Uncertainty as a Predictor of Classification Accuracy in Machine Learning-Assisted Measurements [Measurement Methodology]

Like other sectors of the industry, instrumentation and measurement (I&M) in the past few years has witnessed the significant rise of Machine Learning (ML) in its domain [1]. In I&M, ML is used for indirect measurement, as explained in detail in [2]. Like all measurements, an ML-assisted measurement also needs to quantify its uncertainty. In fact, the uncertainty of ML's inference can be correlated to its classification accuracy [3]. The impact of this is significant, because it enables the system or its operator to use the uncertainty for catching less-trustworthy measurements. As such, quantifying, communicating, and mitigating the uncertainty in ML-assisted measurement systems is crucial for risk management and ensuring that the system is trustworthy and can be deployed in the real world [4]. This is why recent works have proposed that ML models be trained with loss functions that include the uncertainty [5], thereby minimizing the uncertainty and in turn increasing the classification accuracy.