High-Capacity Complex Convolutional Neural Networks For I/Q Modulation Classification

I/Q modulation classification is a unique pattern recognition problem as the for each class varies in quality, quantified by signal to noise ratio(SNR), and has structure in the complex-plane. Previous work shows treating samples as complex-valued signals and computing complex-valued within deep learning frameworks significantly increases the over comparable shallow CNN architectures. In this work, we claim of the art performance by enabling high-capacity architectures containing and/or dense connections to compute complex-valued convolutions, with classification accuracy of 92.4% on a benchmark classification problem, RadioML 2016.10a dataset. We show statistically significant improvements in networks with complex convolutions for I/Q modulation classification. and inference speed analyses show models with complex convolutions outperform architectures with a comparable number of parameters comparable speed by over 10% in each case.