Robust Bayesian Transfer Learning between Autoregressive Inference Tasks

Abstract

Bayesian transfer learning typically relies on a complete stochastic dependence specification between source and target learners. We advocate a solution to the Bayesian transfer learning paradigm which adopts Fully Probabilistic Design (FPD) to search for an optimal choice of distribution constrained by probabilistic source knowledge. Using this optimal decision-making strategy, an algorithm for accepting source knowledge is identified but is found to be effectively insensitive to source uncertainty. Therefore, we propose an adaptation of the FPD framework which results in a robust transfer learning algorithm.Experimental evidence gathered via synthetic data shows enhanced performance when employing both optimal algorithms in a low source data predictor variance regime. In a high source data predictor variance setting, only our adapted FPD-optimal algorithm achieves robustness.