SpikF: Spiking Fourier Network for Efficient Long-term Prediction
Authors: Wenjie Wu, Dexuan Huo, Hong Chen
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations on eight well-established long-term prediction datasets demonstrate that Spik F achieves an averaged 1.9% reduction in Mean Absolute Error (MAE) compared to state-of-the-art models, while lowering total energy consumption by 3.16 . Our code is available at https://github.com/WWJcreator/Spik F. We conduct empirical evaluations across eight real-world long-term time-series benchmark datasets, demonstrating that Spik F achieves superior accuracy over SOTA ANN models, with a 1.9% reduction in MAE. Our extensive experiments across Electricity, Weather, ETT, Traffic and Exchange datasets validate the robustness of our approach and establish a foundational SNN benchmark for the research community. |
| Researcher Affiliation | Academia | 1Tsinghua University, Beijing, China. Correspondence to: Hong Chen <EMAIL>. All authors are affiliated with Tsinghua University, which is an academic institution, and the email domain 'tsinghua.edu.cn' confirms this. |
| Pseudocode | Yes | Algorithm 1 Fast Fourier Transform |
| Open Source Code | Yes | Our code is available at https://github.com/WWJcreator/Spik F. |
| Open Datasets | Yes | The Electricity dataset (Wu et al., 2022) records hourly interval electricity consumption data from 321 clients. The Weather dataset (Wu et al., 2022) includes meteorological measurements collected every 10 minutes from the Weather Station of the Max Planck Biogeochemistry Institute in 2020, containing 21 variables. The ETT dataset (Zhou et al., 2021) consists of four subsets: ETTh1 and ETTh2 with hourly interval data, and ETTm1 and ETTm2 with 15-minute interval data, each containing two years (2016-2018) of measurements from electrical transformers, including variables like oil temperature and load. The Traffic dataset (Wu et al., 2022) provides hourly road occupancy rates from 862 sensors in the San Francisco Bay area between 2015 and 2016. The Exchange dataset (Wu et al., 2022) comprises daily exchange rate data from eight countries spanning from 1990 to 2016. Solar-Energy (Lai et al., 2018), provides a dataset documenting the solar power generation of 137 photovoltaic plants throughout the year 2006, with measurements recorded at 10-minute intervals. |
| Dataset Splits | Yes | The train-validation-test split method follow the i Transformer (Liu et al., 2024) and i Spikformer (Lv et al., 2024b) approach to ensure fairness. Detailed dataset information is provided in Table 6. Table 6: Dataset descriptions. The dataset size is organized in (Train, Validation, Test). For example, Electricity: (18317, 2633, 5261). |
| Hardware Specification | Yes | The experiments are executed on a single NVIDIA 3090 GPU. |
| Software Dependencies | No | All experiments are built on the frameworks developed on Py Torch (Paszke et al., 2019) and Spiking Jelly (Fang et al., 2023), the latter being an SNN repository built upon Py Torch. The paper mentions PyTorch and Spiking Jelly but does not provide specific version numbers for these software components or the Python version used. |
| Experiment Setup | Yes | The surrogate function for spatio-temporal backpropagation (STBP) is selected as the Sigmoid function, defined as: σ(x) = 1 / (1 + e^-αx), where the parameter α is set to 4.0, consistent with previous studies (Zhou et al., 2023; 2024) across all experiments. The parameters for the spiking neurons are predominantly set to their default values within the Spiking Jelly library, with the time constant τm fixed at 2.0 and the threshold potential vth at 1.0. An exception is made for the generator LIF neurons, where vth was adjusted to 0.1 to improve passing rate of the selector. The Adam optimizer is employed with its parameters configured as follows: β1 = 0.9 and β2 = 0.999. The temporal resolution Ts is set to 16 for datasets exhibiting minimal variables and 4 for those with greater variables. The feature extraction layers of the architecture are chosen from the set {1, 2, 4}, and the patch dimension are selected from {8, 16, 32, 64}. The hidden layer dimension within the encoder s MLP are chosen from {180, 360, 540, 720}. Batch sizes are selected from {4, 8, 16, 32, 64}, and initial learning rates are set from {10^-4, 5 x 10^-4, 10^-3}. The number of training epochs is chosen from {5, 10, 15, 20}. An early-stopping strategy is implemented to optimize the use of the validation set for model evaluation. All model parameters are fine-tuned based on validation set performance. A comprehensive tabulation of parameter selections is provided in Table 7. |