Stay Informed:
Baskin Engineering COVID-19 Information and Resources
Campus Roadmap to Recovery
Zoom Links: Zoom Help | Teaching with Zoom | Zoom Quick Guide

Optimal Transport in Machine Learning and Computer Vision

Speaker Name: 
Soheil Kolouri
Speaker Title: 
Principal Scientist
Speaker Organization: 
HRL Laboratories
Start Time: 
Monday, February 22, 2021 - 11:00am
End Time: 
Monday, February 22, 2021 - 12:15pm
Via Zoom Link:
Lise Getoor


Modeling and comparing probability distributions appears in many machine learning (ML) and computer vision problems, from 2D and 3D shape and texture analysis in medical images, to deep predictive and generative modeling, domain adaptation, and continual learning. Optimal transport (OT) provides a rich theoretical foundation for comparing probability distributions while respecting their underlying geometry, making it a natural choice for many ML applications. Wasserstein distances are a prominent example of geometric tools that arise from the OT theory. In this talk, I will develop a set of computational tools rooted in OT that allow for efficient comparison of probability distributions and show their applications in explainable, label efficient, and continual ML. First, I will go through a brief introduction to OT and Wasserstein distances. Second, I will focus on three main research thrusts: (i) explainable ML on medical images, in which I will introduce Transport-Based Morphometry that allows for exploratory and discriminative analysis of nonlinear variations in data, (ii) unsupervised deep representation learning where I will present generalized sliced-Wasserstein distances as a useful tool for training deep generative models, and (iii) continual learning where I will focus on overcoming catastrophic forgetting in deep neural networks.  Finally, I will conclude my talk with future research directions and their broader impact.  


Soheil Kolouri is an IEEE senior member and a principal scientist at HRL Laboratories, Malibu, CA. The overarching goal of his research is to design the next generation ML algorithms to solve real-world problems.  He is currently the principal investigator on DARPA’s Learning with Less Labels (LwLL) and the co-investigator on DARPA’s Lifelong Learning Machines (L2M) programs. He received his doctorate in Biomedical Engineering from Carnegie Mellon University (CMU) in 2015. At CMU, he received the Bertucci Fellowship Award for exceptional graduate students in 2014 and the Outstanding Dissertation Award in 2015.

Zoom Link:

Event Type: