Abstract

Learning using privileged information and distillation are powerful machinelearning frameworks that allow a machine learning model to be learned from anexisting model or from a classifier trained over another feature space.Existing implementations of learning using privileged information are limitedto classification only. In this work, we have proposed a novel meta-lossfunction that allows the general application of learning using privilegedinformation and distillation to not only classification but also regression andother related problems. Our experimental results show the usefulness of theproposed scheme.

 

Quick Read (beta)