Maximum total correntropy-based broad learning system with robust M-estimator.

Area of Science:

• Machine Learning
• Robust Statistics

Background:

• Broad Learning System (BLS) excels in various applications but struggles with noisy training data.
• Existing BLS variants often show performance degradation when input data contains noise.

Purpose of the Study:

• To develop a noise-resilient Broad Learning System (BLS) capable of handling both input and output noise.
• To introduce novel methods that mitigate the impact of kernel width selection in entropy-based criteria.

Main Methods:

• Proposed Maximum Total Correntropy-based BLS (MTC-BLS) utilizing the Maximum Total Correntropy (MTC) criterion.
• Introduced MMTC-BLSa by incorporating the M-estimator into the objective function for dual constraint optimization.
• Developed MMTC-BLSb using the M-estimator as a weighting factor in the MTC criterion.
• Employed fixed-point iteration methods for efficient training of the proposed algorithms.

Main Results:

• MTC-BLS effectively handles both input and output noise, outperforming standard BLS.
• MMTC-BLSa and MMTC-BLSb demonstrate reduced sensitivity to kernel width selection deviations.
• Experimental validation on time series, regression, and image datasets confirms the superiority of the proposed methods.

Conclusions:

• The proposed MTC-BLS, MMTC-BLSa, and MMTC-BLSb offer robust and efficient solutions for noisy data in machine learning.
• These methods enhance the reliability and applicability of BLS in real-world scenarios with imperfect data.