expected calibration error
In plain wordsA single number measuring how far a model's confidence is from its actual hit rate, averaged across confidence levels; zero means its confidence can be taken at face value.
A summary measure of miscalibration (calibrationcalibration, n.: How well a model's stated confidence matches reality: when a well-calibrated system says it is 80% sure, it turns out to be right about 80% of the time.) for classifiers. Predictions are grouped into M bins by confidence, the probability given to the predicted class, typically 10 or 15 equal-width bins; in each bin the absolute gap between average confidence and observed accuracyaccuracy, n.: The share of a model's answers that are right, such as 95 out of 100. Simple and intuitive, but misleading when one kind of answer is far more common than the others. is computed, and ECE is the average of these gaps weighted by the share of examples in each bin. The maximum calibration error (MCE) takes the largest gap instead. ECE summarizes the deviations visible in a reliability diagram and became the standard calibration metric for deep networks and language modelslarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code.. It has known flaws: estimates depend on the number and placement of bins (equal-mass binning and debiased estimators help); it looks only at top-class confidence; it can be near zero for a useless model that always predicts the base rate; and it is not a proper scoring ruleproper scoring rule, n.: A way of grading probability forecasts that gives the best expected score only to a forecaster who reports what they really believe, so there is nothing to gain by exaggerating or hedging., so it should be reported alongside log losslog loss, n.: A score for probability predictions that punishes confident mistakes severely: saying '99% sure' and being wrong costs far more than saying '60% sure' and being wrong. Lower is better. or the Brier scoreBrier score, n.: A score for probability forecasts: the average squared gap between each forecast probability and what actually happened (1 if it did, 0 if not). Zero is perfect; lower is better..
Instruction tuning raised the model's ECE on trivia questions from 0.04 to 0.12.
Origin
Defined in its now-standard binned form by Mahdi Pakdaman Naeini, Gregory Cooper and Milos Hauskrecht in 'Obtaining Well Calibrated Probabilities Using Bayesian Binning' (AAAI 2015). Chuan Guo and colleagues' 2017 study of overconfident neural networks made it the default calibration metric, and critiques of its binning bias, such as Jeremy Nixon and colleagues' 'Measuring Calibration in Deep Learning' (2019), followed.
Related
- calibrationcalibration, n.: How well a model's stated confidence matches reality: when a well-calibrated system says it is 80% sure, it turns out to be right about 80% of the time.
- Brier scoreBrier score, n.: A score for probability forecasts: the average squared gap between each forecast probability and what actually happened (1 if it did, 0 if not). Zero is perfect; lower is better.
- log losslog loss, n.: A score for probability predictions that punishes confident mistakes severely: saying '99% sure' and being wrong costs far more than saying '60% sure' and being wrong. Lower is better.
- proper scoring ruleproper scoring rule, n.: A way of grading probability forecasts that gives the best expected score only to a forecaster who reports what they really believe, so there is nothing to gain by exaggerating or hedging.
- accuracyaccuracy, n.: The share of a model's answers that are right, such as 95 out of 100. Simple and intuitive, but misleading when one kind of answer is far more common than the others.
- abstentionabstention, n.: When an AI declines to answer because it is not sure enough — saying 'I don't know' or asking for more information — instead of guessing and risking a confident mistake.
Sources
- Naeini, M. P., Cooper, G. F. & Hauskrecht, M. (2015). Obtaining Well Calibrated Probabilities Using Bayesian Binning. Proc. AAAI 2015.
- Guo, C., Pleiss, G., Sun, Y. & Weinberger, K. Q. (2017). On Calibration of Modern Neural Networks. Proc. ICML 2017.
- Nixon, J., Dusenberry, M., Zhang, L., Jerfel, G. & Tran, D. (2019). Measuring Calibration in Deep Learning. Proc. CVPR Workshops 2019.
Mentioned in
Brier scoreBrier score, n.: A score for probability forecasts: the average squared gap between each forecast probability and what actually happened (1 if it did, 0 if not). Zero is perfect; lower is better., calibrationcalibration, n.: How well a model's stated confidence matches reality: when a well-calibrated system says it is 80% sure, it turns out to be right about 80% of the time., proper scoring ruleproper scoring rule, n.: A way of grading probability forecasts that gives the best expected score only to a forecaster who reports what they really believe, so there is nothing to gain by exaggerating or hedging.