Bias and Generalization in Deep Generative Models: An Empirical Study