Improving neural networks by preventing co-adaptation of feature detectors
This paper introduces the "dropout" technique, which randomly omits feature detectors during training to prevent complex co-adaptations and overfitting, thereby significantly improving neural network performance on tasks like speech and object recognition.