Looping back: Circular nodes revisited with novel applications in the radio frequency domain

Tegan Emerson, Tim Marrinan, Bill Kay, Rachel Wofford, Audun D Myers

Track: Full Paper (8 pages)

Keywords: phase estimation, radio frequency, modulation classification, circular nodes, neural networks

TL;DR: Paper revisits a paired node architecture intended to leverage both phase and amplitude content with applications in the RF domain.

Abstract: In domains with complex-structured data, some relationships cannot be easily modeled using only real-valued Euclidean features. In spite of this misalignment, most modern machine learning methods default to representing data in just that way. By failing to appropriately encode the data structure, the performance and reliability of the resulting machine learning models can be degraded. In prior work, Kirby and Miranda introduced the concept of a circular node, a type of artificial neuron engineered to represent periodic data or angular information~\cite{kirby1996circular}. These nodes can be implemented directly in many traditional neural network architectures to more faithfully model periodic relationships. However, since they have garnered relatively little attention compared to their non-circular counterparts, circular nodes have largely been excluded from open-source machine learning libraries. In this paper, we re-investigate circular nodes in the context of modern machine learning libraries, and demonstrate the advantages of circular representations in applications with complex-structured data. Our experiments center around radio frequency signals, which naturally encode circular relationships. We illustrate that a neural network composed of a single circular node can learn the phase offset of a radio frequency signal. We show that a fully-connected neural network made up of multiple layers of circular nodes can successfully classify digital modulation constellation points, and demonstrates accuracy gains over its traditional counterpart when the model size is small. Finally, we demonstrate notable performance improvements on the task of automatic modulation classification through the integration of a circular node layer into traditional convolutional networks.

Submission Number: 13