Question 21/30 fast.ai v3 lecture 10

When should you and shouldn't you use softmax?

Answer

Softmax is a terrible idea unless each of your items has one, no more than one and at least one example of the thing you care about

Relevant part of lecture

supplementary material

Softmax likes to pick one thing (because of the exponential nature it pushes the biggest value to one) and it always needs to pick something. If the answer does not hold for your dataset, go for the binomial.