Robust learning methods for out-of-distribution tasks

Statistical learning methods leverage past training data to perform tasks that either assist or automate human decision-making.

Most methods are built on the assumption that data distributions from the training environment and future test environments will match each other.In practice, however, many tasks will be out-of-distribution, that is, data distributions from the training and test environments will diverge from each other.

In such cases, the task performance of state-of-the-art learning methods will degrade and may result in erroneous decisions.

Recent advances in distributionally robust optimization have led to improved generalization properties of learning methods using finite samples.

While demonstrating promising results, these techniques are developed for in-distribution tasks.This research project will develop new robust learning theory and methods for out-of-distribution tasks.

Specifically, it will focus on achieving robustness with respect to structured distributional divergences that arise in three different problems: corrupted training data, missing covariate test data and interventionally shifted test data.

These appear in a wide range of applications, including wirless receiver localization, health diagnostics, and medical decision support.

Through a cross-fertilization of ideas and tools from different branches of machine learning, statistics and optimization, the project will advance the generalization capabilities of statistical learning methods.