Rate-Distortion Bounds on Bayes Risk in Supervised Learning