CS-QCFS: Bridging the performance gap in ultra-low latency spiking neural networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Deep Learning

Background:

• Spiking Neural Networks (SNNs) emulate biological systems for energy-efficient and plausible AI.
• Converting Artificial Neural Networks (ANNs) to SNNs is promising but often requires long time-steps for comparable performance.
• Existing conversion methods face challenges with channel heterogeneity and negative thresholds, hindering efficiency.

Purpose of the Study:

• To investigate critical issues in ANN-to-SNN conversion that lead to the requirement of long time-steps.
• To introduce a novel activation function that addresses channel disparities and maintains positive thresholds.
• To enable high-performance SNNs with significantly reduced time-steps.

Main Methods:

• Thorough investigation of the ANN-to-SNN conversion process, identifying channel heterogeneity and negative thresholds as key problems.
• Introduction of the Channel-wise Softplus Quantization Clip-Floor-Shift (CS-QCFS) activation function.
• Experimental validation on CIFAR datasets using the proposed CS-QCFS activation function.

Main Results:

• The CS-QCFS activation function effectively handles inter-channel disparities and ensures positive thresholds.
• The proposed method achieves state-of-the-art performance on CIFAR-10 and CIFAR-100 datasets.
• Achieved 95.86% top-1 accuracy on CIFAR-10 and 74.83% on CIFAR-100 with only 1 time-step.

Conclusions:

• The novel CS-QCFS activation function successfully resolves critical issues in ANN-to-SNN conversion.
• This approach enables the creation of highly efficient and accurate SNNs with ultra-low time-steps.
• The method sets a new benchmark for SNN performance, particularly in resource-constrained applications.