Softmax is an activation function that transforms a vector of real numbers into a probability distribution. It's commonly used in the output layer of multi-class classification models. The highest probability corresponds to the predicted class label.