Improving Deep Neural Network Training With Knowledge Distillation