Luiz Manella Pereira

Florida International University

Lecture Information:
  • February 20, 2024
  • 12:30 AM
  • CASE 349 and Zoom

Speaker Bio

Luiz Manella Pereira is a Ph.D. candidate at the Sustainability, Optimization, and Learning for InterDependent networks laboratory (solid lab), Knight Foundation School of Computing and Information Sciences at Florida International University, working under the supervision of Dr. M. Hadi Amini. He is also a Graduate Assistance in Areas of National Need (GAANN) Fellow. His work spans leveraging mathematical tools from algebraic topology, homology, probability theory, and optimal transport to develop novel machine learning techniques and apply them to computer vision, federated learning, and complex network resilience. He received the “2021 Best Journal Paper Award” from SN Operations Research Forum for his work on topological data analysis for network resilience quantification. Prior to his Ph.D., Luiz received a BBA in Finance and a BS in Applied Mathematics where he graduated Magna Cum Laude and was awarded FIU Worlds Ahead Graduate, Dean’s High Achievers Society, and BGS Honor Society.


Optimal transport (OT) has ties to optimization and probability theory, two domains that are critical to machine learning. The main objective of this dissertation is to leverage OT to develop geometry-aware machine learning techniques. First, OT and topological data analysis are integrated to develop a novel approach to quantify resilience in complex networks. This work has broader applications in large-scale critical infrastructure resilience. Next, a WB-based nonparametric classification model for federated learning (FL) is developed. The algorithm is robust against noise when transitioning from a homogeneous to a heterogeneous data distribution. The proposed research established a case for using Wasserstein barycenters as a representation of a cluster of data, which was subsequently leveraged to develop an OT-based preprocessing pipeline. The framework aims to minimize the distribution discrepancy between agents in the network, in turn facilitating the training process. While results were shown for image classification, this framework can be generalized and applied to other FL problems. This dissertation explores how the preprocessing pipeline improves federated learning by aligning data in a privacy-preserving fashion. Most importantly, this dissertation demonstrates the versatility and utility of optimal transport in the field of machine learning, particularly in federated learning.