Jo-SNC: Combating Noisy Labels Through Fostering Self- and Neighbor-Consistency.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Label noise presents a significant challenge in supervised deep learning, as deep networks are prone to memorizing corrupted samples.
• Existing methods for handling label noise often overlook imbalances in noise distribution across mini-batches and struggle with out-of-distribution noisy data.

Purpose of the Study:

• To develop a robust method, Jo-SNC (Joint sample selection and model regularization based on Self- and Neighbor-Consistency), to effectively address label noise in deep learning.
• To improve the identification and handling of both in-distribution and out-of-distribution noisy data.

Main Methods:

• Utilizing Jensen-Shannon divergence to assess sample likelihood of being clean or out-of-distribution, incorporating nearest neighbors for enhanced reliability.
• Implementing a self-adaptive, data-driven thresholding scheme for per-class selection and applying partial label learning for in-distribution noise and negative learning for out-of-distribution noise.
• Introducing triplet consistency regularization to bolster self-prediction, neighbor-prediction, and feature consistency.

Main Results:

• Jo-SNC demonstrates superior performance compared to state-of-the-art methods across various benchmark datasets.
• Extensive ablation studies confirm the effectiveness and robustness of the proposed approach in handling label noise.

Conclusions:

• Jo-SNC offers a significant advancement in noise-robust deep learning by effectively identifying and treating different types of noisy data.
• The method's ability to handle label noise, including out-of-distribution samples, makes it a valuable contribution to the field.