Fast and reliable uncertainty quantification with neural network ensembles for industrial image classification

2024-03-15

Annals of Operations Research

AI Summary

To address the challenge of efficiently and reliably quantifying model uncertainty under out-of-distribution (OOD) data in industrial image classification, this paper proposes a lightweight ensemble training and uncertainty calibration co-optimization framework. The method jointly integrates heterogeneous neural network ensembles, temperature scaling calibration, Monte Carlo DropPath, and distribution-consistency regularization to simultaneously improve confidence calibration accuracy and OOD detection robustness with minimal inference overhead. Evaluated on multiple industrial defect datasets, it reduces Expected Calibration Error (ECE) by over 40%, improves OOD detection F1-score by 12%, and incurs less than 8% additional inference latency. Its core innovation lies in the synergistic modeling of ensemble diversity, calibration mechanisms, and distribution-alignment regularization—enabling high-accuracy, low-latency, and robust uncertainty quantification. This significantly enhances safety and interpretability in mission-critical industrial decision-making.

Technology Category

Machine Learning: Calibration & Uncertainty QuantificationReasoning under Uncertainty: Uncertainty RepresentationsComputer Vision: Adversarial Attacks & Robustness

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingGraph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphsSystems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Problem

• Neural Networks
• Out-of-Distribution Data
• Uncertainty Quantification

Innovation

• Neural Network Ensembles
• Efficiency and Resource Saving
• Diversity Quality Indicator