Broad Learning System under Label Noise: A Novel Reweighting Framework with Logarithm Kernel and Mixture Autoencoder.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Broad Learning System (BLS) shows high performance but is sensitive to label noise.
• Label noise significantly degrades the performance of BLS models.
• Robustness in noisy datasets is crucial for real-world applications.

Purpose of the Study:

• To enhance the robustness of BLS against label noise.
• To develop a novel BLS variant using Logarithm Kernel (LK) for improved sample weighting.
• To introduce a Mixture Autoencoder (MAE) for better feature representation in noisy image databases.

Main Methods:

• Designed a Logarithm Kernel (LK) function to reweight samples during BLS training, creating Logarithm Kernel-based BLS (L-BLS).
• Developed a Mixture Autoencoder (MAE) to generate more representative feature nodes for BLS in complex label noise scenarios.
• Proposed MAEBLS and L-MAEBLS variants integrating MAE with BLS and L-BLS.

Main Results:

• Extensive experiments validated the robustness and effectiveness of the proposed L-BLS.
• The MAE demonstrated its capability to provide more representative feature nodes for BLS.
• L-MAEBLS showed superior performance in handling complex label noise environments.

Conclusions:

• L-BLS offers a robust alternative to standard BLS in the presence of label noise.
• MAE significantly improves feature representation for BLS, especially in challenging noisy datasets.
• The proposed methods provide effective solutions for robust learning in noisy environments.