Why Intel Is Tweaking Xeon Phi For Deep Learning