Complex-Valued Convolutions for Modulation Recognition using Deep Learning

Natural signals are inherently comprised of two components, real and imaginary components. Due to recent successes and progress in Deep Learning, specifically Convolutional Neural Networks (CNNs), this field of machine learning has become extremely popular when handling a wide variety of data, including natural signals. However, deep learning frameworks have been developed to deal with exclusively real-valued data and are unable to compute convolutions for complex-valued data. In this work, we present a linear combination that enables deep learning architectures to compute complex convolutions and learn features across the real and imaginary components of natural signals. When implemented into existing I/Q modulation classification architectures, this small change increases classification accuracy across a range of SNR levels by up to 35%.