Sparse extreme learning machine for classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Extreme Learning Machine (ELM) was originally designed for single-hidden-layer feedforward neural networks (SLFNs).
• Unified ELM offers a broad framework but results in dense solutions, demanding significant storage and testing time for large datasets.
• Traditional unified ELM methods involve matrix inversion, leading to high computational complexity (quadratic to cubic) for large-scale training.

Purpose of the Study:

• To introduce a sparse Extreme Learning Machine (ELM) as an efficient alternative for classification.
• To reduce storage space and testing time requirements in large-scale machine learning applications.
• To develop an efficient training algorithm for the proposed sparse ELM.

Main Methods:

• Proposed a sparse ELM model to address the limitations of dense solutions in unified ELM.
• Developed an efficient training algorithm that decomposes the quadratic programming problem into smaller, analytically solvable sub-problems.
• Compared the performance of sparse ELM against Support Vector Machines (SVM) and unified ELM.

Main Results:

• Sparse ELM demonstrates superior generalization performance and significantly faster training speeds compared to SVM.
• For binary classification, sparse ELM achieves comparable generalization performance to unified ELM.
• Sparse ELM exhibits substantially faster training speeds than unified ELM, particularly for large-scale binary classification problems.

Conclusions:

• Sparse ELM provides an effective and efficient approach for large-scale classification tasks.
• The proposed efficient training algorithm enhances the practicality of sparse ELM for big data.
• Sparse ELM offers a compelling alternative to existing methods like SVM and unified ELM, balancing performance and computational efficiency.