Why the Sigmoid Function Destroys Uncertainty

(and why naive probability calibration fails)

Most machine-learning practitioners believe that probability calibration is a matter of “fixing” model outputs after training.