Federated Sensing hands-on tutorial
Stefanos Laskaridis, Dimitris Spathis, Mario Almeida
TL;DR: Showcase of a distributed FL setup from scratch for an activity recognition task.
Abstract:
In this hands-on session, we are building a federated training setup from scratch, destined to perform the task of activity recognition by using a mobile phone’s accelerometer data. To this direction, we will build upon the Flower framework, PyTorch and the MotionSense HAR dataset as a case in point.
During this session, we go in detail through the creation of the sever and client-side functionality, the lifecycle of each entity and the state of learning. We discuss about aggregation and optimisation schemes in practice, as well as issues such as client selection, hyperparameter search and secure computation.
Participants are free to download the code and dataset and participate in the tutorial with their laptops.