Coursebooks 2017-2018

PDF
 

Mathematics of data: from theory to computation

EE-556

Lecturer(s) :

Cevher Volkan

Language:

English

Summary

This course reviews recent advances in convex optimization and statistical analysis in the wake of Big Data. We provide an overview of the emerging convex formulations and their guarantees, describe scalable solution techniques, and illustrate the role of parallel and distributed computation.

Content

The course consists of the following topics

Lecture 1:    'Objects in Space': Definitions of norms, inner products, and metrics for vector, matrix and tensor objects. Basics of complexity theory.

Lecture 2:    Maximum likelihood principle as a motivation for convex optimization. Fundamental structures in convex analysis, such as cones, smoothness, and conjugation.

Lecture 3:    Unconstrained, smooth minimization techniques. Gradient methods. Variable metric algorithms. Time-data tradeoffs in ML estimation.

Lecture 4:    Convex geometry of linear inverse problems. Structured data models (e.g., sparse and low-rank) and convex gauge functions and formulations that encourage these structures.  Computational aspects of gauge functions.

Lecture 5:    Composite convex minimization. Regularized M-estimators. Time-data tradeoffs in linear inverse problems.

Lecture 6:    Convex demixing. Statistical dimension. Phase transitions in convex minimization. Smoothing approaches for non-smooth convex minimization.

Lecture 7:    Constrained convex minimization-I. Introduction to convex duality. Classical solution methods (the augmented Lagrangian method, alternating minimization algorithm, alternating direction method of multipliers, and the Frank-Wolfe method) and their deficiencies

Lecture 8:    Constrained convex minimization-II. Variational gap characterizations and dual smoothing. Scalable, black-box optimization techniques. Time data-tradeoffs for linear inverse problems.

Lecture 9:    Classical black-box convex optimization techniques. Linear programming, semidefinite programming, and the interior point method (IPM). Hierarchies of classical formulations. Time and space complexity of the IPM.

Lecture 10: Time-data tradeoffs in machine learning.

Lecture 11: Convex methods for Big Data I: Randomized coordinate descent methods. The Page Rank problem and Nesterov's solution. Composite formulations.

Lecture 12: Convex methods for Big Data II: Stochastic gradient descent methods. Least squares: conjugate gradients vs. a simple stochastic gradient method. Dual and gradient averaging schemes. Stochastic mirror descent.

Lecture 13: Randomized linear algebra routines for convex optimization. Probabilistic algorithms for constructing approximate low-rank matrix decompositions. Subset selection approaches. Theoretical approximation guarantees.

Lecture 14: Role of parallel and distributed computing. How to avoid communication bottlenecks and synchronization. Consensus methods. Memory lock-free, decentralized, and asynchronous algorithms.   

 

Learning Prerequisites

Important concepts to start the course

Previous coursework in calculus, linear algebra, and probability is required.

Familiarity with optimization is useful. 

 

Learning Outcomes

By the end of the course, the student must be able to:

Assessment methods

Homework assignments. (Continuous control)

In the programs

Reference week

 MoTuWeThFr
8-9     
9-10     
10-11 BC01   
11-12    
12-13     
13-14     
14-15     
15-16     
16-17    CE1103
CO4
CO5
INM202
17-18    
18-19     
19-20     
20-21     
21-22     
 
      Lecture
      Exercise, TP
      Project, other

legend

  • Autumn semester
  • Winter sessions
  • Spring semester
  • Summer sessions
  • Lecture in French
  • Lecture in English
  • Lecture in German