I have a hunch that half of Andrew Ng's approach to really invest in better data is actually about lowering the Bayes risk of the associated learning problem.