High-accuracy deep ANN-to-SNN conversion using quantization-aware training framework and calcium-gated bipolar leaky...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Deep Learning

Background:

• Spiking Neural Networks (SNNs) offer efficient event-driven computing.
• Artificial Neural Network (ANN)-to-SNN conversion is a leading method for SNN accuracy.
• Existing ANN-to-SNN methods require complex post-conversion steps and long inference times.

Purpose of the Study:

• To develop an improved ANN-to-SNN conversion technique.
• To reduce the discrepancy between ANN and SNN neuron behaviors.
• To minimize inference latency in converted SNNs.

Main Methods:

• Proposed a calcium-gated bipolar leaky integrate and fire (Ca-LIF) spiking neuron model.
• Developed a quantization-aware training (QAT)-based framework for direct ANN-to-SNN weight export.
• Utilized an off-the-shelf QAT toolkit for streamlined conversion.

Main Results:

• The Ca-LIF model effectively approximates ReLU neuron functions.
• The QAT framework enabled direct ANN-to-SNN conversion without post-processing.
• Converted SNNs achieved competitive accuracy with significantly shorter inference time steps across various network structures.

Conclusions:

• The proposed Ca-LIF neuron model and QAT framework offer an efficient and effective ANN-to-SNN conversion method.
• This approach overcomes limitations of traditional ANN-to-SNN techniques, reducing complexity and inference latency.
• The findings pave the way for more practical and performant SNN applications.