Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Being Bayesian, Even Just a Bit, Fixes Overconfidence in ReLU Networks
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 |
|---|---|---|---|
| Apply last-layer Laplace approximation instead of MAP | Binary ReLU classifiers · far-away uniform-noise inputs (δ = 100), four binary datasets | MMC: 99.9$\pm$0.0 to 79.4$\pm$0.9 (Binary-MNIST); 98.9$\pm$1.0 to 50.6$\pm$0.1 (Binary-CIFAR10) | Helped |
| Apply last-layer Laplace approximation instead of MAP | Binary ReLU classifiers · far-away uniform-noise inputs (δ = 100), four binary datasets | AUR: 0.2$\pm$0.1 to 99.6$\pm$0.1 (Binary-MNIST); 11.3$\pm$10.3 to 99.5$\pm$0.1 (Binary-CIFAR10) | Helped |
| Use temperature scaling instead of MAP | Binary ReLU classifiers · far-away uniform-noise inputs (δ = 100) | MMC: 100.0$\pm$0.0 to 100.0$\pm$0.0 (Binary-MNIST); 98.9$\pm$1.0 to 97.6$\pm$2.2 (Binary-CIFAR10) | No clear effect |
| Use temperature scaling instead of MAP | Binary ReLU classifiers · far-away uniform-noise inputs (δ = 100) | AUR: 0.2$\pm$0.1 to 45.1$\pm$5.8 (Binary-MNIST); 11.3$\pm$10.3 to 0.0$\pm$0.0 (Binary-CIFAR100) | Mixed |
| Apply last-layer Laplace approximation instead of MAP | Multi-class CIFAR10 classifier · CIFAR10 - Noise (δ = 2000) | MMC: 98.7$\pm$0.2 to 17.4$\pm$0.0 | Helped |
| Apply last-layer Laplace approximation instead of MAP | Multi-class CIFAR10 classifier · CIFAR10 - Noise (δ = 2000) | AUR: 10.9$\pm$0.4 to 100.0$\pm$0.0 | Helped |
| Use all-layer diagonal or Kronecker-factored Laplace approximation instead of MAP | Multi-class CIFAR10 classifier · CIFAR10 - Noise (δ = 2000) | MMC: 98.7$\pm$0.2 to 60.7$\pm$2.0 (DLA); 98.7$\pm$0.2 to 61.8$\pm$1.5 (KFLA) | Helped |
| Use all-layer diagonal or Kronecker-factored Laplace approximation instead of MAP | Multi-class CIFAR10 classifier · CIFAR10 - Noise (δ = 2000) | AUR: 10.9$\pm$0.4 to 89.6$\pm$1.1 (DLA); 10.9$\pm$0.4 to 87.6$\pm$0.9 (KFLA) | Helped |
| Use LLLA rather than all-layer DLA or KFLA | Multi-class CIFAR10 classifier · CIFAR10 - Noise (δ = 2000) | AUR: 89.6$\pm$1.1 to 100.0$\pm$0.0 (DLA); 87.6$\pm$0.9 to 100.0$\pm$0.0 (KFLA) | Helped |
Closest papers by meaning
- An Infinite-Feature Extension for Bayesian ReLU Nets That Fixes Their Asymptotic Overconfidence (NeurIPS 2021 · similarity 0.95)
- On the Expressiveness of Approximate Inference in Bayesian Neural Networks (NeurIPS 2020 · similarity 0.91)
- Precise characterization of the prior predictive distribution of deep ReLU networks (NeurIPS 2021 · similarity 0.90)
- Do Bayesian Neural Networks Actually Behave Like Bayesian Models? (ICML 2025 · similarity 0.89)
- Posterior Refinement Improves Sample Efficiency in Bayesian Neural Networks (NeurIPS 2022 · similarity 0.89)
- A Study of Bayesian Neural Network Surrogates for Bayesian Optimization (ICLR 2024 · similarity 0.88)
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.