MLHPC 2016 | Communication Quantization for Data-parallel Training of Deep Neural Networks