Bayesian transfer learning between Gaussian process regression tasks

Bayesian knowledge transfer in supervised learning scenarios often relies on a complete specification and optimization of the stochastic dependence between source and target tasks. This is a critical requirement of completely modelled settings, which can often be difficult to justify. We propose a strategy to overcome this. The methodology relies on fully probabilistic design to develop a target algorithm which accepts source knowledge in the form of a probability distribution. We present this incompletely modelled setting in the supervised learning context where the source and target tasks are to perform Gaussian process regression. Experimental evaluation demonstrates that the transfer of the source distribution substantially improves prediction performance of the target learner when recovering a distorted nonparametric function realization from noisy data.