Research map / Architectures and efficiency / Spiking networks
FSTA-SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks
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.
Analyzes spike-output frequency spectra of SNNs and proposes FSTA, a plug-and-play DCT-based spatial plus temporal attention module that boosts accuracy and cuts spike firing rate.
Problem
SNNs' intermediate output spikes are rarely analyzed or optimized, which limits their energy efficiency and spatiotemporal feature extraction; spike redundancy is not understood from a network-level learning-preference perspective.
Approach
Frequency-domain analysis of spike outputs reveals layer-wise learning preferences (shallow layers vertical, deeper horizontal) and time-step consistency; based on this, a serial temporal attention plus DCT full-spectrum spatial attention module suppresses redundant spikes and is inserted between SNN layers.
Key results
- On CIFAR-10 with ResNet20 (T=4), FSTA-SNN reaches 94.72% and with ResNet19 (T=2) 96.52%, improving prior best by 0.44% and 0.05%.
- On ImageNet, FSTA-SNN achieves 68.21% (ResNet18) and 70.23% (ResNet34) at T=4, and on CIFAR10-DVS 82.70% at T=16.
- The module reduces overall network spike firing rate by about 33.99%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FSTA spatial-temporal attention | ResNet19 · CIFAR-10, T=2 | accuracy: 95.51% -> 96.52% | Helped |
| Add FSTA spatial-temporal attention | ResNet20 · CIFAR-10, T=4 | accuracy: 94.28% -> 94.72% | Helped |
| Add FSTA spatial-temporal attention | ResNet19 · CIFAR-100, T=2 | accuracy: 79.51% -> 80.42% | Helped |
| Add FSTA spatial-temporal attention | ResNet34 · ImageNet, T=4 | accuracy: 67.69% -> 70.23% | Helped |
| Add FSTA spatial-temporal attention | ResNet20 · CIFAR10-DVS, T=10 | accuracy: 78.70% -> 81.50% | Helped |
| Add FSTA spatial-temporal attention | SNN · Across the network | spike firing rate reduction: 33.99% | Helped |
| Use serial rather than parallel submodule combination (ablation) | FSTA-SNN, ResNet20 · CIFAR-100, T=2 | accuracy: +1.32% | Helped |
| Add channel attention enhancement (ablation) | FSTA-SNN, ResNet20 · CIFAR-100, T=2 | accuracy: 72.15% -> 72.68% | Helped |
| Use a 5x5 rather than 3x3 frequency kernel (ablation) | FSTA-SNN, ResNet20 · CIFAR-10 | accuracy: 92.36% -> 92.58% | Helped |
| Use a 7x7 rather than 5x5 frequency kernel (ablation) | FSTA-SNN, ResNet20 · CIFAR-10 | accuracy: 92.58% -> 93.01% | Helped |
Surprising findings
- Spectral analysis suggests SNN learning preferences are largely insensitive to time step, so increasing time steps has limited effect on feature learning.
Limitations
- Gains on the largest comparisons are small (e.g., 0.05% on CIFAR-10 ResNet19) and reported with ±std only for the proposed method.
- Attention modules add computational cost; channel attention variant increases MACs and parameters.
- Analysis relies on empirical spectral observations rather than a formal theoretical justification.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Benchmarks: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Metrics: top-1 accuracy spike firing rate ACs MACs energy (mJ)
Baselines: SpikeNorm Hybrid-Train PTL DSR KDSNN Joint A-SNN RMP-Loss RecDis-SNN TET Real Spike MPBN GLIF
Builds on: Leaky Integrate-and-Fire neuron TET MPBN RecDis-SNN SEW ResNet
Closest papers by meaning
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.92)
- Unveiling the Spatial-temporal Effective Receptive Fields of Spiking Neural Networks (NeurIPS 2025 · similarity 0.89)
- Advancing Spiking Neural Networks Towards Multiscale Spatiotemporal Interaction Learning (AAAI 2025 · similarity 0.89)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.89)
- TS-SNN: Temporal Shift Module for Spiking Neural Networks (ICML 2025 · similarity 0.89)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.88)
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.