Learning Factored Representations in a Deep Mixture of Experts