Work in Progress. Last update: 7th Nov.
MaxProb describes the max probability in a distribution.
We will start from two representative paper.
One for surface competition [PDF@aclweb], one for domain shift in QA tasks [PDF@aclweb].
We all know max
function has no gradient.
max = $argmax$ $D$
Can we make a differentiable type of MaxProb?
Go back to main webpage for MLP Blog.