Efficient DNN Training with Knowledge-Guided Layer Freezing