It is used to transform a list of classes prediction scores into probabilities

The Softmax function is usually used in classification problems such as neural networks and multinomial logistic regression, this is just generalisation of the logistic function:

f(x) = 1/(1 + e^(-k(z-z0)))

The output of the Softmax is used as a categorical distribution meaning that for each outcome we will get the possibility of it, it takes into account both the result we are checking but all other results together to create the probability.

Yi = e^Zi/ sum(e^Zi)

The beauty of this function is that if you create the derivative according to Zi you will get an elegant solution :

Yi(1-Yi)

So it is very easy to work with.

http://www.talkinghightech.com/en/softmax-function/