Rethinking Pretraining as a Bridge From ANNs to SNNs

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer brain-inspired computation with low power consumption but face challenges in achieving high accuracy efficiently.
• Current SNN training methods include converting Artificial Neural Networks (ANNs) or direct SNN training, each with limitations like long inference times or high computational cost.
• The efficiency and accuracy of SNNs are critical for their adoption in various applications.

Purpose of the Study:

• To propose a new, efficient training paradigm for Spiking Neural Networks (SNNs).
• To address the trade-offs between inference time and training efficiency in current SNN training methodologies.
• To demonstrate the effectiveness of the proposed pipeline on large-scale datasets.

Main Methods:

• A novel SNN training paradigm is introduced, integrating pretraining techniques with a backpropagation (BP)-based deep SNN training mechanism.
• The proposed pipeline, comprising pipe-S for static and pipe-D for dynamic data transfer, combines ANN-to-SNN conversion and direct SNN training concepts.
• The method was evaluated on benchmark datasets including ImageNet-1K and the large-scale event-driven ES-ImageNet and ES-UCF101 datasets.

Main Results:

• The proposed SNN training pipeline achieved state-of-the-art (SOTA) results on the ES-ImageNet dataset.
• Significant training acceleration was observed, achieving comparable or higher accuracy than existing leaky-integrate-and-fire (LIF)-SNNs with substantially reduced training times (1/8 on ImageNet-1K, 1/2 on ES-ImageNet).
• A time-accuracy benchmark was established for the new ES-UCF101 dataset, showcasing the pipeline's performance on dynamic data transfer tasks.

Conclusions:

• The developed SNN training paradigm offers a more efficient and effective approach compared to existing methods.
• Experimental results suggest a strong similarity between the functional parameters of ANNs and SNNs, validating the hybrid training approach.
• The proposed pipeline demonstrates broad potential for various applications requiring efficient and accurate SNNs.