BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:America/New_York X-WR-TIMEZONE:America/New_York BEGIN:VEVENT UID:822@fds.yale.edu DTSTART;TZID=America/New_York:20241216T130000 DTEND;TZID=America/New_York:20241216T140000 DTSTAMP:20250916T142145Z URL:https://fds.yale.edu/events/fds-seminar-kumar-kshitij-patel-ttic/ SUMMARY:FDS Seminar: Kumar Kshitij Patel (TTIC) DESCRIPTION:\n"Re-inventing Machine Learning for Multiple Distributions: O ptimization\, Privacy\, and Incentives"\n\n\n\nAbstract: Federated Learni ng (FL) has emerged as a transformative framework for multi-distribution l earning\, driving breakthroughs in healthcare\, research\, finance\, and c onsumer technologies. FL enables agents to train models on private data wi thout sharing raw information\, offering a basic yet crucial step toward s afeguarding privacy while complying with regulations like GDPR. However\, the practical deployment of FL poses significant challenges\, including da ta heterogeneity\, high communication costs\, partial client participation \, privacy concerns\, and strategic behavior. In this talk\, I will presen t my research addressing these challenges and bridging the gap between the theory and practice of FL and\, more broadly\, multi-distribution learnin g.\n\n\n\nIn the first part of the talk\, I will examine the optimization challenges in FL. Specifically\, I will outline conditions under which loc al update algorithms—the most widely used optimization methods in FL—o utperform traditional distributed optimization approaches. By formalizing notions of data heterogeneity\, I provide practical insights into selectin g the appropriate algorithm for real-world applications. I will then explo re how personalization mitigates heterogeneity's adverse effects and discu ss its application to private image generation using personalized diffusio n models. I will also briefly touch on my work in the online setting and h ow performative effects can further complicate optimization.\n\n\n\nThe se cond part of the talk will focus on incentive design and fairness in colla borative learning. I will discuss two natural approaches to fairness for m ulti-distribution learning and illustrate their implications using my own work. I will also address client defections\, a critical issue undermining FL's effectiveness\, and introduce a novel algorithm that prevents defect ions by ensuring sustainable collaboration. Finally\, I will outline my vi sion for future research on data markets and economic frameworks that ince ntivize equitable and efficient data sharing.\n\n\n\nBio: \;Kumar Kshi tij Patel is a PhD candidate at the Toyota Technological Institute at Chic ago (TTIC)\, where he is advised by Professors Nathan Srebro and Lingxiao Wang. His research focuses on federated learning\, privacy\, optimization\ , and fairness\, with an emphasis on addressing challenges related to data heterogeneity and decentralized collaboration. Kumar Kshitij’s work has been recognized with the Distinguished Paper Award at IJCAI 2024 and the Best Paper Honorable Mention Award at the ICML 2023 Federated Learning Wor kshop. In addition to his research\, he has organized workshops and tutori als at venues such as UAI and TTIC\, fostering interdisciplinary collabora tions in machine learning and data science. Before joining TTIC\, he earne d a BTech in Computer Science from the Indian Institute of Technology (IIT ) Kanpur\, where he received the Honda YES Award in 2018. \;\n CATEGORIES:FDS Events,Postdoctoral Applicants LOCATION:https://yale.zoom.us/j/99456973798 END:VEVENT BEGIN:VTIMEZONE TZID:America/New_York X-LIC-LOCATION:America/New_York BEGIN:STANDARD DTSTART:20241103T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:EST END:STANDARD END:VTIMEZONE END:VCALENDAR