Impact of Neuron Models on Spiking Neural Network Performance: A Complexity-based Classification Approach.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Bio-signal Processing

Background:

• Spiking Neural Networks (SNNs) show promise for bio-signal processing due to their temporal processing capabilities.
• The performance of SNNs is highly dependent on the choice of neuron models and learning rules.
• A standardized evaluation framework for SNNs, especially for bio-signal classification, is lacking.

Purpose of the Study:

• To systematically investigate how different neuron models and learning rules impact SNN classification performance in bio-signal processing.
• To introduce and validate a novel complexity-based evaluation metric, Lempel-Ziv Complexity (LZC), for SNNs.
• To provide guidelines for selecting optimal SNN configurations for diverse neural data complexities.

Main Methods:

• Systematic comparison of Leaky Integrate-and-Fire, metaneurons, and Levy-Baxter (LB) neurons.
• Evaluation across spike-timing dependent plasticity, tempotron, and reward-modulated learning rules.
• Integration of Lempel-Ziv Complexity (LZC) for evaluating spike-train regularity and classification performance on synthetic and real (MNIST) datasets.

Main Results:

• SNN performance is strongly influenced by the interplay between neuron model, learning rule, and network size.
• The Levy-Baxter neuron with tempotron learning demonstrated superior performance for complex temporal patterns.
• Leaky Integrate-and-Fire neurons with Bio-inspired Active Learning provided efficient classification with lower computational cost.

Conclusions:

• The study establishes a systematic mapping of neuron model-learning rule synergies in SNNs for bio-signal classification.
• LZC offers a robust and interpretable benchmark for evaluating SNNs, particularly in noisy or weak signal conditions.
• Actionable guidelines are provided for designing next-generation SNNs capable of handling complex and variable neural data.