Title:	FedMem - Predict event memorability from life-logger data
DNr:	Berzelius-2024-110
Project Type:	LiU Berzelius
Principal Investigator:	Monowar Bhuyan <monowar.bhuyan@umu.se>
Affiliation:	Umeå universitet
Duration:	2024-03-15 – 2024-10-01
Classification:	10201
Homepage:	https://people.cs.umu.se/monowar/
Keywords:

Abstract

Automated systems that measure the ability of a person to recall episodic events from their past are immensely useful to monitor and treat patients suffering from memory disorders. Maintaining patient privacy in a centralized system where all the users’ data is aggregated is challenging, which in turn makes patients cautious about sharing personal data. One solution is training a model on a user's data on their client device. However, as the nature and amount of data available for each user are different, the performance of each client model will vary largely. Therefore, we propose to share the client models to improve their performance by leveraging clustered federated learning. Instead of sharing all the client models globally, we devised FedMEM to model similarity to detect social similarities among clients and group them into clusters. We show that our approach produces personalized client models that have similar performance across clients. We also demonstrate that our approach produces better personalized models than state-of-the-art, federated learning algorithms.