Heterogeneity Aware Distributed Machine Learning at the Wireless Edge for Health IoT Applications: An EEG Data Case Study

In this book chapter, we design and develop a mobile edge learning (MEL) framework that enables multiple end user devices or “learners” to cooperatively train a machine learning (ML) model in a wireless edge environment. We will focus on designing and developing the heterogeneity aware synchronous (HA-Sync) approach with time constraints and extend the framework to consider dual-time and energy constraints. The proposed MEL framework will include the commonly known federated learning (FL) as well as parallelized learning (PL). After discussing the system model and a brief convergence proof for both FL and PL, we will formulate the problem as a quadratically constrained integer linear program (QCILP), relax it to a QCLP, and propose analytical solutions based on Lagrangian analysis, Karush-Kuhn-Tucker (KKT) conditions, and partial fraction expansion. For the problem with dual-time and energy constraints, we will propose solutions based on the suggest and improve (SAI) approach. Results based on achievable local updates, validation accuracy progression, and the optimization algorithm’s execution time will be used to demonstrate the superiority of the proposed HA-Sync compared to the heterogeneity unaware (HU) approaches. As an application focus, we will demonstrate a real-time medical event prediction by showing the applicability of personalized MEL for epileptic seizure detection and predictions – using EEG data.