Dimensionality Reduction for Supervised Learning with Reproducing Kernel Hilbert Spaces