Is it possible to sample from a distribution inside a neural network forward function? Assume that there is a NN and a sample is needed to be derived from it at every forward-pass to randomly set a layer-specific hyper-parameter.
Is this operation differentiable
