Title: Vine copula autoencoders
Authors: Thibault Vatter - Columbia University (United States) [presenting]
Natasha Tagasovska - University of Lausanne (Switzerland)
Damien Ackerer - Swissquote Bank Ltd (Switzerland)
Abstract: A vine copula autoencoder is proposed to construct flexible generative models for high-dimensional distributions in a straightforward three-step procedure. First, an autoencoder compresses the data using a lower dimensional representation. Second, the multivariate distribution of the encoded data is estimated with vine copulas. Third, a generative model is obtained by combining the estimated distribution with the decoder part of the autoencoder. This approach can transform any already trained autoencoder into a flexible generative model at a low computational cost. This is an advantage over existing generative models such as adversarial networks and variational autoencoders which can be difficult to train or impose strong assumptions on the latent space. Experiments on MNIST, Street View House Numbers and Large-Scale CelebFaces Attributes datasets show that vine copulas autoencoders achieve competitive results.