Dropout

Randomly zeroing a fraction of activations during training so the network cannot lean on any single unit. A regularizer; switched off at inference.