STAT260: Mean Field Asymptotics in Statistical Learning

Description

Instructor: Song Mei (songmei@berkeley.edu)
Lectures: Monday/Wednesday 10:00-11:30 am (on Zoom)
Office Hours: Tuesday and Friday 4-5 pm (on Zoom)

This course focuses on the computational and statistical aspects of statistical models in the high dimensional asymptotics (the mean-field asymptotics). We will cover a few useful tools including concentration inequalities, replica methods in statistical physics, Gaussian comparison inequalities, moment methods and Stieltjes transforms for random matrices, and approximate message passing algorithms. A few applications of these methods include the spiked matrix model, the LASSO problem, the double-descent phenomenon, random features models, and the phase retrieval problem.

Announcements

• We will use Bcourses for adminisitrative matters.

• We will use Piazza for discussions and questions.

• The recordings can be found here. The recordings will only be available to Berkeley affiliates.

• Zoom link for classes and office hours will be sent through the bCourse announcements. If you would like to sit in the class, please send me an email. The Zoom link is not available to non-Berkeley affiliates.

• Your feedback will be greatly appreciated: Feedback form of Lecture 1-3.

Prerequisite

• Solid background of matrix calculus, probability theory, theoretical statistics, and convex optimization. Some useful prior knowledge includes statistical learning theory.

Grading

• Class attendance is required. Each enrolled student is expected to scribe the notes for at least one lecture, which is due in one week from the lecture. LaTeX template is available here.

• There will be four problem sets.

• No mid/final exam.

• Course project: literature review or original research.

• For pass/no pass students, project is optional (but encouraged).

• Final score will be max{scribe ×10 % plus assignment ×40 % plus course project ×50 %, course project}.

• See project.pdf for more information on the course project.