EAGER:Real Time Federated Learning using Kernel Methods

We are increasingly seeing increasing amounts of dynamic data coming from a variety of different applications ranging from the electric power grid to transportation systems to healthcare networks to social network data. This data comes from a variety of sources including sensor networks and IoT applications. A common feature is that much of this data is that it is distributed (at the edge) and methodologies need to be established to process, learn, and make decisions from this data.

One approach is to gather all the data together at a central processor (in the cloud) and to process, learn, and make decisions here where computing resources are greater. The edge devices however are often constrained by communication costs which results in considerable power consumption especially if edge devices need to send information over wireless networks.

Federated learning has been proposed where edge devices do not send data to the central processor, but send model parameters (weights) of model to edge devices. Each edge device modifies the weights based on data it receives and then send the modified weights back to the central processor. This proposal develops simple real-time learning algorithms at the edge processor based on unconstrained optimization methods using principles of federated learning and studies applications for power systems.

The research results will be incorporated to both graduate and undergraduate courses in machine learning and signal processing. Special attention will be given to recruitment and retention of underrepresented students through the Native Hawaiian Science and Engineering Mentorship Program (NHSEMP) and the Society of Woman Engineers (SWE).

The project makes advances in the field of real-time distributed learning and decision making using principles of federated learning, adaptive signal processing, graph signal processing, optimization, and kernel methods. The research is convergent bringing in the fields of signal processing, mathematics, statistics, and computer science together. The contributions are in three areas: algorithm development, analysis, and applications

to the electric power grid. A focus is to design simple online kernel algorithms that are applicable for both supervised learning (regression, prediction, classification) and unsupervised learning (principal component analysis, probability density estimation). The algorithms focus on edge computing using optimization methods from online least squares kernel methods using variants of stochastic gradient and the temporal and spatial relationships between nodes represented as graphs.

Tradeoffs are considered between optimizing objective functions, convergence, computational complexity, and communication costs. The online distributed kernel algorithms are then applied to detect bad data on the power grid and providing distributed learning capabilities for demand response (DR) programs.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.