53adaf494dc89ef7196d73636eb2451b-Reviews.html

Summary The authors present a directed model for estimating the density of continuous random variables by exploiting the chain rule of probability theory. They propose to use particular weight-sharing constraints which have proven useful for modeling discrete data and combine it with mixture density networks. They show that their model can generally outperform large mixtures of Gaussians when applied to image patches, speech signals, and several smaller datasets. Comments Weight sharing -------------- Since the main difference to related work appears to be in the RBM-inspired weight-sharing, it would be interesting to see a more thorough investigation of its effects. While it is clear that it can reduce computational costs, its effects on performance have not been fully explored. One would expect the weight-sharing to reduce overfitting where data is scarce, and to hurt performance where data is plenty.