An efficient automated parameter tuning framework for spiking neural networks.

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Artificial Neural Networks

Background:

• Biologically realistic spiking neural networks (SNNs) are increasingly desired for modeling complex neural circuits and phenomena.
• The complexity of SNNs, including realistic plasticity and dynamics, makes parameter tuning a significant challenge.
• SNNs are well-suited for neuromorphic hardware, enabling large-scale brain-inspired architectures.

Purpose of the Study:

• To develop an automated parameter tuning framework for accelerating the development of large-scale SNNs.
• To efficiently tune SNNs using evolutionary algorithms (EA) and graphics processing units (GPUs).
• To demonstrate the framework's capability in achieving specific neural functions, such as V1 simple cell-like responses.

Main Methods:

• Implemented an automated parameter tuning framework utilizing evolutionary algorithms (EA).
• Leveraged graphics processing units (GPUs) for parallel computation to accelerate the tuning process.
• Evaluated the framework on a sample SNN of 4104 neurons, aiming for V1 simple cell-like tuning curves and self-organizing receptive fields (SORFs).

Main Results:

• The GPU-accelerated framework achieved a 65x speedup compared to a single-threaded CPU implementation (0.35 h/generation vs. 23.5 h/generation).
• The tuned SNN successfully produced V1 simple cell-like tuning curve responses and self-organizing receptive fields (SORFs).
• Parameter solutions derived from the tuning process were found to be stable and repeatable.

Conclusions:

• The automated parameter tuning framework significantly enhances the speed and efficiency of constructing and tuning large-scale SNNs.
• This approach facilitates the exploration of complex SNN models for computational neuroscience and neuromorphic engineering applications.
• The framework makes building and tuning biologically realistic SNNs more accessible and practical.