What Happens on the Edge, Stays on the Edge: Toward Compressive Deep Learning