Reliable & Efficient UQ (uncertainty quantification)
■ UQ via various regression models
▲ Deep ensembles (DE): deep learning-based regression model that can quantify uncertainty
■ Reliable UQ by uncertainty calibration
▲ Visualization of uncertainty calibration
(DE-bef: DE model before calibration & DE-aft: DE model after calibration )
▲ Uncertainty error criteria to investigate the effects of uncertainty calibration on DE models
(from BEF to AFT, uncertainty error criteria decrease significantly)
■ Comparison of GPR and DE
▲ DE model after uncertainty calibration outperforms GPR in terms of
regression performance (−56% NLL & −55% RMSE), reliability of UQ (−77% AUCE & −38% ENCE), and training efficiency (−78% time)