Complex-Valued Stacked Denoising Autoencoders

Stacking layers of denoising autoencoders, which are trained to reconstruct corrupted versions of their inputs, results in a type of deep neural network architecture called stacked denoising autoencoders. This paper introduces a model of complex-valued stacked denoising autoencoders, which can be used to build complex-valued deep neural networks. Experiments done using the MNIST and FashionMNIST datasets show superior performance of the complex-valued stacked denoising autoencoders with respect to their real-valued counterparts, both in terms of reconstruction error, and in terms of classification error.