Plummer Autoencoders
share
Estimating the true density in high-dimensional feature spaces is a well-known problem in machine learning. This work shows that it is possible to formulate the optimization problem as a minimization and use the representational power of neural networks to learn very complex densities. A theoretical bound on the estimation error is given when dealing with finite number of samples. The proposed theory is corroborated by extensive experiments on different datasets and compared against several existing approaches from the families of generative adversarial networks and autoencoder-based models.
READ FULL TEXT