference time per input: 13.83s -> 0.08s | Improved |
| Use ACNML rather than naive CNML | VGG16 · CIFAR10 inference | inference time per input: 102.0s -> 0.37s | Improved |
| Use ACNML rather than naive CNML | WRN28x10 · CIFAR10 inference | inference time per input: 359.1s -> 1.1s | Improved |
| Use ACNML with KFAC-Laplace instead of diagonal SWAG posterior (ablation) | VGG16 on CIFAR10 · CIFAR10 test | ECE: 0.0115±0.0010 -> 0.0361± 0.0016 | Hurt |
| Use ACNML-SWAGD instead of SWAGD Bayesian model averaging (ablation) | VGG16 on CIFAR10 · CIFAR10 test | ECE: 0.0284±0.0002 -> 0.0115±0.0010 | Improved |
| Use ACNML-SWAGD instead of SWAGD Bayesian model averaging (ablation) | WideResNet28x10 on CIFAR10 · CIFAR10 test | ECE: 0.0042 ± 0.0003 -> 0.0122 ± 0.0006 | Hurt |
Closest papers by meaning
- Bayesian Uncertainty Estimation for Batch Normalized Deep Networks (ICML 2018 · similarity 0.89)
- Density-Softmax: Efficient Test-time Model for Uncertainty Estimation and Robustness under Distribution Shifts (ICML 2024 · similarity 0.86)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.86)
- Uncertainty Quantification for Deep Regression using Contextualised Normalizing Flows (NeurIPS 2025 · similarity 0.86)
- Neural Conditional Probability for Uncertainty Quantification (NeurIPS 2024 · similarity 0.85)
- Uncertainty Quantification with the Empirical Neural Tangent Kernel (NeurIPS 2025 · similarity 0.85)
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.
