Mixed-timestep spiking neural networks: a temporal alignment framework for high-efficiency neuromorphic computing

Spiking Neural Networks (SNNs) are energy-efficient due to their event-driven computation and sparse spike-based communication. However, most existing SNN models adopt a uniform time-step across all layers, which limits temporal flexibility and leads to suboptimal accuracy latency trade-offs under low-latency inference. To address this limitation, we propose Mixed-Timestep Spiking Neural Networks (MT-SNNs), a paradigm in which different layers operate with different time-steps, enabling layer-wise temporal allocation for efficient inference. To construct MT-SNNs, we develop a quantization-aware ANN-to-SNN conversion framework. Specifically, we establish a theoretical relationship between ANN activation bit-width and SNN time-steps, showing that the SNN time-step T can be approximated by the activation quantization level 2 n of the source ANN. Based on this relationship, layer-wise activation bit-widths in ANNs can be mapped to layer-wise time-steps in SNNs. We then jointly optimize quantization levels and firing thresholds to obtain compact mixed-timestep configurations that reduce inference latency while preserving accuracy. Finally, we identify the temporal dimension mismatch issue caused by heterogeneous layer-wise time-steps and propose Scaled Synaptic Current Accumulation (SSCA), a fused temporal alignment mechanism that aligns inter-layer propagation in the synaptic-current domain without instantiating a standalone temporal-alignment kernel. Extensive experiments on CIFAR-10, CIFAR-100, ImageNet-1K, CIFAR10-DVS, and DVS-Gesture demonstrate the effectiveness of the proposed method. On ImageNet-1K, MT-SNNs based on ResNet-34 achieve 73.63% Top-1 accuracy with only 4.88 average time-steps, demonstrating a favorable accuracy latency trade-off for low-latency SNN conversion.