From Logits to Probabilities | Visual AI Learning Guide

Those logits can be any size, so a score of eight next to a score of two hundred tells you little on its own. To make them comparable, the model runs them through a function called softmax.

0:00 / 0:00

Auto-play

Logits are unbounded, so they cannot be read as likelihoods directly. Softmax converts them: it exponentiates each logit, which makes every value positive, then divides by the sum of all exponentiated values. The result is a probability distribution in which every value lies between zero and one and the whole set sums to exactly one. Now the numbers are comparable likelihoods rather than arbitrary scores.