Statistical Error Analysis of HTL Algorithm via LTO and RO Stability Hypothesis

摘要

There axe abundant parameters to estimate if training and testing samples are drawn from the different probability distribution in supervised machine learning algorithm setting. Hypothesis transfer learning is a trick to get over this limitation in practical engineering, and the source hypotheses trained on a source domain can only be utilized. In this paper, we define the LTO and RO stability hypothesis for hypothesis transfer learning algorithm, and the statistical error is bounded in terms of these stability hypotheses.