On the accuracy and computational cost of spiking neuron implementation

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks
• Biophysics

Background:

• Widely accepted theories on spiking neuron (SN) model efficiency (Hodgkin-Huxley, Izhikevich, Leaky Integrate-and-Fire) are debated.
• Previous consensus on computational cost and accuracy of SN models like HH, IZH, and LIF is challenged by recent findings.
• Lack of consensus necessitates a refined approach to evaluate SN simulation capacities.

Purpose of the Study:

• To introduce a refined approach using multiobjective optimization theory to describe SN simulation capacities.
• To identify optimal simulation parameters balancing computational cost and accuracy.
• To challenge and update existing theories on SN model efficiency and computational cost.

Main Methods:

• Utilized multiobjective optimization theory to analyze SN simulation capacities.
• Employed normalized metrics for accuracy, computational cost, and efficiency for cross-model comparisons.
• Conducted tests under various boundary conditions and current stimuli (constant and random) for regular spiking modes.

Main Results:

• The Hodgkin-Huxley (HH) model is the most accurate, computationally inexpensive, and efficient.
• Izhikevich (IZH) and Leaky Integrate-and-Fire (LIF) models are found to be the most inaccurate and expensive.
• Computational cost is better measured by a proposed metric than FLOPS, and optimal parameters are problem-specific.

Conclusions:

• Established theories on SN simulation capacities are refuted, with HH model showing superior performance.
• The computational cost and accuracy of SN models are more nuanced than previously understood, influenced by factors like table discretization and spike discontinuity.
• Optimal simulation parameter selection is a problem-specific, multiobjective optimization challenge, not a one-size-fits-all solution.