Multivariate Bayesian Last Layer for Regression: Uncertainty Quantification and Disentanglement
This work addresses uncertainty quantification and domain transfer in regression tasks, but it is incremental as it builds on existing Bayesian Last Layer methods by extending them to multivariate settings.
The authors tackled multivariate regression under heteroscedastic noise by proposing Bayesian Last Layer models, which enable uncertainty quantification with a single forward pass and disentangle aleatoric and epistemic uncertainty, allowing transfer to new data domains with uncertainty-aware capability.
We present new Bayesian Last Layer models in the setting of multivariate regression under heteroscedastic noise, and propose an optimization algorithm for parameter learning. Bayesian Last Layer combines Bayesian modelling of the predictive distribution with neural networks for parameterization of the prior, and has the attractive property of uncertainty quantification with a single forward pass. The proposed framework is capable of disentangling the aleatoric and epistemic uncertainty, and can be used to transfer a canonically trained deep neural network to new data domains with uncertainty-aware capability.