Yoshua Bengio
Universite de Montreal
Using the outputs of a neural networks to represent the parameters of the conditional distribution of a random variable, and using the same neural network to model many conditional distributions, one can learn the joint distribution of high-dimensional data (even high-dimensional discrete data) with a reasonable number of free parameters, while taking advantage of hidden units to capture commonalities between the different conditional distributions. Preliminary comparative experiments yielding excellent results in terms of out-of-sample log-likelihood will be described.