Half-time seminar by Li Ju: "Learning from distributed and heterogeneous data"

Welcome to a half-time seminar in scientific computing presented by Li Ju. The seminar will be held in English.

External reviewer: José Mairton Barros da Silva Júnior (Department of Information Technology; Division of Computer Systems).

Abstract:

In modern data-driven applications, addressing the challenges posed by heterogeneous and distributed data is crucial due to the increasing variation in data sources, focuses, and formats. Learning from such data across multiple clients or domains, while ensuring efficiency fairness and certainty, remains a significant challenge.

This half-time seminar highlights three research works that contribute to this goal. The first work addresses the fairness problem in heterogeneous federated learning, proposing methods to mitigate performance imbalances arising from client-specific data disparities. The second work empirically examines the impact of heterogeneous data on federated learning performance within pharmaceutical studies. The third study presents a post-hoc framework for obtaining uncertainty estimates in pre-trained vision-language models by exploiting the asymmetric uncertainty structure of textual and visual information. Together, these studies advance the development of learning algorithms capable of handling the inherent complexities of distributed, heterogeneous data.