Collaborative learning with corrupted labels.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel in supervised learning but require extensive, high-quality labeled data.
• Manual data annotation is costly, leading to the use of cheaper, low-quality datasets with corrupted labels.
• DNNs are prone to overfitting when trained on noisy or mislabeled datasets.

Purpose of the Study:

• To develop a robust deep learning method that mitigates the negative impact of corrupted labels.
• To improve the generalization performance of DNNs when trained on datasets with a significant proportion of mislabeled samples.
• To introduce a cost-effective approach for training DNNs on readily available, albeit noisy, datasets.

Main Methods:

• A novel collaborative learning (co-learning) framework is proposed, utilizing a deep network with two distinct branches.
• A relabeling mechanism is integrated to identify and correct mislabeled samples within the dataset.
• The co-learning approach aims to recover true labels, preventing overfitting to noise and leveraging all available data.

Main Results:

• The proposed co-learning algorithm demonstrates high generalization performance even with substantial label corruption.
• Co-learning effectively prevents DNNs from overfitting to noisy labels by recovering accurate sample labels.
• Experimental results show consistent outperformance of co-learning against state-of-the-art methods on benchmark datasets.

Conclusions:

• Collaborative learning offers a simple yet effective solution for training DNNs on datasets with corrupted labels.
• The method enhances model robustness and generalization, making it suitable for real-world scenarios with imperfect data.
• Co-learning presents a promising direction for improving the practical applicability of deep learning models in data-scarce or noisy environments.