Scholarlink

Convolutional Neural Networks (CNNs) are a fundamental architecture in deep learning, especially in computer vision. They excel in tasks like image classification and object detection. Still, despite their success, CNNs typically provide point estimates without quantifying uncertainty, which poses critical issues for safety-centric applications and model interpretability. This paper reviews various methods developed to quantify uncertainty in CNNs, including Bayesian Neural Networks, Monte Carlo Dropout, Deep Ensembles, Evidential Learning, Deterministic Uncertainty Quantification, Stochastic Weight Averaging-Gaussian, Spectral-normalized Neural Gaussian Processes, and Uncertainty-Aware CNNs. We discuss each method’s principles, advantages, and limitations and highlight their applicability in real-world scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Review of Uncertainty Quantification in Convolutional Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sina Montazeri,
  • Pooya Tavallali

摘要

Convolutional Neural Networks (CNNs) are a fundamental architecture in deep learning, especially in computer vision. They excel in tasks like image classification and object detection. Still, despite their success, CNNs typically provide point estimates without quantifying uncertainty, which poses critical issues for safety-centric applications and model interpretability. This paper reviews various methods developed to quantify uncertainty in CNNs, including Bayesian Neural Networks, Monte Carlo Dropout, Deep Ensembles, Evidential Learning, Deterministic Uncertainty Quantification, Stochastic Weight Averaging-Gaussian, Spectral-normalized Neural Gaussian Processes, and Uncertainty-Aware CNNs. We discuss each method’s principles, advantages, and limitations and highlight their applicability in real-world scenarios.