On Learning Label Noise Robust Networks via Regularization: A Topological View.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Topology

Background:

• Neural networks struggle with real-world label noise, which degrades generalization by disrupting local fit values.
• Existing regularization methods primarily address global constraints, neglecting the local impact of label noise.
• The specific influence of label noise on neural network function requires deeper investigation.

Purpose of the Study:

• To analyze the local effects of label noise on neural networks using a topological perspective.
• To introduce a novel regularization method, network boundary topology regularization (NBTR), to mitigate label noise.
• To improve the generalization performance and robustness of neural networks against label noise.

Main Methods:

• Developed network boundary topology regularization (NBTR) based on persistent homology.
• Focused on simplifying the topology of class boundaries to address local fit value disturbances.
• Conducted extensive experiments across diverse datasets, network architectures, and noise types.

Main Results:

• NBTR effectively reduces the network's tendency to memorize label noise.
• The method surpasses strong baselines in generalization accuracy, particularly under asymmetric noise conditions (average improvement of 7.72%).
• NBTR enhances the anti-noise capabilities of traditional methods when used complementarily.

Conclusions:

• Network boundary topology regularization (NBTR) offers a novel and effective approach to address the local impacts of label noise in neural networks.
• The topological perspective provides valuable insights into understanding and mitigating label noise.
• NBTR demonstrates significant improvements in generalization and robustness, making it a promising technique for real-world applications.