Co-Training Broad Siamese-Like Network for Coupled-View Semi-Supervised Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multiview semi-supervised learning utilizes cross-view knowledge to address limited labeled data.
• Current deep learning methods are time-consuming due to complex structures and backpropagation.
• Efficient methods are needed to improve semi-supervised classification accuracy and reduce training time.

Purpose of the Study:

• Propose a novel Co-training broad Siamese-like network (Co-BSLN) for coupled-view semi-supervised classification.
• Develop a faster and more accurate alternative to existing deep learning approaches.
• Demonstrate the effectiveness of Co-BSLN on popular datasets.

Main Methods:

• Co-BSLN employs a shallow network based on the Broad Learning System (BLS) for simplified structure.
• Replaces backpropagation with direct pseudo-inverse calculation for reduced training time.
• Utilizes cross-view consistency by treating different views of the same instance as positive pairs, guiding training via logit vector mapping.

Main Results:

• Co-BSLN significantly reduces training time compared to deep learning methods.
• The proposed method achieves improved classification accuracy on benchmark datasets.
• Feature concatenation enables Co-BSLN's application to general multiview data.

Conclusions:

• Co-BSLN provides an efficient and effective solution for multiview semi-supervised classification.
• The simplified network architecture and direct computation offer substantial speed advantages.
• Co-BSLN successfully leverages cross-view consistency to enhance learning performance.