[Theory] NOW: [TTIC Talks] 6/10 Thesis Defense: Kumar Kshitij Patel, TTIC

[Brandie Jones via Theory]

*When:     *Wednesday, June 11th at *11am CT*

*Where:*     Talk will be given *live, in-person* at

                       TTIC, 6045 S. Kenwood Avenue

                       5th Floor, *Room 529*


*Virtually:  *via Zoom
<https://uchicago.zoom.us/j/96211938253?pwd=T6huTrQkHiJCnIwFcmzUuik9OlvNSZ.1>

*Who*:           Kumar Kshitij Patel, TTIC

*Title: *What Makes Local Updates Effective: The Role of Data Heterogeneity
and Smoothness

*Abstract:* Over the past decade, Federated Learning (FL) has emerged as a
powerful framework that enables learning from multiple decentralized
datasets without the need to exchange raw data. This paradigm has led to
significant advancements across various fields, including healthcare,
finance, research, and consumer technologies. However, despite extensive
research in FL, important foundational questions remain unanswered. One
such question pertains to the effectiveness of local update algorithms,
particularly in understanding when these algorithms outperform traditional
distributed optimization methods.

This thesis addresses this question by focusing on Local Stochastic
Gradient Descent (Local SGD), also known as Federated Averaging, which is
the simplest and most commonly used local update algorithm in FL. We
present non-asymptotic convergence analyses for Local SGD in convex
settings, deriving tight upper and lower bounds that reveal how
second-order data heterogeneity and third-order smoothness dictate its
communication efficiency.

Additionally, we enhance Local SGD by incorporating variance reduction
techniques, resulting in an optimal algorithm specifically designed for
distributed non-convex optimization. Throughout this exploration, the
thesis also characterizes minimax complexity for several significant
distributed optimization problems. Finally, we examine how distribution
shifts and adaptive adversaries affect distributed online and bandit
optimization frameworks, proposing new algorithms and offering insightful
analyses.

-- 
*Brandie Jones *
*Executive **Administrative Assistant*
Toyota Technological Institute
6045 S. Kenwood Avenue
Chicago, IL  60637
www.ttic.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20250611/1742e77b/attachment.html>