Surrogate approach to uncertainty quantification of neural networks for regression

Uncertainty quantification is essential in preventing inaccurate predictions of neural networks. A vanilla neural network for regression does not intrinsically provide explicit information about pre-diction uncertainty. To quantify the prediction uncertainty for regression problems, we can build an alternative prediction model specialized for uncertainty quantification. However, this requires the use of training data, which are inaccessible in many real-world situations. To address such situations, this study presents a surrogate approach to quantify the prediction uncertainty of a regression network without using training data. A regression network tends to have high prediction uncertainty when its output is sensitive to its input. Based on this intuition, we quantify the sensitivity and use it as a surrogate measure of the prediction uncertainty. To do so, we introduce four surrogate measures that capture the sensitivity in different ways: Input perturbation, Gradient norm, MC-dropout, and Knowledge distillation. For a query instance, each surrogate measure can be calculated by using the regression network only to estimate the prediction uncertainty. We demonstrate the respective effectiveness of the proposed surrogate measures on nine regression datasets.& COPY; 2023 Elsevier B.V. All rights reserved.

Automated summary

Quantifying prediction uncertainty in neural networks is crucial and can be achieved by building specialized models or using surrogate measures. This study proposes a surrogate approach that quantifies prediction uncertainty without using training data. The proposed surrogate measures capture sensitivity and are effective in estimating prediction uncertainty in regression networks. The study demonstrates their effectiveness on nine regression datasets. © 2023 Elsevier B.V. All rights reserved.