Relaxation Networks for Large Supervised Learning Problems