A Commentary on the Unsupervised Learning of Disentangled Representations