Dropout

Randomly zeroes activations during training so no single unit can be relied on, then scales the survivors so the expected output is unchanged. At inference nothing is dropped, and the scaling is what makes the two regimes agree.