Hypothesis Disparity Regularized Mutual Information Maximization

We propose a hypothesis disparity regularized mutual information maximization (HDMI) approach to tackle unsupervised transfer---as an effort towards unifying transfer learning (HTL) and domain adaptation (UDA)---where the knowledge from source is transferred solely through hypotheses adapted target in manner. In contrast prevalent HTL UDA approaches that typically use single hypothesis, HDMI employs multiple leverage underlying distributions of hypotheses. To better utilize crucial relationship among different hypotheses---as opposed unconstrained optimization each independently---while adapting unlabeled maximization, incorporates regularization coordinates jointly learn representations while preserving more transferable with better-calibrated prediction uncertainty. achieves state-of-the-art performance on benchmark datasets for context HTL, without need access data during adaptation.