BASS: Broad Network Based on Localized Stochastic Sensitivity.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Standard Broad Learning System (BLS) training optimizes output weights by minimizing mean square error (MSE) and penalty terms.
• This optimization can degrade BLS generalization and robustness in noisy or complex environments, especially with input data perturbations.

Purpose of the Study:

• To propose a novel algorithm, Broad network based on localized stochastic sensitivity (BASS), to enhance noise robustness and generalization.
• To address limitations of standard BLS in handling noisy and perturbed input data.

Main Methods:

• Introduced localized stochastic sensitivity (LSS) to improve network robustness by considering unseen samples near training data.
• Developed three incremental learning algorithms for efficient BASS updates without full retraining.
• Evaluated BASS on 13 benchmark datasets for regression and classification tasks.

Main Results:

• BASS demonstrated superior accuracy on various regression and classification problems compared to standard methods.
• On ImageNet (ILSVRC2012), BASS achieved 1% higher Top-1 accuracy than AlexNet using significantly fewer parameters (12.6 million vs. 60 million).
• Extensive experiments confirmed BASS's enhanced generalization capability with noisy and perturbed data.

Conclusions:

• BASS effectively tackles noise and input perturbations by leveraging localized stochastic sensitivity.
• The proposed incremental learning algorithms enable efficient model updates.
• BASS offers a more robust and accurate alternative to standard BLS, particularly in challenging data environments.