Distributed and Quantized Kernel-based Learning over Interconnected Sensing Systems

University Of California-Irvine, Irvine CA

Investigators

Abstract

Kernel-based learning is widely used for nonlinear function learning, which is a general task in various machine learning problems, e.g., classification and regression. This leads to its wide application in pattern recognition and data analysis in many challenging tasks, including but not limited to time series prediction in various interconnected sensing systems such as sensor and IoT networks. For example, sensors measure the temperature, humidity, pressure, or other physical phenomena to predict future measures, and cameras take pictures, or videos to recognize an object. In many applications, multiple access nodes collect and/or disseminate information over a certain geographical area of interest. In addition, sensing systems include many computationally capable devices like smartphones, UAVs, cars, and so on. In such distributed networks, both data and computational power are distributed. Transmitting the collected data back to a central entity for processing is not desired. Also, it is impossible to transmit the massive amount of collected data in real-time over networks. In addition, in many applications, there are valid security and privacy concerns about transmitting personal data, for example in medical and finance applications. Therefore, the proposal aims to design distributed and quantized learning algorithms that do not transmit the collected data over networks. The results of this research will be disseminated broadly through traditional scholarly venues to the entire research community, government, and industry. The results of the proposed activity will be integrated into coursework and educational research initiatives at UCI, which is a Hispanic-serving institution (HSI). We propose the design of online distributed and quantized kernel-based learning algorithms that calculate some "local" updates and communicate the corresponding "updates" to their neighbors such that collectively the network can learn a "global" model. This is done without transmitting the collected data over sensing systems with static and dynamic network architectures. We propose to formally present different distributed and quantized function learning algorithms and study their performance including their convergence and regret analysis. Our algorithms will be designed for different network structures while accounting for network delays and dynamics. We also propose the design of adaptive distributed and quantized algorithms and study their performance and regret analysis. In addition, we study the optimal network resource allocation in these scenarios and the corresponding trade-off between computation accuracy and network resources. Our goal is to design robust distributed and quantized kernel learning algorithms over distributed sensing systems that only need to communicate with neighboring nodes and are less sensitive to network characteristics, like network topology and delays. 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.