Bayesian Conditional Density Estimation for Active Learning


Conor Durkan

Uncertainty in probabilistic models of data can decomposed into two parts: epistemic uncertainty, corresponding to uncertainty about the model, and aleatoric uncertainty, corresponding to inherent stochasticity in the data we are modeling.
Bayesian conditional density estimation provides a principled way to disentangle these sources of uncertainty, leveraging flexible neural density estimators and Bayesian inference methods for these networks. With this approach, we propose heuristics for active learning, exploiting uncertainty in our model to learn in a data-efficient manner.