What is inverted dropout?

Scale activations during training so no extra scaling is needed at inference

Drop all weights permanently after training

Increase dropout probability during inference

Normalize labels before applying dropout