STAT547U: Topics in Deep Learning Theory

Jan-Feb 2025

Schedule

Subject to change; check announcements and Canvas!.

Class/DateClass TopicReading DueResources
Part 0: Introduction
Diagnostic problem set due on Tues, Jan 14. (TeX template and math_commands.tex macros file. )
Class 1 (Tue, Jan 07)Course logistics, introduction, failures of classical learning theoryLecture Notes
Class 2 (Thu, Jan 09)Introduction to functional analysis, reproducing kernel Hilbert spacesLecture Notes
Class 3 (Tue, Jan 14)Double descent, implicit bias of gradient descent for regressionFit Without Fear (Sections 1-3.9)
(Required reading, but no reading summary due.)
Lecture Notes
Part 1: Overparameterized linear regression
Class 4 (Thu, Jan 16)Introduction to high-dimensional asymptotics and random matrix theoryHigh Dimensional Regression (Section 2)Lecture NotesDemo
Class 5 (Tue, Jan 21)Effective regularization, risk of overparameterized ridge regressionHigh Dimensional Regression (Section 4)Lecture Notes
Class 6 (Thu, Jan 23)Benign overfittingBenign, Tempered, or Catastrophic: A Taxonomy of Overfitting (Section 2-3)
Part 2: Approximating neural networks as kernel machines
Final project intermediate check-in due on Tues, Feb 04.
Class 7 (Tue, Jan 28)Infinite width NNs: Neural Tangent Kernel (Pt 1)Deep Learning Theory Lecture Notes (Ch. 4)
Class 8 (Thu, Jan 30)Infinite width NNs: Neural Tangent Kernel (Pt 2)Understanding the Neural Tangent Kernel
Class 9 (Tue, Feb 04)Feature learning, rich vs lazy regimesThe lazy (NTK) and rich (µP) regimes: A gentle tutorial
(Optional reading; no reading summary due.)
Part 3: Advanced topics
Class 10 (Thu, Feb 06)Neural networks for classificationDeep Learning Theory Lecture Notes (Ch. 10)
Class 11 (Tue, Feb 11)Effects of depth
Class 12 (Thu, Feb 13)Topic TBD (scaling laws, in-context learning, etc.)
Feb 14: oral presentations for final paper reading assignment