Performance evaluation of Complex-Valued Neural Networks on real and complex-valued classification and reconstruction tasks

Abstract

Complex-Valued Neural Networks (CVNNs) are reported to be more efficient in different applications than Real-Valued Neural Networks (RVNNs) in many papers. In this study, we aim to characterize the cases when it holds true in order to assist the selection of proper tools for two specific tasks: classification and reconstruction.

Among the various ways to compare CVNNs and RVNNs, we apply the one based on the number of parameters of the respective Neural Networks (NNs), assuming that a complex parameter is composed of two real ones. The performed experimentation revealed many surprising differences in the performance of CVNNs and RVNNs compared to the ones discussed in the preceding literature. This drives us to the general conclusion that the performance of RVNNs is similar or better than the performance of CVNNs in the majority of the cases, and the seldom cases when CVNNs achieve better performance are hard to characterize.