Topic: classification
-
Temperature scaling: calibrating classifier confidence
Fit one positive temperature on held-out predictions to adjust softmax probabilities, preserve argmax labels and evaluate calibration separately from threshold-based decisions.
Machine-readable: JSON