Post-training quantization for efficient ANN-SNN conversion.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) mimic biological neurons for efficient computing.
• Current SNN training involves direct optimization or ANN-to-SNN conversion.
• ANN-to-SNN conversion often suffers from significant transformation errors.

Purpose of the Study:

• Investigate and mitigate conversion errors in ANN-to-SNN transformation.
• Propose a novel method for training deep SNNs.
• Enhance the accuracy and efficiency of SNNs.

Main Methods:

• Theoretical analysis of conversion errors.
• Implementation of channel-wise thresholds versus layer-wise thresholds.
• Application of post-training quantization (PTQ) for efficient calibration without retraining.

Main Results:

• Channel-wise thresholds are theoretically more effective than layer-wise thresholds in reducing conversion error.
• Post-training quantization (PTQ) enables efficient calibration of SNNs.
• The proposed method significantly reduces training time compared to direct training and conventional ANN-to-SNN conversion.
• Improved accuracy on both static image and neuromorphic datasets.

Conclusions:

• Channel-wise thresholds and PTQ offer an effective strategy for accurate and efficient ANN-to-SNN conversion.
• This approach advances the practical application of SNNs for next-generation computing.
• The method presents a viable alternative to direct SNN training, offering reduced computational cost.