Softmax
Softmax converts a set of scores into probabilities that sum to 1, commonly used for multi-class prediction outputs.
Definition
Softmax is a function that turns a vector of scores into probabilities that sum to 1. It is widely used when a model outputs multiple competing outcome scores.
Why it matters
Softmax is a standard way to produce valid probability distributions from model scores. It is conceptually related to logit in the binary case and is often paired with proper scoring rules such as log loss.
Common pitfalls
Overconfidence from sharp distributions: Very peaked softmax outputs can be miscalibrated.
Assuming probabilities are calibrated: Softmax produces probabilities, but calibration still must be measured and corrected.