Multidimensional Uncertainty Quantification for Deep Neural Networks

Xujiang ZhaoarXiv 2023 · OpenReview

Abstract

Deep neural networks (DNNs) have achieved remarkable success across diverse domains, yet their tendency toward overconfidence remains a critical barrier to deployment in high-stakes environments. Standard DNNs often fail to distinguish between different sources of uncertainty, such as a lack of training data (vacuity) versus conflicting features (dissonance). This research develops a comprehensive framework for multidimensional uncertainty quantification. We introduce methods for Graph Neural Networks (GNNs) to account for structural evidence, a robust semi-supervised learning (SSL) approach to mitigate out-of-distribution (OOD) interference, and a temporal evidential framework for early event detection. Our findings demonstrate that decomposing uncertainty into specific types significantly improves model reliability, misclassification detection, and robustness against noisy data.

Introduction

In applications ranging from autonomous driving to medical diagnostics, a model’s ability to say "I don't know" is as important as its predictive accuracy. However, traditional deep learning models typically output a single probability score via a softmax layer. This approach collapses all forms of uncertainty into a single metric, making it impossible to determine if a low-confidence prediction is due to noisy input data or an entirely unfamiliar scenario.

This research addresses the challenge of "uncertainty decomposition" by leveraging Subjective Logic (SL). Unlike standard probability, SL allows models to quantify distinct components of uncertainty: vacuity (uncertainty due to a lack of evidence or ignorance) and dissonance (uncertainty due to conflicting evidence). By identifying the root causes of uncertainty, we can build AI systems that are more transparent and resilient.

Quantifying Uncertainty in Graph Neural Networks

Existing Graph Neural Networks (GNNs) often ignore how the underlying graph structure influences a model's confidence. This work introduces Graph-based Kernel Dirichlet distribution Estimation (GKDE), a novel framework that estimates evidence based on the "distance" between nodes.

Through mathematical analysis (Proposition 1), we demonstrate that graph distance serves as a reliable proxy for vacuity; nodes located far from the training data naturally exhibit higher vacuity. Our experiments on citation and social graphs (such as Cora and Citeseer) show that our proposed model, S-BGCN-T-K, outperforms state-of-the-art GNNs in both AUROC and AUPR for misclassification detection. A key insight from this work is that dissonance is the most effective metric for identifying potential misclassifications, while vacuity is superior for detecting Out-of-Distribution (OOD) samples.

Robust Semi-Supervised Learning against OOD Data

Semi-supervised learning (SSL) relies on unlabeled data to improve performance, but it is highly vulnerable to "poisoning" if that unlabeled data contains OOD samples. We identified a specific vulnerability in standard Batch Normalization (BN): OOD data points shift the mean and variance of the entire batch, significantly degrading accuracy.

To solve this, we developed Weighted Batch Normalization (WBN) and a bi-level optimization meta-learning approach. By automatically learning weights for unlabeled data, the model can ignore "noisy" or OOD samples during the normalization process. In tests using CIFAR and SVHN datasets, our "Ours-VAT" method maintained stable performance even when 75% of the unlabeled data consisted of OOD samples, a scenario where traditional SSL models typically fail.

Temporal Uncertainty for Early Event Detection

In time-series analysis, such as audio or video monitoring, models often make premature and incorrect guesses before sufficient evidence has arrived. We proposed the Multi-Label Temporal Evidential Neural Network (MTENN) to address this.

MTENN predicts a Beta distribution for every label, allowing the model to track uncertainty over time. By using novel uncertainty fusion operators, the model can wait until vacuity drops below a certain threshold before committing to a detection. This approach, combined with Uncertainty Mean Scan Statistics (UMSS), significantly reduces detection delay in human action localization without increasing false positives. Furthermore, the framework is highly efficient, achieving an inference time of just 5ms for audio segments, making it viable for real-time industrial applications.

Conclusion and Future Directions

This research proves that entropy is a blunt instrument for uncertainty quantification. By splitting uncertainty into vacuity and dissonance, we provide DNNs with a fine-grained understanding of their own limitations. While Dirichlet and Beta distributions provide a strong foundation, future work will explore more complex distributions, such as Logistic Normal or mixture models, to capture even more nuanced types of uncertainty like "vagueness."

Ultimately, multidimensional uncertainty quantification paves the way for safer AI integration in Reinforcement Learning and autonomous systems, where knowing when to explore (high vacuity) versus when to refine a strategy (high dissonance) is essential for success.

Summary updated

Abstract

Deep neural networks (DNNs) have received tremendous attention and achieved great success in various applications, such as image and video analysis, natural language processing, recommendation systems, and drug discovery. However, inherent uncertainties derived from different root causes have been realized as serious hurdles for DNNs to find robust and trustworthy solutions for real-world problems. A lack of consideration of such uncertainties may lead to unnecessary risk. For example, a self-driving autonomous car can misdetect a human on the road. A deep learning-based medical assistant may misdiagnose cancer as a benign tumor. In this work, we study how to measure different uncertainty causes for DNNs and use them to solve diverse decision-making problems more effectively. In the first part of this thesis, we develop a general learning framework to quantify multiple types of uncertainties caused by different root causes, such as vacuity (i.e., uncertainty due to a lack of evidence) and dissonance (i.e., uncertainty due to conflicting evidence), for graph neural networks. We provide a theoretical analysis of the relationships between different uncertainty types. We further demonstrate that dissonance is most effective for misclassification detection and vacuity is most effective for Out-of-Distribution (OOD) detection. In the second part of the thesis, we study the significant impact of OOD objects on semi-supervised learning (SSL) for DNNs and develop a novel framework to improve the robustness of existing SSL algorithms against OODs. In the last part of the thesis, we create a general learning framework to quantity multiple uncertainty types for multi-label temporal neural networks. We further develop novel uncertainty fusion operators to quantify the fused uncertainty of a subsequence for early event detection.
BibTeX
@misc{zhao2023multidimensional,
  author = {Xujiang Zhao},
  title = {{Multidimensional Uncertainty Quantification for Deep Neural Networks}},
  year = {2023},
  eprint = {2304.10527},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.10527},
}
Figure