Sample-efficient and domain robust machine learning using privileged information

Abstract

In problem domains where data availability is limited, sample-efficient machine learning is critical. Yet, for many learning problems, standard practice routinely leaves substantial information unused. One example is the prediction of an outcome at the end of a time series based on variables collected at a baseline time point, for example, the 30-day risk of mortality for a patient upon admission to a hospital. In applications, it is common that intermediate samples, collected between baseline and endpoints, are discarded, as they are not available as input for prediction when the learned model is used. We say that this information is privileged, as it is available only at training time. In this talk, I show that making use of known causal structure and privileged information from intermediate time series can lead to much more sample efficient learning. I give conditions under which it is provably preferable to classical learning, and a suite of empirical results to support these findings. Additionally, I give preliminary results indicating that learning using privileged information may increase robustness in domain adaptation with examples from medical image classification. 

Presenter

Fredrik Johansson , Chalmers University of Technology