An intuitive training point sampling framework is presented, which biases towards training points whose labels have higher gradient norms. The sampling framework is applied to both active learning and coreset learning, which are both in turn applied to learn a SN2 reaction energy surface, molecular normal mode energies, and various analytical test functions. In all cases, up to a two fold improvement in out of sample error is observed. The framework's failure in higher dimensional spaces is explored, and explained.
Morris Trestman