Efficient Multispike Learning for Spiking Neural Networks Using Probability-Modulated Timing Method.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Traditional Spiking Neural Networks (SNNs) face challenges in multispike learning due to the discontinuous nature of spiking neurons.
• Existing error functions rely on spike timing, leading to information loss when output and target spike counts differ.
• Current methods often neglect spikes to match counts, hindering effective learning in SNNs.

Purpose of the Study:

• To develop a novel learning mechanism for Spiking Neural Networks (SNNs) that addresses information loss in multispike scenarios.
• To introduce a probability-modulated approach for converting discontinuous spike patterns into likelihoods of desired output spike trains.
• To construct a Probability-Modulated Spiking Neural Network (PMSNN) for improved supervised learning.

Main Methods:

• Developed a probability-modulated timing mechanism utilizing stochastic neurons to represent spike patterns as generation likelihoods.
• Constructed a Probability-Modulated Spiking Neural Network (PMSNN) by integrating this mechanism into a spiking classifier.
• Implemented a multilayer, multispike learning structure with a clustering rule connection mechanism in the reservoir for efficient synaptic transmission.

Main Results:

• The proposed PMSNN effectively maps more inputs to target spike trains within its multilayer structure.
• The clustering rule connection mechanism enhances information transmission efficiency by mapping correlated inputs to adjacent neurons.
• Comparisons show the PMSNN achieves higher efficiency and requires fewer parameters than popular SNN algorithms.

Conclusions:

• The probability-modulated timing mechanism provides a robust solution for multispike learning in SNNs, preserving information.
• The PMSNN demonstrates superior performance in terms of efficiency and parameter reduction compared to existing SNN methods.
• This approach offers a promising advancement for developing more effective and efficient spiking neural network models.