Instructor
Prof. Volkan Cevher
Description
This course describes theory and methods to address three key challenges in data sciences: estimation, prediction, and computation. We use convex analysis and methods as a common connecting theme, and illustrate the main ideas on concrete applications from machine learning and signal processing.
Learning outcomes
By the end of the course, students must be able to follow up advanced research in machine learning and optimization.
Prerequisites
Multivariable calculus, linear algebra and probability theory.
Outline
The course consists of the following topics
|
Overview of the course. Learning-based compressive subsampling. Introduction to submodularity, examples, and properties. | |
|
Discrete optimization. Submodular function maximization. The greedy algorithm. Relations between submodularity and convexity. The Lovazs extension. | Scribe |
|
Introduction to structured sparsity. Convex relaxation by biconjugation. Structured sparsity via submodular functions. | Scribe |
|
Integer and linear programming. Structured sparsity via totally unimodular constraints. Robust submodular function maximization. The Saturate algorithm. | Scribe |
|
Stochastic optimization. Stochastic subgradient method. Stochastic proximal method. Stochastic accelerated methods. | Scribe |
|
Variance reduction. Coordinate descent methods for smooth objectives. | Scribe |
Lecture 7 | Coordinate descent methods for composite functions. Coordinate descent primal-dual algorithms. Randomized Linear Algebra. Randomized matrix decompositions. Comparison to classical methods. | Scribe |
Lecture 8 | Randomized Linear Algebra. Row extraction method. Power method. Column selection methods. Stochastic quasi-Newton Method. | Scribe |
Lecture 9 | Introduction to statistical learning theory. Statistics vs statistical learning. Approximation error. Excess risk, Hoeffding’s inequality. | Scribe |
Lecture 10 | Concentration of measure inequalities. Chernoff-type bounds. Azuma-Hoeffding inequality. Bounded differences inequality. | Scribe |
Lecture 11 | Uniform Convergences in Statistical Learning Theory. Classical VC Theory for Binary Classification. | Scribe |
Lecture 12 | Project presentations |