A complex-valued widening spiking neural network.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Traditional spiking neural networks (SNNs) primarily utilize spike timing, neglecting crucial subthreshold dynamics.
• This limitation restricts the representational capacity and information processing capabilities of SNNs.

Purpose of the Study:

• To introduce a novel neural network architecture, the Complex-valued Widening Spiking Neural Network (CWSNN).
• To enhance SNNs by integrating subthreshold membrane potential dynamics alongside spike timing information.
• To improve network capacity and performance without increasing network depth.

Main Methods:

• Developed CWSNN by encoding temporal information in the real part and spatial information (subthreshold membrane potential derivative) in the imaginary part of complex neurons.
• Evaluated CWSNN performance on seven tabular and four image datasets.
• Compared CWSNN against existing SNN models in terms of accuracy, convergence speed, and generalization.

Main Results:

• CWSNN demonstrated superior accuracy and faster convergence across all tested datasets compared to conventional SNNs.
• The proposed model exhibited robust generalization capabilities.
• CWSNN achieved competitive performance on regression tasks, an area where traditional SNNs often struggle.

Conclusions:

• CWSNN effectively integrates spatiotemporal features, significantly widening network capacity.
• The novel architecture overcomes limitations of traditional SNNs, offering improved performance and broader applicability.
• CWSNN presents a promising advancement for SNNs, particularly for complex tasks including regression.