A Noise-Based Novel Strategy for Faster SNN Training.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer low power consumption and bioplausibility but face optimization challenges.
• Existing methods like ANN-to-SNN conversion and spike-based backpropagation (BP) have limitations in inference time and computational cost, respectively.

Purpose of the Study:

• To propose a novel and efficient SNN training approach.
• To reduce the computational resources and time required for SNN training and inference.
• To enhance the bioplausibility of SNN neuron models.

Main Methods:

• A single-step SNN(T = 1) is trained by approximating neural potential distribution with Gaussian noise.
• The trained single-step SNN is losslessly converted to a multistep SNN(T = N).

Main Results:

• The proposed method significantly reduces SNN training time by 65%–75%.
• Inference speed is improved by over 100 times compared to existing methods.
• High accuracy is maintained after conversion, with noise enhancing performance.

Conclusions:

• This novel training approach effectively addresses the efficiency limitations of SNNs.
• The noise-augmented neuron model demonstrates increased bioplausibility.
• The method offers a promising direction for practical SNN applications.