Can Bayesian Neural Networks Explicitly Model Input Uncertainty?

2025-01-14

AI Summary

Bayesian neural networks (BNNs) are widely assumed to inherently model input uncertainty—e.g., measurement noise or distributional shift—but this capability remains empirically unverified across common approximate inference methods. Method: We systematically investigate whether standard BNN approximations—ensembles, MC-Dropout, and Flipout—can explicitly capture input uncertainty, introducing a dual-input BNN architecture that ingests both mean and standard deviation of noisy inputs. Contribution/Results: We provide the first empirical evidence that only ensembles and Flipout reliably encode input uncertainty, whereas MC-Dropout fails catastrophically in this role. This finding clarifies the fundamental limitations of popular BNN approximations for input uncertainty quantification and corrects a longstanding misconception in the literature. Our results offer principled guidance for selecting, deploying, and certifying BNNs in safety-critical applications such as autonomous driving and medical diagnosis, where robust uncertainty awareness is essential.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Natural Language Processing: Safety and Robustness

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Inputs to machine learning models can have associated noise or uncertainties, but they are often ignored and not modelled. It is unknown if Bayesian Neural Networks and their approximations are able to consider uncertainty in their inputs. In this paper we build a two input Bayesian Neural Network (mean and standard deviation) and evaluate its capabilities for input uncertainty estimation across different methods like Ensembles, MC-Dropout, and Flipout. Our results indicate that only some uncertainty estimation methods for approximate Bayesian NNs can model input uncertainty, in particular Ensembles and Flipout.

Problem

• Bayesian Neural Networks
• Input Uncertainty
• Simplified Versions

Innovation

• Bayesian Neural Networks
• Uncertainty Quantification
• Flipout Method