Streamlined and Resource-Efficient Estimation of Epistemic Uncertainty in Deep Ensemble Classification Decision via Regression

Abstract

Ensemble deep learning (EDL) has emerged as a leading tool for epistemic uncertainty quantification (UQ) in predictive modelling. Our study focuses on the utilization of EDL, composed of auto-encoders (AEs) for out-of-distribution (OoD) detection. EDL offers straightforward interpretability and valuable practical insights. Conventionally, employing multiple AEs in an ensemble requires regular training for each model whenever substantial changes occur in the data, a process that can become computationally expensive, especially when dealing with large ensembles. To address this computational challenge, we introduce an innovative strategy that treats ensemble UQ as a regression problem. During initial training, once the uncertainty distribution is established, we map this distribution to one ensemble member. This approach ensures that during subsequent trainings and inferences, only one ensemble member and the regression model are needed to predict uncertainties, eliminating the need to maintain the entire ensemble. This streamlined approach is particularly advantageous for systems with limited computational resources or situations that demand rapid decision-making, such as alert management in cybersecurity. Our evaluations on five benchmark OoD detection data sets demonstrate that the uncertainty estimates obtained with our proposed method can, in most cases, align with the uncertainty distribution learned by the ensemble, all while significantly reducing the computational resource requirements.