Bayesian Task-Level Transfer Learning for Non-linear Regression

Pei Yang, Qi Tan, Yehua Ding

Abstract: Multi-task learning utilizes labeled data from other ldquosimilarrdquo tasks and can achieve efficient knowledge-sharing between tasks. Previous research mainly focused on multi-task learning for linear regression. In this paper, a novel Bayesian multi-task learning model for non-linear regression, i.e. HiRBF, is proposed. HiRBF is constructed under a hierarchical Bayesian framework. In the model all tasks are combined in a single RBF network. The input-to-hidden weights are shared between tasks, and the hidden-to-output weights are assumed to be sampled randomly from a certain prior distribution. The HiRBF algorithm is compared with two transfer-unaware approaches. The experiments demonstrate that HiRBF significantly outperforms the others.