Robust Learning Methods for Imprecise Data and Cautious Inference

Developing Machine Learning models that are robust in the face of uncertainty is currently one of the most significant problems for the advancement of Machine Learning in both research and application. In this presentation I will discuss the problem of dealing with a specific form of uncertainty, called imprecision, and summarize my contributions with respect to this topic. I will focus, in particular, on two particular forms of imprecision: input imprecision, which leads to the problem of learning from imprecise data (that is, learning models from incomplete or partially specified datasets), and output imprecision, which leads to the problem of cautious inference (that is, learning models that are able to partially abstain on uncertain instances). For both problems I will describe some algorithmic solutions for common machine learning tasks (including feature selection, classification and ensembling) as well as illustrate their computational properties (computational complexity and learning-theoretic guarantees) and results on synthetic and real-world benchmarks.

NOTE: This presentation is online-only at meet.google.com/jbj-tdsr-aop