A novel correlation Gaussian process regression-based extreme learning machine.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Statistics

Background:

• Extreme Learning Machine (ELM) prediction performance is sensitive to random initialization.
• Gaussian Process Regression (GPR) integration into ELM (GPRELM) aims to improve robustness.
• Kernel-based GPRELM (kGPRELM) suffers from a significant overfitting problem.

Purpose of the Study:

• Investigate the theoretical causes of kGPRELM overfitting.
• Propose a novel correlation-based GPRELM (cGPRELM) to mitigate overfitting.
• Enhance the generalization performance and computational efficiency of ELM-based models.

Main Methods:

• Utilized correlation coefficients to measure hidden-layer output vector similarity.
• Developed cGPRELM to prevent covariance matrix from becoming an identity matrix.
• Evaluated cGPRELM on real-world classification and regression datasets.

Main Results:

• cGPRELM effectively controls overfitting by managing covariance matrix behavior.
• cGPRELM demonstrates superior performance compared to ELM and kGPRELM, especially in challenging initialization scenarios.
• Experimental results confirm cGPRELM's improved generalization and reduced computational complexity.

Conclusions:

• The proposed cGPRELM offers a robust solution to overfitting in GPRELM.
• cGPRELM provides better prediction accuracy and efficiency than existing methods.
• This approach enhances the practical applicability of ELM and GPR integration.