M-estimation activation functions for high-performance extreme learning machine ensemble classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Machine learning, particularly AI in Software-Defined Networking, is crucial for cybersecurity tasks like traffic monitoring and anomaly detection.
• Existing ensemble methods often struggle with noisy or contaminated data, limiting their effectiveness in real-world security scenarios.

Purpose of the Study:

• To develop a robust ensemble framework for Extreme Learning Machines (ELMs) that is resilient to data irregularities.
• To improve the generalization, predictive precision, and stability of neural classifiers.

Main Methods:

• Proposed a novel ensemble framework for ELMs incorporating redescending ψ-activation functions based on M-estimation theory.
• Utilized grid search to determine the optimal hidden-node count by minimizing the Brier score.
• Combined ensemble outputs using least-squares optimization instead of traditional voting for precise parameter estimation.

Main Results:

• The proposed method demonstrated consistently superior accuracy and reduced variance across five benchmark datasets compared to existing ELM ensembles.
• Validated performance gains through rigorous statistical testing, including Kruskal-Wallis and Dunn's post-hoc analyses.
• The framework showed marked improvements in generalization, predictive precision, and resilience to data irregularities.

Conclusions:

• Embedding robust M-estimator-based activations within a controlled ensemble significantly enhances ELM performance.
• The developed framework offers a substantial advancement in designing efficient and resilient neural classifiers for machine learning applications.