An Interclass Margin Maximization Learning Algorithm for Evolving Spiking Neural Network.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) offer a biologically plausible model for information processing.
• Pattern classification remains a challenging task for existing SNN learning algorithms.
• Maximizing interclass margins is a proven strategy for improving classifier generalization.

Purpose of the Study:

• To introduce a novel learning algorithm, the two stage margin maximization spiking neural network (TMM-SNN), for pattern classification.
• To enhance the generalization performance of SNNs through margin maximization.
• To develop a new learning rule, the normalized membrane potential learning rule, for SNNs.

Main Methods:

• The TMM-SNN algorithm features a two-stage learning process: structure learning and output weights learning.
• The structure learning stage evolves hidden layer neurons and updates their weights using a normalized membrane potential learning rule.
• The output weights learning stage maximizes interclass margins based on output neuron responses.

Main Results:

• TMM-SNN demonstrated superior generalization performance across ten benchmark datasets from the UCI machine learning repository.
• Statistical analysis using the Friedman test and Fisher's LSD confirmed TMM-SNN's improved performance compared to existing SNN learning algorithms.
• The normalized membrane potential learning rule effectively utilizes local spike train information and existing synaptic weights.

Conclusions:

• The TMM-SNN algorithm represents a significant advancement in SNN-based pattern classification.
• The proposed normalized membrane potential learning rule contributes to enhanced interclass margin maximization.
• TMM-SNN offers a robust and effective solution for pattern classification tasks using spiking neural networks.