Improved Boundary Uncertainty Estimation For Classifier Selection
David Ha, Shigeru Katagiri, Miho Ohsaki · 2020 IEEE MTT-S International Wireless Symposium (IWS) · 2020
Classifier evaluation is a key component of optimal classifier design that traditionally requires validation data separate from the training data. In contrast, a recent classifier evaluation method enabled classifier evaluation directly on the training data, by introducing a new evaluation measure called boundary uncertainty. However, the boundary uncertainty estimation proposed by this method suffered costs that hindered its applicability and accuracy. Therefore, we propose an improved boundary uncertainty estimation method and demonstrate its high utility. Compared to the original method, our improved method is more accurate, transparently applicable, hyperparameter-free, and roughly 104times faster on the presented benchmark dataset.