Output details
11 - Computer Science and Informatics
University of Exeter
A Bayesian framework for active learning
<24> Active learning approaches mitigate the cost associated with obtaining data samples by actively selecting which data to sample. This paper presents the first Bayesian framework for probabilistic active learning of non-separable data. Results show that the framework, which incorporates a query density to explicitly model the sampling process, and crucially makes no assumption about independence between queried data points, performs significantly better than passive learning (even when the overlap regions are wide). It has led to BBSRC and EPSRC proposals with engineers and biologists on design of experiments problems and surrogate modelling for evolutionary optimisation.