CoDC: Accurate Learning with Noisy Labels via Disagreement and Consistency

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) excel in many tasks but struggle with label noise.
• Existing co-teaching methods for noisy labels are sample-inefficient for generalization.
• Label noise can negatively impact DNN memorization and performance.

Purpose of the Study:

• To propose CoDC, a novel co-teaching method for accurate learning with label noise.
• To improve generalization performance in noisy scenarios using both disagreement and consistency strategies.
• To leverage knowledge from large-loss samples for enhanced learning.

Main Methods:

• CoDC maintains feature-level disagreement and prediction-level consistency using a balanced loss function.
• A weighted cross-entropy loss is utilized, informed by historical training data.
• Pseudo-labels are assigned to large-loss samples to utilize their knowledge.

Main Results:

• CoDC achieved 72.81% accuracy on the Clothing1M dataset.
• CoDC attained 76.96% Top-1 accuracy on the WebVision1.0 dataset.
• Experiments demonstrated superior performance across synthetic and real-world noise types.

Conclusions:

• CoDC effectively addresses the challenge of learning with noisy labels in DNNs.
• The proposed method shows robustness and improved generalization capabilities.
• CoDC offers a promising approach for reliable deep learning in noisy environments.