[talolard]

I don’t know if it’s correct , but I often think of a classification model as learning the parameters of a dirchlet distribution with the final softmax layer being a sample from it