Broad Learning System Based on Maximum Correntropy Criterion.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Broad Learning System (BLS) is a powerful discriminative learning method for regression and classification.
• Standard BLS, based on Minimum Mean Square Error (MMSE), is sensitive to outliers, limiting its real-world applicability.
• Robustness is crucial for machine learning models operating in environments with noisy data.

Purpose of the Study:

• To enhance the robustness of the Broad Learning System (BLS) against data outliers.
• To introduce a novel correntropy-based BLS (C-BLS) leveraging the Maximum Correntropy Criterion (MCC).
• To develop efficient incremental learning algorithms for the proposed C-BLS.

Main Methods:

• The Maximum Correntropy Criterion (MCC) was adopted to train the output weights of the BLS, creating the C-BLS.
• Three incremental learning algorithms were developed for C-BLS, based on weighted regularized least-squares solutions.
• The proposed methods were evaluated on diverse regression and classification datasets.

Main Results:

• The correntropy-based BLS (C-BLS) demonstrated excellent robustness to outliers.
• C-BLS maintained the performance of the standard BLS in Gaussian or noise-free environments.
• Incremental learning algorithms enabled quick updates for C-BLS without full retraining.

Conclusions:

• The proposed C-BLS effectively addresses the outlier sensitivity of standard BLS.
• MCC provides a robust alternative criterion for training BLS, enhancing its practical utility.
• The developed incremental learning algorithms facilitate efficient adaptation and expansion of the C-BLS.