[Paper Review] A Survey on Uncertainty Quantification Methods for Deep Learning

Wenchong He, Zhe Jiang+3 more|arXiv (Cornell University)|Feb 26, 2023

Adversarial Robustness in Machine Learning26 citations

TL;DR

This survey categorizes DNN uncertainty quantification methods by uncertainty sources (data vs. model) and discusses their pros, cons, and applications.

ABSTRACT

Deep neural networks (DNNs) have achieved tremendous success in computer vision, natural language processing, and scientific and engineering domains. However, DNNs can make unexpected, incorrect, yet overconfident predictions, leading to serious consequences in high-stakes applications such as autonomous driving, medical diagnosis, and disaster response. Uncertainty quantification (UQ) estimates the confidence of DNN predictions in addition to their accuracy. In recent years, many UQ methods have been developed for DNNs. It is valuable to systematically categorize these methods and compare their strengths and limitations. Existing surveys mostly categorize UQ methodologies by neural network architecture or Bayesian formulation, while overlooking the uncertainty sources each method addresses, making it difficult to select an appropriate approach in practice. To fill this gap, this paper presents a taxonomy of UQ methods for DNNs based on uncertainty sources (e.g., data versus model uncertainty). We summarize the advantages and disadvantages of each category, and illustrate how UQ can be applied to machine learning problems (e.g., active learning, out-of-distribution robustness, and deep reinforcement learning). We also identify future research directions, including UQ for large language models (LLMs), AI-driven scientific simulations, and deep neural networks with structured outputs.

Motivation & Objective

• Motivate trustworthy AI by addressing overconfident DNN predictions in high-stakes domains.
• Propose a taxonomy of UQ methods based on uncertainty sources (data vs. model) rather than solely on architectures or Bayesian viewpoints.
• Summarize advantages and disadvantages of UQ methods within each uncertainty-source category.
• Connect UQ taxonomy to applications such as active learning, robustness, and reinforcement learning.
• Identify research gaps and outline future directions in UQ for DNNs.

Proposed method

• Present a systematic taxonomy of UQ methods categorized by uncertainty sources: data uncertainty, model uncertainty, and their combination.
• Describe data uncertainty as aleatoric and model uncertainty as epistemic, with representations for each (discriminative vs. generative for data; Bayesian neural networks, ensembles, and sample-density models for data).
• Discuss Bayesian neural networks with posterior approximation techniques (variational inference, Laplace approximation, MCMC) and their trade-offs.
• Explain ensemble and evidential learning approaches as alternatives to full Bayesian treatment for capturing uncertainty.
• Highlight challenges in calibrating predictive uncertainty and in scaling posterior approximations to modern DNNs.

Experimental results

Research questions

• RQ1 How can uncertainty in DNN predictions be systematically categorized by uncertainty sources (data vs. model) rather than architectures or Bayesian perspective?
• RQ2 What are the relative advantages and limitations of methods that address data uncertainty, model uncertainty, or their combination across different applications?
• RQ3 How can the proposed uncertainty-source taxonomy guide method selection in tasks like active learning, robustness, and reinforcement learning?
• RQ4 What are the current gaps and promising directions for uncertainty quantification in deep learning?
• RQ5 How do data and model uncertainties interact in high-stakes domains such as medicine, geoscience, and transportation?

Key findings

• A new taxonomy organizes UQ methods by uncertainty sources rather than solely by model architecture or Bayesian framing.
• Data uncertainty is treated via discriminative and generative approaches that model output distributions to reflect inherent data noise.
• Model uncertainty is addressed through Bayesian neural networks, ensembles, and sample-density aware models to capture parameter, architecture, and distribution-shift uncertainties.
• Combining data and model uncertainty provides a comprehensive view but increases computation and storage demands; evidential deep learning offers a more efficient alternative.
• The survey connects UQ taxonomy to real-world domains (medicine, geoscience, transportation, biochemistry) and discusses application-specific challenges and future directions.