뉴스 및 공지사항
- 관리자
- 2024.05.24
- 310
[세미나] [ISE/GSDS Seminar] May 24 11:00 / Online Seminar / Recent Advances in Asynchronous Federated Learning Algorithms / Dr. Kibaek Kim / Computational Mathematician / Mathematics and Computer Science Division / Argonne National Laboratory
ISysE / GSDS 세미나가 다음과 같이 진행될 예정입니다.
# 날짜/시간: 2024년 5월 24일 금요일 11:00~12:00
# 장소 : 온라인 세미나
# Zoom link : https://kaist.zoom.us/j/
# 연사 : Dr. Kibaek Kim / Computational Mathematician / Mathematics and Computer Science Division / Argonne National Laboratory
# 제목 : Recent Advances in Asynchronous Federated Learning Algorithms
# 초록 : Federated learning (FL) is a distributed machine learning paradigm that enables the training of models across decentralized data sources while preserving data privacy. This approach is particularly advantageous when data cannot be centralized due to privacy concerns. However, FL introduces unique challenges, including handling heterogeneous client devices and diverse local objectives. In this talk, we explore recent advances in asynchronous FL algorithms, summarizing our two recent studies that address these critical challenges. The first study introduces an innovative scheduling algorithm designed to dynamically adapt to the computing capabilities of client devices. This scheduler ensures balanced workloads and efficient resource utilization, significantly enhancing the overall training performance in cross-silo FL environments. The second study focuses on the issue of asynchronous updates in FL. We introduce a new stochastic optimization method that incorporates exact averaging, effectively managing the differences in local objective functions. This method not only speeds up convergence in the presence of stragglers but also maintains robustness against the heterogeneity of client data distributions. Together, these studies provide a comprehensive framework for advancing FL techniques, emphasizing the importance of advanced algorithms to handle heterogeneous environments. The insights from these works create new opportunities for more efficient, scalable, and resilient FL systems, suitable for a wide array of real-world applications.
# 본 세미나는 영어로 진행됩니다.
------------------------------
ISysE/GSDS dept. office invites you to the following seminar.
# Time/Date : May 24, 2024 (Fri.) 11:00~12:00
# Location : Online Seminar
# Zoom link : https://kaist.zoom.us/j/
# Presenter : Dr. Kibaek Kim / Computational Mathematician / Mathematics and Computer Science Division / Argonne National Laboratory
# Title : Recent Advances in Asynchronous Federated Learning Algorithms
# Abstract : Federated learning (FL) is a distributed machine learning paradigm that enables the training of models across decentralized data sources while preserving data privacy. This approach is particularly advantageous when data cannot be centralized due to privacy concerns. However, FL introduces unique challenges, including handling heterogeneous client devices and diverse local objectives. In this talk, we explore recent advances in asynchronous FL algorithms, summarizing our two recent studies that address these critical challenges. The first study introduces an innovative scheduling algorithm designed to dynamically adapt to the computing capabilities of client devices. This scheduler ensures balanced workloads and efficient resource utilization, significantly enhancing the overall training performance in cross-silo FL environments. The second study focuses on the issue of asynchronous updates in FL. We introduce a new stochastic optimization method that incorporates exact averaging, effectively managing the differences in local objective functions. This method not only speeds up convergence in the presence of stragglers but also maintains robustness against the heterogeneity of client data distributions. Together, these studies provide a comprehensive framework for advancing FL techniques, emphasizing the importance of advanced algorithms to handle heterogeneous environments. The insights from these works create new opportunities for more efficient, scalable, and resilient FL systems, suitable for a wide array of real-world applications.
# The seminar will be in English