CIF: Small: Latent Neural Factor Models for Radio Cartography From Bits

Oregon State University, Corvallis OR

Investigators

Abstract

In the next generation of intelligent, cognitive and software-defined wireless systems, everything is expected to be connected, literally. Advanced radio frequency (RF) awareness techniques will be the cornerstone of wireless resource management, interference avoidance, transmission optimization, and decision making in a highly crowded, self-organized, and heterogeneous wireless communication environment. To advance RF awareness, spectrum cartography crafts a "radio map" across multiple dimensions (e.g., time, frequency and space) from limited sensors and measurements. Prior approaches often rely on over-simplified RF environment models (e.g., smooth and static radio maps) and problem settings (e.g., using unquantized overhead), which lowers performance when applied in real-world settings. Leveraging recent advances in artificial intelligence, this project aims to develop spectrum cartography theory and methods under complex, heavily shadowed and dynamic environments using limited (i.e., a few bits of) information exchange, which are largely uncharted research waters. In particular, the project seeks to design a class of latent neural factor analysis (LaNFAC) models to represent the RF environments in a parsimonious way. Using the LaNFAC models, the project will offer spectrum cartography approaches to reconstruct realistic RF environments from limited and quantized measurements. Theory and methods developed in this project may find wide application in such disciplines as geoscience, food science, video processing, and medical imaging. The research will bolster undergraduate education and offer training opportunities in optimization, deep learning, tensor analysis, and sensing to students from under-represented and under-served groups with the aim to enhance their career prospects in signal and machine intelligence. This project will develop a suite of analytical and computational tools for provable, robust and efficient spectrum cartography from a small number of measurement bits, by way of developing a variety of LaNFAC tools for radio map modeling. The LNFAC models are a judicious integration of latent factor analysis models (e.g., tensor decomposition) and neural generative models. The work will first develop the basic framework of limited feedback-based and LaNFAC-assisted spectrum cartography in realistic RF environments. Then, the project will consider more challenging scenarios (e.g., no training data) and develop provable spectrum cartography from quantized information feedback/exchange using untrained LaNFAC models. The last research thrust will validate the theory and evaluate the algorithms using carefully designed simulators and software-defined radio experiments using real data. Using untrained neural models retains strong expressiveness without relying on training data, which will facilitate distributed, exchange-limited, and adaptive spectrum cartography. Real-data acquisition and releasing will assist the research community to develop effective and reproducible spectrum cartography approaches, and ultimately advance understanding of the RF awareness problem in a collective way. 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.