Improving Disentangled Representation Learning with the Beta Bernoulli Process