Why does randomly 'breaking' a neural network during training actually make it smarter?

The answer

Dropout, introduced by Geoffrey Hinton and his students Nitish Srivastava and others in 2014 (first presented in 2012), is a regularization technique that randomly deactivates a fraction of neurons (typically 50%) during each training step. This forces the network to develop redundant representations and prevents neurons from co-adapting too tightly, dramatically reducing overfitting. The technique was inspired partly by biological theories about the role of noise in neural development. Dropout was a key ingredient in AlexNet's success and remains one of the most widely used regularization methods in deep learning.