Research map / Architectures and efficiency / Spiking networks
SpikF: Spiking Fourier Network for Efficient Long-term Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SpikF, an attention-free spiking neural network for long-term time-series forecasting that combines patch-based spiking encoding with a spiking Fourier frequency selector, reporting lower MAE and energy than ANN baselines on eight datasets.
Problem
Spiking neural networks are underexplored for long-term time-series prediction because existing spike encoders handle long input histories poorly and spiking self-attention lacks effective positional encoding.
Approach
Encode input series as patches into spike trains, then select relevant frequency components via a spiking Fourier selector (grouped S-FFT, spiking max-pooling selection mask, S-iFFT) instead of self-attention, plus a temporal-synchronization MLP decoder.
Key results
- SpikF averages a 1.9% MAE reduction versus iTransformer across eight long-term datasets, with a 6.1% lower MAE on ETTm1.
- Total energy consumption is 3.16x lower than iTransformer (1.37e4 µJ vs 4.34e4 µJ) on ECL with look-back 96 and horizon 720.
- On Solar-energy short-term prediction, SpikF reduces RSE by 2.3% vs iSpikformer and 2.8% vs iTransformer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Spiking Patch Encoding instead of delta encoding (ablation) | SpikF · ETTh1 | MAE: 0.450 -> 0.428 | Helped |
| Use Spiking Patch Encoding instead of convolution encoding (ablation) | SpikF · ETTh1 | MAE: 0.446 -> 0.428 | Helped |
| Add the Spiking Frequency Selector rather than removing feature extraction (ablation) | SpikF · ETTm1 | MAE: 0.387 -> 0.385 | Helped |
| Use the Spiking Frequency Selector instead of inverted spiking self-attention (ablation) | SpikF · ETTh1 | MAE: 0.440 -> 0.428 | Helped |
| Add temporal synchronization to the decoder (ablation) | SpikF · ETTh1 | MAE: 0.430 -> 0.428 | Helped |
| Use SpikF instead of DLinear for feature extraction | SpikF · Electricity | operational energy consumption: 205.19 -> 117.66 | Helped |
| Use SpikF instead of iTransformer across all modules | SpikF · ECL | operational energy consumption: 75.05% reduction | Helped |
Surprising findings
- Deeper spiking frequency selector layers tend to discard more frequency components, while shallower layers retain useful frequencies.
Limitations
- Energy figures are theoretical estimates based on SOP/FLOP counts and per-operation energy constants rather than measured hardware energy.
- Gains over baselines are modest (about 1.9% MAE on average), and performance is sensitive to the temporal resolution parameter Ts within a few percent.
Experimental setup
Datasets: ECL Weather ETTh1 ETTh2 ETTm1 ETTm2 Traffic Exchange Solar-energy
Benchmarks: ECL Weather ETTh1 ETTh2 ETTm1 ETTm2 Traffic Exchange Solar-energy
Metrics: MSE MAE RSE energy consumption (SOPs/energy in µJ)
Baselines: iTransformer PatchTST Crossformer TimesNet DLinear RLinear SCINet Autoformer iSpikformer
Builds on: PatchTST FITS FEDformer Spikformer iSpikformer
Closest papers by meaning
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.89)
- Unveiling the Spatial-temporal Effective Receptive Fields of Spiking Neural Networks (NeurIPS 2025 · similarity 0.88)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.88)
- FSTA-SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks (AAAI 2025 · similarity 0.88)
- LongSpike: Fractional Order Spiking State Space Models for Efficient Long Sequence Learning (NeurIPS 2026 · similarity 0.88)
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.