Dynamic extreme learning machine and its approximation capability.

Area of Science:

• Machine Learning
• Artificial Neural Networks

Background:

• Extreme learning machines (ELMs) are efficient feedforward networks for regression and classification.
• Determining optimal ELM architecture is critical for successful application.

Purpose of the Study:

• To propose a dynamic ELM (D-ELM) that self-adapts its architecture.
• To theoretically prove the approximation capabilities of D-ELM.
• To validate D-ELM's performance in reducing network size and preserving generalization.

Main Methods:

• Developed a D-ELM model with dynamic recruitment/deletion of hidden nodes based on performance significance.
• Provided theoretical proof of D-ELM's ability to approximate Lebesgue p-integrable functions using Lebesgue p-integrable hidden activation functions on compact input sets.
• Conducted simulations on various test problems to evaluate D-ELM performance.

Main Results:

• The D-ELM dynamically adjusts network architecture by adding or removing hidden nodes.
• Theoretical analysis confirms D-ELM's function approximation capabilities.
• Empirical results demonstrate significant reduction in network size without compromising generalization performance.

Conclusions:

• The proposed D-ELM offers an effective approach for self-adaptive network architecture optimization.
• D-ELM successfully balances network size reduction with robust generalization.
• This method enhances the practical applicability of ELMs in diverse machine learning tasks.