IRNet: Iterative Refinement Network for Noisy Partial Label Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Partial Label Learning (PLL) is a form of weakly supervised learning.
• A key assumption in PLL is that the true label is within the candidate set.
• This assumption is often violated due to annotator error, leading to noisy PLL.

Purpose of the Study:

• To address the challenge of noisy Partial Label Learning (PLL).
• To propose a novel framework, the Iterative Refinement Network (IRNet), for noisy PLL.
• To relax the assumption that the ground-truth label must be in the candidate set.

Main Methods:

• Developed the Iterative Refinement Network (IRNet) framework.
• Incorporated noisy sample detection and label correction modules within IRNet.
• Utilized smoothness constraints to minimize prediction errors in the modules.

Main Results:

• IRNet demonstrated the ability to reduce dataset noise levels.
• Theoretical analysis showed IRNet approximates the Bayes optimal classifier.
• IRNet serves as a plug-in strategy compatible with existing PLL methods.

Conclusions:

• IRNet significantly outperforms state-of-the-art methods on noisy PLL tasks.
• The proposed framework offers a robust solution for handling label noise in PLL.
• Experimental validation on benchmark datasets confirms IRNet's effectiveness.