Evolutionary Spiking Neural Networks for Solving Supervised Classification Problems.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer a biologically plausible model for computation.
• Designing effective SNNs for supervised classification often requires complex training procedures.

Purpose of the Study:

• To develop an automated methodology for designing third-generation artificial neural networks (ANNs), specifically SNNs.
• To enable SNN design for supervised classification without explicit training by optimizing network topology and synaptic connections.

Main Methods:

• A grammatical evolution (GE)-based approach was employed to explore the design space of three-layered feedforward SNNs.
• The methodology focused on configuring synaptic weights and delays, incorporating partial connections between input and hidden layers.
• Fitness functions were refined to enhance the design process and improve SNN performance.

Main Results:

• The proposed GE-based methodology automatically designed SNNs with competitive and consistent results on benchmark datasets.
• Statistical comparisons indicated that the designed SNNs outperformed those generated by existing methodologies for both second-generation ANNs and adapted SNN approaches.
• Partial connections in the designed SNNs potentially reduce redundancy and input feature dimensionality.

Conclusions:

• The GE-based methodology provides an effective, automated approach for designing high-performing SNNs for supervised classification.
• The findings suggest that automated design, leveraging GE and optimized fitness functions, can yield superior results compared to traditional methods.
• This work contributes to advancing the practical application of SNNs in machine learning tasks.