Mobicom 2021 postponed.
MobiCom 2021 is being postponed again to March 28 through April 1, 2022, with plans to hold it then in person in the conference’s originally planned location of New Orleans. The MobiCom and SIGMOBILE committees are also exploring ways to make a hybrid option available for presenters or participants unable to attend in person. More information to follow, but also keep updated with the Mobicom’21 website.
Tutorial on Federated Mobile Sensing for Activity Recognition, hosted by Mobicom 2021
Smartphones have become omnipresent nowadays, a necessary professional, entertainment and communication tool. With the advent of smartphones and their sensing capabilities, varying from multiple camera sensors to accelerometers, gyroscopes and a multitude of microphones, users produce more data than ever before. This data can be beneficial for various tasks, ranging from activity or context recognition to intelligent assistants and mobile health. A key enabler for such tasks has been deep learning and the simultaneous breakthroughs on SIMD hardware, embedded (SoC) or not (discrete accelerators, e.g. GPUs).
Nevertheless, a crucial counterweight to these innovations is privacy; in particular privacy of the user data when performing intelligent tasks. To this direction, there has been a late series of works based upon the paradigm of Federated Learning (FL), which aims to create globally robust models without accessing the user device data directly. By combining the novel distributed training architecture with techniques such as Differential Privacy, Secure Multi-Party Computation and Homomorphic Encryption, this approach has the potential of guaranteeing the privacy of user data without sacrificing utility and accuracy.
Recognising the role and momentum of Federated Learning in recent research, in this tutorial we aim to illustrate how a researcher or engineer can bootstrap their work, by implementing a distributed federated training pipeline on mobile sensing data, with a benchmark task of activity recognition through mobile accelerometer (MotionSense HAR). We will cover and show to participants techniques for robust 1) data preparation, 2) model development, 3) federated training, 4) optimization, 5) evaluation, and 6) execution on resource-constrained environments. In addition, we plan to have three additional lectures on the topics of on-device sensing, federated learning and its challenges as well as a more practical “how to deploy” and scale such distributed infrastructures for production environments.
The tutorial is co-located with Mobicom and planned to be hosted in a hybrid (virtual and in-person) fashion, in the beautiful city of New Orleans.
The tutorial will take place on the 1st April 2022
4th of February 2022 29th of October 2021.