Real Time Kernel Learning for Sensor Networks using Principles of Federated Learning

This paper considers online kernel learning at the edge. Increasingly, we are seeing more networks consisting of edge processors and a central processor or coordinator. The edge processor could be a sensor, mobile phone, cognitive assistant, and/ or IoT devices. These processors likely have limited power and communication capabilities, but have enough processing ability to make intelligent decisions and learn. With the advent of federated learning, these edge processors do not pass data to the central processor, but assist in learning by adjusting parameters of the learning model based on data the edge processor receives before passing information about the updated model to the central processor. Here we consider where the edge processors use kernel methods to perform the learning. The learning algorithms are online with edge processors adjusting the model parameters using stochastic gradient algorithms. We establish a framework for online federated kernel learning.