The softmax function normalizes a vector so its entries are positive and sum to

$.$ This is useful for creating a probability distribution across the vector, so it is commonly used as the final activation layer for neural networks $($especially since the gradient of the softmax function is relatively easy to compute$).$