Fast-SNN: Fast Spiking Neural Network by Converting Quantized ANN.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer computational and energy efficiency advantages over artificial neural networks (ANNs) due to event-driven processing and simpler operations.
• Training deep SNNs is challenging due to their discrete spiking nature, often necessitating ANN-to-SNN conversion.
• ANN-to-SNN conversion typically suffers from quantization and accumulating errors, leading to high inference latency and reduced performance.

Purpose of the Study:

• To propose Fast-SNN, a novel method for training deep SNNs that achieves high performance with significantly reduced inference latency.
• To address the limitations of traditional ANN-to-SNN conversion methods by minimizing quantization and sequential errors.

Main Methods:

• Establishing an equivalent mapping between temporal quantization in SNNs and spatial quantization in ANNs.
• Transferring the minimization of quantization error to the training phase of quantized ANNs.
• Identifying sequential error as the primary cause of accumulating error and introducing a signed Integrate-and-Fire (IF) neuron model and layer-wise fine-tuning to mitigate it.

Main Results:

• Fast-SNN achieves state-of-the-art performance on various computer vision tasks, including image classification, object detection, and semantic segmentation.
• The proposed method significantly reduces inference latency compared to existing ANN-to-SNN conversion techniques while maintaining high accuracy.
• Demonstrated effective minimization of quantization error through optimized ANN training.

Conclusions:

• Fast-SNN offers a viable and efficient approach for deploying deep SNNs in practical applications.
• The method successfully balances high performance with low latency, realizing the inherent advantages of SNNs.
• The findings pave the way for more efficient and powerful event-driven neural network architectures.