Precise spiking neurons for fitting any activation function in ANN-to-SNN Conversion

Abstract

Spiking Neural Networks (SNNs) are recognized for their energy efficiency due to spike-based communication. In this regard, the shift towards SNNs is driven by their ability to significantly reduce energy consumption while maintaining the performance of ANNs. Converting Artificial Neural Networks (ANNs) to SNNs is a key research focus, but existing methods often struggle with balancing conversion accuracy and latency, and are typically restricted to ReLU activations. We introduce Precision Spiking (PS) neurons, a novel dynamic spiking neuron model that can precisely fit any activation function by jointly regulating spike timing, reset voltage, and membrane potential threshold. This capability enables exact parameter optimization via iterative methods, achieving low-latency, high-accuracy ANN-to-SNN conversion. Experiments on image classification and natural language processing benchmarks confirm state-of-the-art results, with a maximum conversion loss of 0.55% and up to 0.38% accuracy improvement over the original ANN. To the best of our knowledge, this method offers a significant advancement over existing approaches by achieving high-precision fitting of arbitrary activation functions with low latency and minimal conversion loss, thus considerably expanding the range of feasible ANN-to-SNN conversions.