Watch a 55-minute lecture from Harvard CMSA's GRAMSIA series where Oxford's Patrick Rebeschini explores the relationship between uniform convergence and implicit regularization in learning algorithms. Delve into the statistical analysis of early-stopped mirror descent applied to unregularized empirical risk with squared loss for linear models and kernel methods. Discover how the potential-based analysis of mirror descent from optimization theory connects to uniform learning, and learn how the path traced by mirror descent can be characterized through localized Rademacher complexities. Understand how these complexities are influenced by various factors including mirror map selection, initialization point, step size, and iteration count.
Implicit Regularization via Uniform Convergence in Linear Models and Kernel Methods
Harvard CMSA via YouTube
Overview
Syllabus
Patrick Rebeschini | Implicit regularization via uniform convergence
Taught by
Harvard CMSA
Related Courses
-
Towards Understanding the Implicit Regularization Effect of SGD
-
Implicit and Explicit Regularization in Deep Neural Networks
-
Implicit Regularization for General Norms and Errors - Lorenzo Rosasco, MIT
-
Generalization in Deep Learning Through the Lens of Implicit Rank Minimization
-
Random Initialization and Implicit Regularization in Nonconvex Statistical Estimation - Lecture 2
-
Statistical Learning Theory and Neural Networks
