A universal ANN-to-SNN framework for achieving high accuracy and low latency deep Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer biological plausibility, low power consumption, and neuromorphic hardware potential.
• Converting Artificial Neural Networks (ANNs) to SNNs is a cost-effective method for SNN development.
• Existing ANN-to-SNN conversion methods often suffer from errors and lack generality, requiring modifications to the original ANNs.

Purpose of the Study:

• To present a universal and effective framework for ANN-to-SNN conversion that minimizes errors.
• To address conversion inaccuracies caused by discreteness and asynchrony in network transmission.
• To improve the accuracy and performance of converted SNNs without altering the source ANNs.

Main Methods:

• Developed the DNISNM framework comprising Data-based Neuronal Initialization (DNI) and Signed Neuron with Memory (SNM).
• DNI mechanism tackles errors from discreteness disparities.
• SNM mechanism addresses errors from asynchrony disparities.

Main Results:

• The DNISNM framework successfully converts ANNs to SNNs with improved accuracy.
• Achieved high precision and low inference latency in converted SNNs.
• Demonstrated effectiveness on challenging datasets like CIFAR10, CIFAR100, and ImageNet-1k.

Conclusions:

• The DNISNM framework offers a universal solution for ANN-to-SNN conversion, overcoming limitations of previous methods.
• The framework enhances SNN accuracy and performance without modifying the original ANN architecture.
• DNISNM enables high-precision, low-latency SNNs suitable for demanding applications like object recognition.