Compute the expected output after dropout from the keep probability, plus the inference-time scaling factor that keeps train and test expectations aligned.
Dropout 率通常取 0.2-0.5。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| Training behaviour | training drops units and rescales survivors by the keep probability | Dropout paper |
| Inference behaviour | inference uses all weights without dropping | Dropout paper |
| Regularisation view | dropout approximates an ensemble of subnetworks | Dropout paper |