A Curriculum-Driven Approach to Client Selection in Federated Learning
PhD. Seminar by: Soroush Ziaeinejad
Date: Friday, November 21, 2025
Time: 1:00 PM
Location: Essex Hall, Room 122
Federated learning often faces performance degradation in non-IID settings, where the choice of participating clients strongly influences training stability and model quality. This work examines the use of curriculum learning as a structured approach to improving client selection. Instead of sampling clients randomly, we order their participation based on meaningful indicators derived from their previous performance and data characteristics. Using the MedMNIST dataset, we analyze how this curriculum-driven strategy affects accuracy, fairness, and training dynamics. Our initial findings indicate that incorporating curriculum learning leads to a more stable optimization process and achieves better overall performance compared to conventional client selection methods.
Federated Learning, Client Selection, Curriculum Learning, Non-IID Data
Internal Reader: Dr. Jianguo Lu
Internal Reader: Dr. Muhammad Asaduzzaman
External Reader: Dr. Majid Ahmadi
Advisor(s): Dr. Saeed Samet