Dynamic Graph Regularized Broad Learning With Marginal Fisher Representation for Noisy Data Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Broad learning system (BLS) offers an effective, non-deep neural network architecture but is susceptible to noisy data.
• Existing robust broad learning models often fail to adequately handle noise and outliers, leading to suboptimal feature representation.
• This fragility necessitates the development of more resilient approaches for real-world applications with imperfect data.

Purpose of the Study:

• To propose a novel discriminative and robust network, dynamic graph regularized broad learning (DGBL), specifically designed for noisy data classification.
• To introduce a new algorithm, robust and dynamic marginal fisher analysis (RDMFA), for effective noise elimination prior to feature mapping.
• To enhance the discrimination capacity of broad learning models by incorporating dynamic graph regularization.

Main Methods:

• The proposed DGBL model utilizes marginal fisher representation for robust classification.
• A key innovation is the robust and dynamic marginal fisher analysis (RDMFA) algorithm, which extracts representations from a latent clean data space by dynamically generating graphs.
• The dynamic graphs derived from RDMFA are integrated as regularization terms within the DGBL framework.

Main Results:

• Extensive experiments on benchmark datasets demonstrate the superior performance of the proposed DGBL model.
• The RDMFA algorithm effectively mitigates the impact of noise and outliers, leading to more informative feature representations.
• DGBL significantly outperforms several state-of-the-art methods in noisy data classification tasks.

Conclusions:

• The proposed DGBL model, enhanced by RDMFA, offers a robust and discriminative solution for noisy data classification.
• The dynamic graph regularization effectively improves the feature discrimination capabilities of the broad learning system.
• This approach represents a significant advancement in handling noisy data within the broad learning framework.