For example, AFAIK pooling layer in CNN is not differentiable, but it can be used because it's not learning. Is it always true?