Calibration, reliability diagrams, proper scores, repairs and abstention on held-out text
|
Next Words - Lesson 16 - Calibration, Uncertainty (pdf)
|
389.92 K |
Among 1,630 held-out cases given a confidence of 0.5 or more, the mean confidence is 0.601 and the share correct is 0.336, so the height of a softmax output is a claim that can be audited. Calibration is a property of a set of predictions and is separate from accuracy and sharpness. Reliability diagrams are drawn on 11,933 predictions, and the expected calibration error ranges from 0.0471 to 0.0785 with the binning alone. Cross-entropy is a proper scoring rule and accuracy is not, since log loss goes from 0.4450 to 5.7091 when every confidence is replaced by 0.999, and accuracy stays at 0.1736. Miscalibration is lowered from 0.0734 to 0.0050 by histogram binning, with accuracy unchanged at 0.1741. Keeping the most confident twentieth of the cases gives an accuracy of 0.519 against 0.174 for answering everything.