Active learning with label quality control.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural network training demands extensive labeled data, incurring high costs through manual annotation and quality control.
• Current active learning strategies aim to optimize sample selection for labeling but can be further refined for efficiency.

Purpose of the Study:

• To develop a practical active learning method that significantly reduces overall data labeling costs.
• To improve the efficiency of labeling by adaptively allocating resources to valuable unlabeled and potentially mislabeled samples.

Main Methods:

• Designed an active learning method that dynamically assigns labeling resources.
• Incorporated a strategy to identify and re-evaluate potentially mislabeled samples.
• Developed a theoretical guarantee that limits redundant sample labeling within batches (probability < 1/k).

Main Results:

• Achieved state-of-the-art results on benchmark datasets.
• Demonstrated strong performance in a real-world industrial application for automatic optical inspection.
• Significantly reduced labeling costs compared to traditional methods.

Conclusions:

• The proposed active learning method offers a cost-effective solution for deep neural network training.
• The adaptive resource allocation and mislabeled sample identification contribute to reduced labeling expenses.
• The method shows practical utility in both academic benchmarks and industrial settings.