A Fast Reduced Kernel Extreme Learning Machine.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Support Vector Machines (SVM) and Least Square SVM (LS-SVM) require iterative identification of support or weight vectors.
• Iterative methods can be computationally expensive, particularly for large datasets.
• Efficient algorithms are needed for complex pattern recognition and regression tasks.

Purpose of the Study:

• To introduce a novel, fast, and accurate kernel-based supervised learning algorithm named Reduced Kernel Extreme Learning Machine (RKELM).
• To demonstrate RKELM's ability to approximate nonlinear functions accurately.
• To evaluate RKELM's performance against established methods like SVM and LS-SVM.

Main Methods:

• RKELM employs a non-iterative approach by randomly selecting a subset of data samples as support vectors.
• The algorithm is theoretically grounded in the universal learning capabilities of reduced kernel-based SLFNs.
• Rigorous proofs establish the accuracy of function approximation with sufficient support vectors.

Main Results:

• RKELM achieves significant computational cost savings compared to iterative SVM/LS-SVM methods, especially on large datasets.
• Experimental results show RKELM performs competitively across binary classification, multi-class problems, and regression tasks.
• The algorithm demonstrates accurate nonlinear function approximation under specific conditions.

Conclusions:

• RKELM provides a computationally efficient alternative to traditional SVM/LS-SVM algorithms.
• The method offers a favorable trade-off between accuracy and computational effort for various machine learning applications.
• RKELM is suitable for both small and large-scale real-world datasets.