Coincidence detection and integration behavior in spiking neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) are increasingly studied for their potential energy efficiency and dynamic adaptability, mimicking biological neural networks.
• The leaky-integrate-and-fire (LIF) neuron is a common model in SNNs, exhibiting dual operation modes: coincidence detection and integration, depending on membrane decay times.
• The precise emergence and impact of these LIF neuron operation modes on SNN performance remain unclear.

Purpose of the Study:

• To investigate the effect of varying decay times on LIF neuron operation modes within SNNs.
• To propose and validate quantitative measures for characterizing these operation modes.
• To explore the relationship between decay time and the identified operation modes.

Main Methods:

• Trained SNNs using a surrogate-gradient-based approach with varying LIF neuron decay times.
• Introduced two novel measures: number of contributing input spikes and effective integration interval, to quantify neuron operation modes.
• Analyzed the correlation between these measures and the neuron's decay time.

Main Results:

• Coincidence detection mode is characterized by fewer input spikes and shorter integration intervals.
• Integration mode is associated with numerous input spikes and extended integration intervals.
• A linear correlation was found between the two measures, with a correlation factor exhibiting a power-law dependency on decay time.

Conclusions:

• The identified measures effectively distinguish between coincidence detection and integration modes in LIF neurons.
• The power-law relationship suggests an intrinsic property of LIF networks related to decay time.
• This research provides a foundation for optimizing SNNs for enhanced efficiency and biological realism by understanding and controlling neuron operation modes.