Manifold Learning by a Deep Gaussian Process Variational Autoencoder

This paper presents a Deep Gaussian Process Variational Autoencoder (DGPVA), i.e., a novel Manifold Learning algorithm based on Deep Gaussian Processes. Besides, in order to assess quantitatively the performance of the proposed algorithm, an experimental validation protocol has been proposed. Experimental tests on synthetic and real data sets show that DGPVA compares favourably with state-of-the-art Manifold Learning algorithms.