A unified uncertainty-informed approach for risk management of deep learning models in the open world

Abstract

Equipping deep learning models with a principled uncertainty quantification (UQ) has become essential to ensure their reliable performance in open-world environments. To address uncertainty arising from two prevalent sources - distribution shifts and out-of-distribution (OOD) inputs, this paper presents a unified, uncertainty-informed approach for quantifying and managing the risks these factors pose to deep learning models. Toward this goal, we leverage a principled UQ approach, Spectral-normalized Neural Gaussian Process (SNGP), to quantify the epistemic uncertainty associated with model predictions. Unlike other UQ methods in the literature, SNGP offers two distinctive properties: (1) spectral normalization applied to hidden layer weights to preserve relative distances among data points throughout feature transformations, and (2) replacement of the output layer with a Gaussian process to produce distance-aware uncertainty estimates. Using the uncertainty estimates from SNGP, we employ Youden's index to derive an optimal threshold that categorizes predictions into different risk levels, enabling uncertainty-informed decision making. Experiments on two datasets of varying scale demonstrate that the proposed method facilitates effective risk assessment and management in open-world settings. Computational results show that the proposed method achieves predictive performance comparable to Monte Carlo dropout and deep ensembles, while providing more computationally efficient, consistent, and principled uncertainty estimates under no shift, distribution shift, and OOD conditions.