Compressing Deep Neural Networks Via Knowledge Distillation