Quantization Framework for Fast Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer temporal dynamics but have lower information transmission rates than artificial neural networks (ANNs).
• ANN-to-SNN conversion links ANN activation bit precision to SNN representation time, suggesting quantization can reduce SNN inference latency.
• Directly applying ANN quantization techniques to SNNs is challenging due to fundamental differences between the network types.

Purpose of the Study:

• To propose a novel quantization framework for fast SNNs (QFFS) to enhance SNN latency and minimize accuracy loss during ANN-to-SNN conversion.
• To ensure compatibility between ANN quantization methods and SNNs while suppressing noise to preserve accuracy.
• To demonstrate that QFFS can overcome accuracy degradation in SNNs with limited time steps.

Main Methods:

• Developed a quantization framework for fast SNNs (QFFS) to bridge ANN quantization and SNN conversion.
• Implemented techniques to promote compatibility of ANN quantization with SNNs.
• Introduced methods to suppress "occasional noise" to minimize accuracy loss.

Main Results:

• Achieved an accuracy of 70.18% on ImageNet using SNNs within only 8 time steps.
• Overcame the typical accuracy degeneration observed in SNNs with a limited number of time steps.
• Demonstrated that SNNs built via ANN-to-SNN conversion can achieve latency comparable to directly trained SNNs.

Conclusions:

• The QFFS framework effectively enables the creation of low-latency, high-accuracy SNNs through ANN conversion.
• This work represents the first successful demonstration of achieving competitive latency with directly trained SNNs using ANN-to-SNN conversion.
• QFFS offers a viable method for optimizing SNN performance by leveraging ANN quantization strategies.