Data-Centric Approaches to Radio Frequency Machine Learning.

The successes of machine learning (ML), and in particular deep learning, in other fields has inspired similar research within the radio frequency (RF) domain. Initial research in RF ML has been largely applied to the application of modulation recognition, with the past several years seeing it expand into other applications as well. The field has slowly evolved from the direct application of models developed in other fields, e.g., convolutional neural networks (CNN), to ones that are better suited for RF signals, e.g., dilated causal convolutions (DCCs). At the same time, the broader ML community has realized the importance data has on deep learning performance and a growing datacentric ML movement has emerged. In this paper, we return to the problem of modulation recognition and provide insights into how a data-centric approach can be coupled with a DCC model. In particular, we look at cases with limited amounts of training data and investigate means to achieve levels of performance typical reserved for larger training datasets. This is done by developing specific SNR models, data augmentation, performing multi-burst processing, and upsampling expected undersampled parts of an unbalanced training dataset. Overall, we present ways to intelligently use sparse available data to achieve the same performance as larger datasets, helping to mitigate a challenge in RF ML where gathering and curating large representative datasets is not always feasible.