A pruning feedforward small-world neural network based on Katz centrality for nonlinear system modeling.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Small-world neural networks enhance artificial neural network generalization but often feature large, predefined architectures.
• Large architectures can lead to overfitting and prolonged training times, hindering optimal structure discovery.

Purpose of the Study:

• To propose a pruning feedforward small-world neural network (PFSWNN) for efficient nonlinear system modeling.
• To address limitations of traditional small-world networks, including overfitting and computational cost.

Main Methods:

• Constructed a feedforward small-world neural network (FSWNN) using the Watts-Strogatz rewiring rule.
• Implemented a pruning strategy based on Katz centrality to identify and merge unimportant hidden neurons.
• Trained connection weights via gradient-based algorithms and theoretically analyzed PFSWNN convergence.

Main Results:

• PFSWNN demonstrated superior generalization performance on diverse nonlinear system modeling tasks.
• The pruning algorithm resulted in a compact network structure, significantly shortening training time.
• Effectiveness validated on function approximation, time-series prediction, benchmark datasets, and wastewater treatment processes.

Conclusions:

• The proposed PFSWNN effectively combines small-world properties with a pruning mechanism for enhanced performance.
• PFSWNN offers an automated approach to discovering optimal network structures, overcoming limitations of predefined architectures.
• This method provides a computationally efficient and high-performing solution for complex nonlinear system modeling.