Dropout
Randomly disabling neurons during training to prevent overfitting.
Dropout randomly zeroes a fraction of units on each training step, forcing the network not to rely on any single neuron and acting like an efficient ensemble. It became a standard regularization technique across deep learning.