Empirical Analysis of Complex-Valued Neural Nets on Real-Valued Tabular Data

Complex-valued neural networks (CVNNs) have gained recent traction for their potential benefits in having better representational capacity and robustness to noise. The CVNNs have been mostly studied on non-tabular datasets and on datasets where complex numbers naturally occur by design, for example, in the field of signal processing. However, real-valued tabular datasets are the most common in academic research and industry specific machine learning tasks. This paper aims to empirically study the performance of CVNNs on real-valued tabular datasets. The performance of the real-valued neural networks (RVNNs) and CVNNs were studied on five different real-valued tabular datasets. We find that CVNNs perform marginally better than RVNNs on all the datasets in terms of R2 score and RMSE with a significantly faster convergence rate. Our results suggest that CVNNs may be a promising avenue for further research.