Prof. Anthony Kuh
IEEE Members: Free
Non-members: FreeDuration: 00:31:50
In recent years there have been many applications where data comes from distributed sources and with the increasing computational capabilities of edge devices more computing, processing and machine learning is done at the edge. Applications range from monitoring the environment to healthcare to transportation to energy to social networking. The edge devices could be sensors from a sensor network, IoT devices, mobile phones, or intelligent assistants. The edge devices learn from data they receive with popular methods including Federated Learning (FL). FL has advantages with learning done by edge devices thereby eliminating the need for data to be transmitted to the cloud or central processor. These advantages include communication savings, increased security, and addressing privacy concerns. In this presentation, we discuss real-time FL using kernel methods. We discuss real-time learning as much of the data in applications is dynamic or streaming requiring online learning and decision-making. Kernel methods are used as online nonlinear linear algorithms can easily be developed using tools from adaptive signal processing and convex optimization approaches. We show how online kernel methods can be modified for distributed learning and FL. Online FL kernel algorithms are developed, analyzed, and compared via simulation studies.