AlphaDropoutBackward

Autograd node for alpha-dropout — element-wise dropout that preserves the self-normalising statistics required by SELU activations.

Dropped elements are not zeroed but replaced with the SELU saturation value $\alpha' = -\lambda\alpha \approx -1.7581$; an affine rescaling $y \mapsto a \cdot y + b$ is then applied so that the mean and variance of the post-dropout distribution match those of the input. The chain rule therefore collapses to $\partial y / \partial x = a \cdot \text{mask}$, which is what apply evaluates using the saved raw mask.