Active Learning for Optimal Minimization of Experimental Characterization Uncertainty

Collecting experimental measurements is rarely an end in itself; rather, measurements inform key outcome statistics. Standard active learning procedures can drive a cumulative decrease in measurement uncertainty, but do not account for the uncertainty of the outcome. Here we present an active learning framework that operates to minimize the uncertainty of the outcome, and demonstrate its applicability with imaging and spectroscopic tasks. We show how our framework can effectively select regions for measurement without iteratively retraining a model. We conclude with two instances where our framework has outperformed standard active learning procedures to accelerate the classification of unknown samples.