A Geometric Method for Improved Uncertainty Estimation in Real-time

Uncertainty calibration is the process of adapting machine learning models' confidence estimations to be consistent with the actual success probability of the model [Guo et al., Machine learning classifiers are probabilistic in nature, 2017a]. The model's confidence evaluation on its classifications, and thus inevitably involve uncertainty. Predicting i.e., the model's prediction of the success ratio on a the probability of a specific input to be specific input, is an essential aspect of mission-critical machine correct is called uncertainty (or confidence) estimation learning applications as it provides a realistic estimate and is crucial for risk management. Posthoc of the classification's success probability and facilitates informed model calibrations can improve models' uncertainty decisions about the current situation. Even a very estimations without the need for retraining, accurate model may run into an unexpected situation, which and without changing the model.