An unsupervised parameter learning model for RVFL neural network.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Neural Networks

Background:

• Random Vector Functional Link (RVFL) networks are effective single-hidden layer feedforward neural networks (SLFNs) known for fast learning.
• A key limitation of RVFL networks is the performance degradation caused by randomly initialized network parameters.
• Existing RVFL implementations often struggle with suboptimal performance due to reliance on random parameter assignment.

Purpose of the Study:

• To propose a novel unsupervised network parameter learning method for RVFL networks.
• To enhance the performance of RVFL networks by adaptively learning superior network parameters.
• To address the challenge of randomly assigned parameters in RVFL networks.

Main Methods:

• Introduced a sparse pre-trained random vector functional link (SP-RVFL) network.
• Utilized a sparse autoencoder with L1-norm regularization for adaptive parameter learning.
• Employed unsupervised learning to embed input data information into network parameters.

Main Results:

• The proposed SP-RVFL method adaptively learns superior network parameters, embedding valuable input data information.
• SP-RVFL effectively alleviates the issue of randomly generated parameters, leading to improved algorithmic performance.
• Experiments on 16 diverse benchmarks confirm the effectiveness of SP-RVFL and demonstrate its superiority over extreme learning machine (ELM).

Conclusions:

• SP-RVFL offers a robust solution for optimizing RVFL network parameters through unsupervised learning.
• The method enhances the generalizability and performance of RVFL networks across various domains.
• SP-RVFL represents a significant advancement in improving the practical applicability and effectiveness of SLFNs.