Bayesian posterior approximation with stochastic ensembles

Abstract

We introduce ensembles of stochastic neural networks to approximate the Bayesian posterior, combining stochastic methods such as dropout with deep ensembles. The stochastic ensembles are formulated as families of distributions and trained to approximate the Bayesian posterior with variational inference. We implement stochastic ensembles based on Monte Carlo dropout, DropConnect and a novel non-parametric version of dropout and evaluate them on a toy problem and CIFAR image classification. For both tasks, we test the quality of the posteriors directly against Hamiltonian Monte Carlo simulations. Our results show that stochastic ensembles provide more accurate posterior estimates than other popular baselines for Bayesian inference.

19 pages, CVPR 2023

References in corpus (6)

• Weight Uncertainty in Neural Networks
• Dropout Inference in Bayesian Neural Networks with Alpha-divergences
• Diversity Matters When Learning From Ensembles
• DropConnect Is Effective in Modeling Uncertainty of Bayesian Deep Networks
• Improving robustness and calibration in ensembles with diversity regularization
• Ensembles for Uncertainty Estimation: Benefits of Prior Functions and Bootstrapping

Cited by in corpus (2)

• Bayesian RG Flow in Neural Network Field Theories
• Looking at the posterior: accuracy and uncertainty of neural-network predictions