A Conditional Analysis of RF ML Latent Features

The use of radio frequency (RF) machine learning (ML) continue to expand into more applications of the field. These models have shown that they are capable of handling ambiguities and complexity in datasets and settings that pose challenges to the domain. While RF ML methods have shown success, we are still seeking to understand how to best design the solutions, what cases they work best in, what potential failures could arise, and how to build trust into the systems, amongst other important questions that need to be answered. In light of that, this paper takes a look at what is produced in the latent feature space of an RF model, instead of studying the resulting performance of such a system. The paper studies the RiftNetT model applied to modulated communication signals. Latent features in a classification task (modulation recognition) are contrasted with one in a generative case (synthesizing modulated signals). The impact of a signal-to-noise ratio, hold-out datasets, and bottleneck layer designs are explored. Results help to shed some light on design choices that should be made for different applications and end goals.