Building Calibrated Deep Models via Uncertainty Matching with Auxiliary Interval Predictors

Abstract

With rapid adoption of deep learning in critical applications, the question of when and how much to trust these models often arises, which drives the need to quantify the inherent uncertainties. While identifying all sources that account for the stochasticity of models is challenging, it is common to augment predictions with confidence intervals to convey the expected variations in a model's behavior. We require prediction intervals to be well-calibrated, reflect the true uncertainties, and to be sharp. However, existing techniques for obtaining prediction intervals are known to produce unsatisfactory results in at least one of these criteria. To address this challenge, we develop a novel approach for building calibrated estimators. More specifically, we use separate models for prediction and interval estimation, and pose a bi-level optimization problem that allows the former to leverage estimates from the latter through an uncertainty matching strategy. Using experiments in regression, time-series forecasting, and object localization, we show that our approach achieves significant improvements over existing uncertainty quantification methods, both in terms of model fidelity and calibration error.

Designing accurate emulators for scientific processes using calibration-driven deep models

“…Similarly, Tagasovska et al recently developed a conditional quantile-based estimator for measuring aleatoric uncertainties 45 . Due to the lack of suitable evaluation mechanisms for validating the quality of these estimates, it is common to utilize empirical calibration as a quality metric 20,[46][47][48][49] . Interestingly, it has been reported in several studies that these estimators are not inherently well-calibrated 47 .…”

“…Following 56 , for each sample, we make T forward passes with the dropout rate set to τ and obtain the final prediction as the average from the T runs. This is known to produce more robust estimates in regression problems 20 . In our experiments, we set T = 20 and the dropout rate τ = 0.3. ig.…”

“…In this paper, we present Learn-by-Calibrating (LbC), a nonparametric approach based on interval calibration for building emulators in scientific applications that are effective even with heterogeneous data and are robust to outliers. The notion of interval calibration comes from the uncertainty quantification literature 20,21 and can be formally defined as follows: let us assume that the model f is designed to produce prediction intervals, in lieu of simple point estimates, for the response y, i.e., ½ŷ À δ l ;ŷ þ δ u . Suppose that the likelihood for the true response y to be contained in the prediction interval is pðŷ À δ l ≤ y ≤ŷ þ δ u Þ, the intervals are considered to be well-calibrated if the likelihood matches the expected confidence level.…”

Training Calibration-based Counterfactual Explainers for Deep Learning Models in Medical Image Analysis

“…To this end, uncertainty estimation methods are being adopted to determine the deficiencies of a model and/or the training data 13 . Meaningful uncertainties can play a crucial role in supporting practical objectives that range from assessing regimes of over (or under)-confidence and active data collection, to ultimately improving the predictive models themselves 14 . However, in practice, uncertainties are known to be challenging to communicate to decision-makers 15 , and the robustness of decisions with respect to uncertainties can vary considerably between use-cases 16 .…”

Uncertainty-Aware Boosted Ensembling in Multi-Modal Settings

“…As such, generating confidence intervals or uncertainty estimates along with the predictions is crucial for reliable § Equal contribution and safe deployment of machine learning systems in safetycritical settings (such as healthcare) [4]- [7]. It is critical to understand what a model does not know when building and deploying machine learning systems to help mitigate possible risks and biases in decision making [8].…”