Graph Embedded Extreme Learning Machine.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Neural Networks

Background:

• Extreme Learning Machine (ELM) is a popular algorithm for training single-hidden layer feedforward neural networks.
• Incorporating subspace learning (SL) criteria can potentially improve ELM's performance by optimizing output weight calculations.
• Existing methods may not fully exploit the potential of SL within the ELM framework, especially in high-dimensional spaces.

Purpose of the Study:

• To propose a novel extension of the ELM algorithm, named Graph Embedded Extreme Learning Machine (GEELM).
• To enable GEELM to incorporate both intrinsic and penalty subspace learning criteria within the graph embedding framework.
• To extend GEELM for application in arbitrary (even infinite) dimensional ELM spaces.

Main Methods:

• Developed the Graph Embedded Extreme Learning Machine (GEELM) algorithm.
• Integrated subspace learning (SL) criteria, including intrinsic and penalty types, into the ELM output weight optimization.
• Extended the GEELM framework to handle arbitrary dimensional ELM spaces.
• Evaluated GEELM on standard classification problems and human behavior analysis datasets (face, expression, activity recognition).

Main Results:

• GEELM effectively incorporates subspace learning criteria into the ELM training process.
• The algorithm demonstrated superior performance compared to other ELM-based classification schemes across all tested datasets.
• Experimental results confirmed the effectiveness of GEELM in classification tasks, including human behavior analysis.

Conclusions:

• The proposed GEELM algorithm offers a significant advancement in ELM-based neural network training.
• GEELM's ability to integrate subspace learning criteria enhances its classification capabilities.
• The approach shows promise for complex pattern recognition tasks, particularly in human behavior analysis.