ConceptTemperature Scaling
Temperature Scaling
A post-hoc probability calibration method that adjusts confidence by dividing the logits output by the model by a learnable scalar temperature parameter T. It has few parameters and does not change the ranking, but it relies on the validation set being drawn from the same distribution as deployment.
Timeline (last 90 days)
Oct 22
Temperature calibration assumes the validation set and deployment distribution come from the same source; once temporal drift occurs, the calibrated T becomes invalid.
Unverified50%