Course Overview

Overview

Why these ideas matter? Understanding machine learning requires more than knowing how to train models or tune hyperparameters. The ideas developed in this course emphasize learning as one component of intelligence, embedded within perception, representation, and action. This perspective clarifies why data alone is insufficient, why generalization—not memorization—is the central challenge, and why model complexity must be carefully controlled. Approaching machine learning through foundational principles allows us to reason about when algorithms work, when they fail, and how to design better ones, rather than treating models as black-box tools.

Course goals This course will take off from around the late 1990s, beginning with kernel methods, and develop ideas in deep learning that bring us to the present day. Our goals are to become proficient in using modern machine learning tools—implementing them, training them, and modeling real problems using ML ideas; understand why many seemingly ad hoc or quixotic ideas in machine learning actually work. After completing this course, students should be able not only to apply machine learning methods, but more importantly to improve existing approaches using a foundational mathematical understanding, and to develop new ideas that advance machine learning theory and practice.