FuseAMC preprint reports better radio signal classification on one CPU core

Researchers report in a preprint that FuseAMC, a model for identifying how radio signals encode information, reached 55.8% overall accuracy on RadioML 2016.10a while training and running on one CPU core. It improved on the best cumulant-feature baseline by 5.1 points and on a convolutional neural network alone by 3.2 points. The system combines a neural network's predictions with those of a classifier using statistical signal features. The network handles noisier signals better, while the statistical classifier performs better on cleaner signals. Tested across 5 seeds on a 44k-frame held-out test set, the pipeline requires roughly 3 MFLOPs per frame; the reported gains concern that benchmark.