Operational manifolds in spiking neural networks

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) offer potential energy savings over traditional deep networks.
• Performance and stability of SNNs are sensitive to neuron hyperparameters and inference strategies.
• Leaky Integrate-and-Fire (LIF) models are common SNN neuron types.

Purpose of the Study:

• To investigate the interplay between LIF neuron hyperparameters and inference policies.
• To define and map an 'operational manifold' for balanced SNN activity and performance.
• To identify accuracy-energy trade-offs and assess robustness under varying conditions.

Main Methods:

• Systematic grid sweeps of membrane time constant (τm) and firing threshold (Vth) to map the operational manifold.
• Synaptic Operation (SOP) cost estimation for quantifying inference energy efficiency.
• Comparison of 'reset' and 'carry' membrane potential policies for state handling.
• Analysis of robustness via input perturbations and examination of spike-train correlations.

Main Results:

• The operational manifold represents a stable operating region for SNNs, balancing activity and performance.
• Accuracy-energy frontiers were identified within the manifold using composite scores.
• 'Reset' policy improved accuracy on static data, while 'carry' introduced interference in streaming data.
• Leaving the operational manifold correlated with increased spike-train synchrony and correlations.

Conclusions:

• Practical guidelines for selecting SNN hyperparameters and inference policies for energy efficiency and stability.
• Correlation statistics of spike trains serve as effective, label-free indicators of SNN health and noise exposure.
• The findings support the development of robust and efficient neuromorphic systems.