Biologically-informed excitatory and inhibitory ratio for robust spiking neural network training

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neuromorphic engineering

Background:

• Spiking neural networks (SNNs) offer energy-efficient AI by mimicking biological brains.
• Training SNNs, particularly with biological constraints on excitatory and inhibitory connections, presents significant challenges.
• Robust training principles for SNNs are crucial for their practical application.

Purpose of the Study:

• To identify key factors influencing the trainability of SNNs with varying excitatory-inhibitory (E:I) neuron ratios.
• To investigate the impact of biological constraints, such as firing rates and inhibitory spiking patterns, on SNN training.
• To assess the robustness of trained SNNs in noisy environments and their reliance on inhibitory neuron function.

Main Methods:

• Simulated spiking neural networks with diverse E:I ratios.
• Analysis of training dynamics under varying initial firing rates and inhibitory spiking patterns.
• Utilized Van Rossum distance to quantify spike train synchrony and network robustness.

Main Results:

• Low initial firing rates and diverse inhibitory spiking patterns are critical for successful SNN training.
• Biologically realistic E:I ratios enable reliable training even at low activity levels and in noisy conditions.
• Inhibitory neurons significantly enhance network robustness to noise, as indicated by Van Rossum distance analysis.

Conclusions:

• Identified key training principles for biologically constrained SNNs.
• Demonstrated that realistic E:I ratios and specific inhibitory dynamics improve training reliability and robustness.
• Findings support the development of biologically-informed large-scale SNNs and energy-efficient hardware.