Fall 2026 Statistics GR8201 section 002

TOPICS IN THEORETICAL STATISTICS - STAT

Theoretical Statistics

Call Number 20973
Day, Time & Location View Class Schedule & Location in Vergil
Points 3
Grading Mode Pass/Fail
Approvals Required None
Instructor Sumit Mukherjee
Type SEMINAR
Method of Instruction In-Person
Course Description

This six-week course provides an introduction to the mathematical and statistical foundations of deep learning. It focuses on three central components of the theory: approximation, generalization, and optimization. The first part studies the expressive power of neural networks and examines how their approximation
properties depend on network depth, width, activation functions, and structural assumptions on the target function. The second part develops statistical risk bounds for neural-network estimators. Particular emphasis is placed on the formulation of prediction problems for which neural networks can provably exploit compositional, hierarchical, or l ow-dimensional structure. This problem-oriented perspective complements approaches that focus primarily on explaining the behavior of a fixed
architecture.

The course further i ntroduces statistical perspectives on generative modeling, including score-based, dilusion, and flow-based methods. The final part covers gradient-based optimization, overparameterization, neural tangent kernels, feature learning, and implicit regularization. Throughout the course, the connections between approximation, statistical estimation, and optimization will be emphasized. The course ends with discussing future research directions.

Department Statistics
Enrollment 0 students (55 max) as of 6:06PM Thursday, November 19, 2026
Subject Statistics
Number GR8201
Section 002
Division Graduate School of Arts and Sciences
Open To Engineering:Graduate, GSAS, Public Health
Note PhD STUDENTS ONLY; Priority Registration for STAT and ENGI s
Section key 20263STAT8201G002