Logo Università degli Studi di Milano



Methods for Pervasive Heterogeneous Environments on Sensors-Based Human Activity Recognition

DATE: Wednesday, October 26, 2022
TIME: 2:30 - 3:30 PM
PLACE: Lab. Laurea Magistrale (Floor 5)
SPEAKER: Sannara Ek (University of Grenoble-Alpes, France)

HOSTS: Claudio Bettini, Gabriele Civitarese


Within the pervasive computing domain, deploying machine learning solutions to clients faces many obstacles in real-world scenarios. Notably, the issue of heterogeneity severely impacts the trained model's ability to perform on real-life data. In this seminar, I will introduce two ways to mitigate heterogeneity's detrimental effects in the domain of wearable Human Activity Recognition (HAR). The first method seeks external solutions, such as having a user's model collaborate withother learners within the framework of federated learning to increase the model's generalization ability. Alternatively, it is possible to adopt internal solutions by relying on advanced learning architectures, such as state-of-the-art transformers, that can learn rich and robust representations, as seen in the vision and natural language processing domains.


Sannara Ek finished his master's in vision and robotics at Université Grenoble Alpes in Grenoble, France. He is now a 2nd year PhD student at the Grenoble Informatics Laboratory, working on heterogeneity within federated learning applied to pervasive computing environments.


10 ottobre 2022
Torna ad inizio pagina