Honors Statistical Machine Learning

This is a rigorous introduction to machine learning from a statistical perspective. While we will cover many of the same introductory elements of machine learning as courses in other departments, the statistical perspective emphasizes the distinction between spurious trends or patterns observed in data, and more stable patterns present in the actual population the data is drawn from. For instance, in prediction problems, while two variables might appear related in observed data, such relation might not be generalizable to the population. Such statistical perspective on ‘generalization’ from sample to population is fundamental to the design of prediction algorithms in modern machine learning, in addition to computational constraints. The course aims to explain how ‘generalization’ together with ‘computation’ drives every aspect of machine learning, from modeling assumptions, to common optimization procedures.  

Major families of algorithms will be covered, from unsupervised procedures for clustering, to supervised procedures for classification and regression, along with an introduction to common optimization techniques. The course requires a good preparation in calculus up to multivariate calculus, and good understanding of linear algebra, and familiarity with basic probability and statistics. At the end of the course, students would be expected to have gained a sense of common approaches in ML, and importantly, the assumptions (on the data and the population) under which such approaches operate.