Benchmarking ANN-to-SNN Conversion: Dataset-Dependent Analysis of Accuracy, Latency, and Spike Efficiency

Spiking Neural Networks (SNNs) offer event-driven, energy-efficient computation well suited to neuromorphic hardware, but direct training remains difficult due to non-differentiable spike dynamics. ANN-to-SNN conversion methods address these challenges by leveraging established artificial neural network (ANN) training frameworks. However, tailoring conversion methods to specific applications requires a clear understanding of the trade-offs in accuracy, latency, and spike efficiency across datasets of varying complexity. In this paper, we benchmark five widely used ANN-to-SNN conversion techniques: Rate-based Coding, Temporal Coding, Threshold Adjustment, Surrogate Gradient Descent, and Hybrid ANN-SNN Training, on MNIST, Fashion-MNIST, and CIFAR-10. Our analysis quantifies the key trade-offs between accuracy, latency, and spike efficiency, showing that Hybrid and Temporal Coding approaches achieve the best balance between accuracy and efficiency, particularly as dataset complexity increases. We further demonstrate that dataset characteristics play a crucial role in determining the optimal conversion method, providing practical guidance for selecting approaches best suited to energy- and latency-constrained environments.