Conditional density estimation with a neural network using the GEM algorithm

Recent theoretical advances have shown the applicability of neural networks in density estimation. However, training in these methods is slow, especially where gaps exist in the data (which is often the case in practical situations). A standard method for attacking this missing data problem is to use the GEM (or generalized expectation-maximization) algorithm. We apply this algorithm to conditional density estimation with missing data, and show that it requires significantly fewer training examples to attain acceptable performance