Learning with Compressible Priors