Explainability through uncertainty: Trustworthy decision-making with neural networks

TL;DR

This work proposes a general uncertainty framework, with contributions being threefold: uncertainty estimation in ML models is positioned as an XAI technique, giving local and model-specific explanations; classification with rejection is used to reduce misclassifications by bringing a human expert in the loop for uncertain observations.

Abstract

Uncertainty is a key feature of any machine learning model and is particularly important in neural networks, which tend to be overconfident. This overconfidence is worrying under distribution shifts, where the model performance silently degrades as the data distribution diverges from the training data distribution. Uncertainty estimation offers a solution to overconfident models, communicating when the output should (not) be trusted. Although methods for uncertainty estimation have been developed, they have not been explicitly linked to the field of explainable artificial intelligence (XAI). Furthermore, literature in operations research ignores the actionability component of uncertainty estimation and does not consider distribution shifts. This work proposes a general uncertainty framework, with contributions being threefold: (i) uncertainty estimation in ML models is positioned as an XAI technique, giving local and model-specific explanations; (ii) classification with rejection is used to reduce misclassifications by bringing a human expert in the loop for uncertain observations; (iii) the framework is applied to a case study on neural networks in educational data mining subject to distribution shifts. Uncertainty as XAI improves the model's trustworthiness in downstream decision-making tasks, giving rise to more actionable and robust machine learning systems in operations research.

Figures (10)

• Figure 1: Overview explainable artificial intelligence. Uncertainty estimation is a third general type of XAI technique. Figure adapted from bai2021explainable.
• Figure 2: General uncertainty framework. The framework consists of two stages: (i) uncertainty estimation as XAI and (ii) classification with rejection. It can be applied to multiple ML models, each having one or more specific uncertainty techniques.
• Figure 3: Two types of uncertainty. Observation A has high data uncertainty; B has high model uncertainty. Figure adapted from hullermeier2021aleatoric.
• Figure 4: Performance measures for classification with rejection. Three performance measures are proposed by condessa2017performance to find the optimal rejection point. Figure adapted from mena2021survey.
• Figure 5: Overview of uncertainty estimation methods. Forward passes are generated differently depending on the method. In MC dropout, different units are dropped out from a NN; in Deep Ensembles, multiple independent NNs are used, with different parameter initializations and noise in the SGD training process.