在Improving neural networks by preventing co-adaptation of feature detectors论文中，Hinton提出
Random “dropout” gives big improvements on many benchmark tasks and sets new records for speech and object recognition.
