An Optimality Principle for Unsupervised Learning