Title
Coded Computing: How to Utilize Fast Workers in Heterogenous Systems?
Abstract
The widespread use of machine learning and big data has led to a surge in demand for computational power. To efficiently train learning models, computation tasks are often offloaded to distributed servers, or workers, which can compute in parallel to speed up the process. However, the workers may not be trustworthy, which raises data privacy concerns. Moreover, they may suffer from straggling or outages, causing delays or even complete failures. Researchers have adopted coding techniques originally designed for error correction to address these challenges to ensure data privacy and enhance system resilience against slow or unresponsive workers.
This seminar will be divided into two parts. The first part will introduce the concept of coded computing and highlight the challenge of dealing with heterogeneous worker speeds, particularly in edge computing scenarios where no prior knowledge of worker speeds is available. The selection of workers is formulated as a multi-armed bandit problem, solved by a reinforcement learning method called Thompson sampling. In the second part, the complementary scenario of multi-cloud computing will be considered, where there is complete knowledge of worker speeds. The fundamental tradeoff between storage and computing under security constraints is analyzed via information theory, and the impact of heterogeneity is characterized via majorization theory. The talk will then conclude by discussing future directions in this emerging field. Attendees will gain insights into the latest advances in coded computing and its potential for improving the efficiency and resilience of distributed learning systems.
Biographies
Albert Sung is an Associate Professor and the Associate Head of Undergraduate Programmes in the Department of Electrical Engineering at City University of Hong Kong. He holds a B.Eng, M.Phil, and Ph.D, all in Information Engineering, from the Chinese University of Hong Kong. His research interests span broad areas of communications, coding, and computing, with a specific focus on resource allocation for mobile networks, code constructions for distributed storage systems, and the design and analysis of evolutionary algorithms. In recognition of his contribution to the field, he has been included in the list of top 2% most highly cited scientists complied by Stanford University since 2020. He also serves on the editorial boards of the ETRI Journal and Electronics Letters.
Jiajun Chen is a Ph.D. student in the Department of Electrical Engineering at City University of Hong Kong. She received the B.S. degree from Sichuan University, Chengdu, China, in 2018. She has a journal paper published in IEEE Transactions on Information Theory (TIT) in 2023. Her research interests include coded distributed computing, information theory, and resource allocation.
