The idea is ridiculously simple (perhaps why it is effective?): randomly skip layers while training • /r/MachineLearning