Spiking Neural Networks (SNNs) exhibit superior energy efficiency and biological plausibility due to their event-driven, sparse computations in temporal tasks. However, conventional convolution-based spiking models struggle to capture long-range dependencies and cross-channel interactions, thereby limiting their ability to represent complex temporal patterns. To address this limitation, this article introduces a parallel spiking temporal modeling framework that synergizes heterogeneous convolutional and gated recurrent architectures (Spiking Temporal Gated, STG) with adaptive positional encoding. Firstly, we design a parallel module consisting of two branches: spiking Heterogeneous Temporal Convolution Network (SHTCN), which fuses 1D causal convolutions with 2D cross-channel convolutions to extract short-term local features; spike-gated recurrent unit (GRU), a spiking-driven GRU that captures long-term global dependencies. Secondly, we propose Gated Adaptive Positional Encoding (GAPE), which employs learnable gating weights to dynamically modulate multi-frequency positional encodings, thereby enhancing the model’s sensitivity to temporal order and periodic structures. Extensive experiments on multiple multivariate time-series forecasting benchmarks demonstrate that STG-GAPE yields notable improvements in predictive accuracy and robustness compared with existing SNN-based approaches, while maintaining low computational complexity and a compact parameter footprint. These results validate the effectiveness of combining parallel heterogeneous modeling with adaptive positional encoding in SNNs.
Wang et al. (Wed,) studied this question.