Bayesian Deep Learning

The Bayesian module provides methods for approximate Bayesian inference in neural networks, enabling uncertainty decomposition into epistemic and aleatoric components.

Methods

MCDropout(model[, num_samples, dropout_rate])	Monte Carlo Dropout for uncertainty estimation.
DeepEnsemble(model_fn[, num_models])	Deep Ensembles for uncertainty quantification.
SWAG(model[, num_samples, max_models, var_clamp])	Stochastic Weight Averaging - Gaussian (SWAG).
LaplaceApproximation(model[, likelihood, ...])	Laplace Approximation for Bayesian Neural Networks.

Metrics

predictive_entropy(predictions)	Compute predictive entropy for batched Bayesian predictions (total uncertainty).
mutual_information(predictions)	Compute mutual information (epistemic uncertainty).
disagreement(predictions[, method])	Compute disagreement score for each sample.

Utilities

decompose_uncertainty(predictions)

Decompose predictive uncertainty into epistemic and aleatoric components.