Prediction Consistency Regularization for Learning with Noise Labels Based on Contrastive Clustering.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Label noise significantly degrades neural network performance in classification tasks by disrupting prediction consistency and reducing accuracy.
• Identifying similar samples is crucial for effective noise mitigation but remains a primary challenge in existing methods.

Purpose of the Study:

• To develop a novel prediction consistency regularization technique to mitigate the impact of label noise on neural network classification.
• To address the challenge of identifying similar samples by formalizing it as a clustering problem.

Main Methods:

• Employed twin contrastive clustering (TCC) for similar sample identification, enhanced by adjusting clustering priors with label information.
• Constructed cluster prototypes based on adjusted TCC results and formulated a prototype-based regularization term to improve prediction consistency.
• Evaluated the method's effectiveness on benchmark datasets across various noise rates.

Main Results:

• Demonstrated significant enhancement in classification accuracy under different label noise scenarios.
• Confirmed that the proposed regularization term effectively mitigates the adverse effects of label noise.
• Showcased that the adjusted TCC improves the quality of similar sample recognition.

Conclusions:

• The novel prediction consistency regularization effectively combats label noise in classification tasks.
• The integration of adjusted TCC and prototype-based regularization offers a robust solution for improving model performance in noisy data environments.