Complex-Valued Deep Belief Networks

Deep belief networks were among the first models in the deep learning paradigm. Their use for unsupervised pretraining allowed deep neural network architectures to perform better than shallow ones. This paper introduces complex-valued deep belief networks, which can be used for unsupervised pretraining of complex-valued deep neural networks. Experiments on the MNIST dataset using different network architectures show better results of the complex-valued networks compared to their real-valued counterparts, when complex-valued deep belief networks are used for pretraining them.