Does the Wake-sleep Algorithm Produce Good Density Estimators?

By Brendan J. Frey, Geoffrey E. Hinton, Peter S. Dayan. 1995

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

RNADE: The real-valued neural autoregressive density-estimator (2013, methodology): “For binary data, each conditional distribution can be modeled with logistic regression, which is called a fully visible sigmoid belief network (FVSBN) [7].”
Connectionist multivariate density-estimation and its application to speech synthesis (2016, background): “Helmholtz machines are capable of capturing complex dependencies between the visible variables (Frey et al., 1996; Mnih and Gregor, 2014).”
Learning to Parse Images (1999, methodology): “Efficient methods of approximat ing the posterior distribution were introduced later [5, 6, 7] and these approaches were shown to yield good density models for binary images of handwritten digits [8].”
Learning with Mixtures of Trees (2000, methodology): “…mixture of factorial distributions (MF), (3) the UNIX " gzip " compression program, (4) the Helmholtz Machine, trained by the wake-sleep algorithm (Frey et al., 1996) (HWS), (5) the same Helmholtz Machine where a mean field approximation was used for training (HMF), (5) a fully observed and fully…”
A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications (2020, background): “PixelCNN [104] and fully visible belief networks (FVBNs) [105], [106].”
A Mean Field Learning Algorithm for Unsupervised Neural Networks (1999, background): “In a recent study, Frey et al[15] reported that learning by Gibbs sampling was an extremely slow process for sigmoid belief networks.”