Machine Learning: Coping with Non-Stationary Environments and Errors

Course Description

This course teaches machine learning methods that 1) consider data distribution shift and 2) represent and quantify the model uncertainty in a principled way. The topics we will cover include machine learning techniques that deal with data distribution shift, including domain adaptation, domain generalization, and distributionally robust learning techniques, and various uncertainty quantification methods, including Bayesian methods, conformal prediction methods, and model calibration methods. We will introduce these topics in the context of building trustworthy machine learning solutions to safety-critical applications and socially-responsible applications. For example, a typical application is responsible decision-making under uncertainty in non-stationary environments. So we will also introduce concepts like fair machine learning and learning under safety constraints, and discuss how robust and uncertainty-aware learning techniques contribute to such more desired systems. Students will learn the state-of-the-art methods in lectures, test their understanding in homeworks, and apply these methods in a project. Prerequisite(s): Students who have taken, or are currently enrolled in EN.601.687 OR EN.601.787, are not eligible to take EN.601.487.;EN.601.475 OR EN.601.675

Code

EN.601.487

Credits

3

Subject

Computer Science