Stacked Extreme Learning Machines.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Extreme Learning Machine (ELM) offers rapid training, good generalization, and implementation ease for various classification and regression tasks.
• Existing ELM methods face challenges with large and complex datasets, demanding significant computational resources.

Purpose of the Study:

• To introduce Stacked Extreme Learning Machines (S-ELMs) designed for efficient processing of large-scale, complex data.
• To enhance the performance and memory efficiency of ELM for big data applications.

Main Methods:

• Developed S-ELMs by segmenting a large ELM into serially connected, smaller ELM modules.
• Integrated ELM autoencoders within the S-ELMs iterative process to boost testing accuracy on big data.
• Evaluated S-ELMs performance against Support Vector Machine (SVM) and Deep Belief Network (DBN).

Main Results:

• S-ELMs, even with random hidden nodes, achieved testing accuracy comparable to SVM while significantly reducing memory requirements.
• The inclusion of ELM autoencoders in S-ELMs led to substantially improved testing accuracy over SVM.
• S-ELMs demonstrated superior training speed compared to SVM and DBN, with accuracy slightly exceeding DBN.

Conclusions:

• S-ELMs provide a scalable and memory-efficient solution for tackling large and complex machine learning problems.
• The proposed S-ELMs framework, enhanced with ELM autoencoders, offers a compelling alternative to traditional methods like SVM and DBN for big data analysis.