Optimal transport and applications

Abstract:

Optimal transport is a mathematical discipline with numerous interesting applications. While traditionally it has found applications in civil engineering (e.g. optimal mass transport) and economics (e.g. traveling salesman problems), optimal transport modeling techniques have recently begun being widely adopted in numerous applications including: signal/image processing, machine learning, control, data science and others. This course will introduce students to the mathematics of optimal transport and its modern applications in signal/image processing and machine learning. Specifically, we will describe both the Monge & Kantorovich formulations for the optimal transport problem, solution methods, Wasserstein distances and geometry. We will also describe the concept of transport embeddings, representations, and transforms using linearized optimal transport. Finally we will describe applications of this theory to problems related to data classification, signal estimation and image modeling. Results using real data, together with software, will be demonstrated.

Meetings: Tuesdays, 4-6:20 pm, MR 5 1041

Instructors: Gustavo K. Rohde (gustavo@virginia.edu), Ivan V. Medri (pzr7pr@virginia.edu)

Office hours: Wednesdays 13:30-4:30 hs, MR4 1116A (GKR), TBD (IVM) Mondays 14:00-16:00 hs, MR4 1181

Syllabus

Link for scribe schedule (Link to edit overleaf)

Content Overview

1. Optimal transport theory: 

o measures, transport maps/plans, push forward, Monge and Kantorovich optimal transport, static & dynamic formulations, Wasserstein distances, geometry,  sliced Wasserstein distances


 2. Numerical methods:  

o Sorting, linear programming solutions, entropic regularization, PDE solutions


3. Transport representations/embeddings: 

o Mathematical transforms, linearized optimal transport (LOT), cumulative distribution transform (CDT), signed CDT, Radon CDT, extensions

 

4. Applications: 

o Detection, estimation, classification, inverse problems, transport-based morphometry, system identification, image and signal processing, machine learning



Weakly plan

Week 1 (8/27) Introduction

Week 2 (9/03) Optimal transport, measure theory

Week 2 (9/10) Optimal transport

Week 4 (9/17) Optimal transport

Week 5 (9/24) 1D Optimal transport

   Week 6 (10/01) Discrete OT solvers

   Week 7 (10/08) Distances, dynamic formulation

   Week 8 (10/15) Exam 1

   Week 9 (10/21) Dynamic formulation, Brenier theorem 

   Week 10 (10/28) 1D Transport embeddings 

   Week 11 (11/12) Transport transforms and applications 

   Week 12 (11/19) Transport transforms and applications Continuation

   Week 13 (11/26) Kantorovich problem & Sinkhorn solutions