When Broad Learning System Meets Label Noise Learning: A Reweighting Learning Framework.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Broad Learning System (BLS) offers efficient learning and expansion but is vulnerable to noisy data.
• Current robust BLS models use scalar weights, potentially discarding useful information from noisy samples.

Purpose of the Study:

• To introduce a novel Broad Learning System with adaptive reweighting (BLS-AR) for improved data classification accuracy in the presence of label noise.
• To address the limitations of existing methods that disregard valuable information within noisy samples.

Main Methods:

• The BLS-AR employs an element-level reweighting strategy, assigning a weight vector to each sample to indicate element-wise noise levels.
• This approach allows for precise identification and down-weighting of noisy elements while emphasizing informative ones.
• The model's separability facilitates efficient training of subnetworks and development of incremental learning algorithms.

Main Results:

• The BLS-AR demonstrates superior effectiveness and robustness in classifying data with label noise compared to existing methods.
• Element-level reweighting enables more accurate representation learning by selectively utilizing informative data points.
• Experimental results validate the proposed model's performance and resilience to noise.

Conclusions:

• The BLS-AR strategy significantly enhances the performance of Broad Learning Systems in noisy data classification tasks.
• Element-level adaptive reweighting is a promising approach for robust deep learning models.
• The developed incremental learning algorithms support the scalability and adaptability of the BLS-AR model.