Calibration: When Confidence Matches Correctness

Calibration is a promise about frequency. When a classifier reports 70 per cent confidence, it is not describing a feeling; it is making a prediction you can check. Collect every case where the model said 70 per cent. If the number is honest, about 70 in every 100 of those answers will be correct. A model whose stated confidence matches its actual rate of being right is called calibrated. This is not the same thing as accuracy. Accuracy asks whether the top answer is right. Calibration asks whether the number printed beside that answer can be believed.

The standard test is a reliability diagram. Sort the model's predictions into bins by stated confidence — everything near 60 per cent in one bin, everything near 90 per cent in another. In each bin, count the fraction of predictions that were actually correct, and plot it against the confidence the model claimed. An honest model traces the diagonal: 60 lands on 60, 90 on 90. Averaging the gaps across bins, weighted by how many predictions each bin holds, gives a single score called Expected Calibration Error.

Modern deep networks fail this test in a consistent direction. When Guo and colleagues ran it on image classifiers, the curves bowed above the diagonal: the models were over-confident, stating confidence well beyond their actual rate of being right. The pattern was not a quirk of one design. The very ingredients that had been raising accuracy — deeper stacks of layers, wider layers, and batch normalisation, a technique that stabilises training by rescaling the values flowing between layers — each went with worse calibration in the settings they examined. One ingredient pushed the other way: weight decay, a penalty that discourages large weights, tended to help.

The mechanism is worth walking through slowly, because it explains why the problem grew as models did. A classifier ends in a row of raw scores, one per class, called logits. A function called softmax converts those scores into probabilities that add up to 1. It favours the largest logit, and the wider the gap between the largest and the rest, the closer its probability gets to certainty. Training minimises a loss called negative log likelihood, which rewards the model for piling probability on the correct answer. Here is the trap. Even after the model has stopped getting more answers right, that loss can still be driven lower — by pushing the correct logit further above its rivals on the answers it already gets. Guo and colleagues observed exactly this: the loss kept falling after the error rate had flattened. Wider gaps between logits mean sharper probabilities out of softmax. So the model prints 0.99 where 0.85 would have been honest — not because it knows more, but because the training objective kept rewarding sharpness after correctness had run out of room to improve.

The repair can be a single number. Temperature scaling takes the finished, trained network and divides every logit by the same positive constant, T, before softmax. Dividing shrinks the gaps between the scores, and smaller gaps come out of softmax as gentler probabilities. Because every class is divided by the same T, their order cannot change: the winning answer stays the winning answer, and accuracy is untouched. Only the confidence moves. The value of T is chosen on a held-out validation set — data the model never trained on — by finding the T that makes the softened probabilities fit that data best. On the benchmarks Guo and colleagues studied, this one-parameter fix cut calibration error about as well as far heavier post-hoc methods — Platt scaling, histogram binning, isotonic regression and Bayesian binning — some of which fit many parameters per class.

Classification is only half the picture. When the target is a continuous quantity — a temperature, a price, a distance — the model's uncertainty takes the form of a range rather than a class probability, and calibration means that a range labelled 90 per cent should contain the true value about 90 per cent of the time. Kuleshov and colleagues showed that deep regression models often break this promise too, and that the fix follows the same shape. On held-out data, compare the coverage the model promised at each level with the coverage it actually delivered. Then fit a small correcting function through those pairs — usually isotonic regression, which finds the best-fitting curve that never decreases — and use it to widen or narrow every future interval. The network's weights are never retrained.

Both repairs share one honest limitation. They assume the validation data resembles the data the model will meet in deployment. Neither method creates new uncertainty; each reshapes the uncertainty the model already expresses so that its numbers can be read as frequencies. A model that meets an input unlike anything in its training data can still be confidently wrong, and the rescaling will preserve that confident error intact. Within those bounds, though, the payoff is large for the price. Every decision built on a confidence score — refusing to answer below a threshold, routing hard cases to a person, weighing one model's vote against another's — silently assumes the score means what it says. Making that assumption true costs a validation set and a few minutes of fitting, after the expensive training is already done.