Dynamic adjustment of hidden node parameters for extreme learning machine.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• Extreme Learning Machines (ELM) are effective for function approximation with single hidden layers.
• ELMs can contain redundant or insignificant hidden nodes, impacting performance.
• Existing ELM variants do not optimally address insignificant node contributions.

Purpose of the Study:

• To introduce a Dynamic Adjustment Extreme Learning Machine (DA-ELM) for enhanced performance.
• To reduce residual errors by tuning input parameters of insignificant hidden nodes.
• To provide a theoretical foundation for the DA-ELM approach.

Main Methods:

• Proposing the DA-ELM algorithm with dynamic adjustment of hidden node parameters.
• Utilizing the recursive expectation-maximization theorem to minimize energy error.
• Updating insignificant hidden node parameters iteratively based on error reduction.

Main Results:

• DA-ELM effectively reduces energy error.
• Experimental results demonstrate superior efficiency compared to state-of-the-art algorithms.
• DA-ELM outperforms Bayesian ELM, optimally-pruned ELM, and other methods.

Conclusions:

• DA-ELM offers a significant improvement over standard ELM and other advanced techniques.
• The dynamic adjustment mechanism effectively handles insignificant hidden nodes.
• DA-ELM presents a more efficient and accurate solution for function approximation problems.