Improving the speed of neural networks on CPUs

   Abstract