Research map / Architectures and efficiency / Spiking networks
S$^2$M-Former: Spiking Symmetric Mixing Branchformer for Brain Auditory Attention Detection
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 S²M-Former, a spiking symmetric dual-branch (CSP spatial and DE frequency) network with 1D token mixers for EEG-based auditory attention detection, achieving competitive accuracy with 0.06M parameters and lower energy than dual-branch ANNs.
Problem
EEG-based auditory attention detection needs accurate yet energy-efficient models that fuse complementary EEG features; existing dual-branch models are parameter-heavy (3D convolutions) and combine branches in isolation.
Approach
A spike-driven symmetric dual-branch architecture (spatial CSP branch and frequency DE branch) with mirrored modules that use 1D token sequences instead of 3D operations, plus cross-branch gated channel mixers and membrane-potential-aware token mixers, and a channel-wise parametric LIF neuron (CPLIF).
Key results
- Within-trial on DTU at 2s: S²M-Former 85.28% ± 6.01 vs DBPNet 83.93% ± 5.17 with 0.06M vs 0.88M parameters.
- Cross-trial on DTU at 2s: S²M-Former 76.74% ± 9.96, the highest among compared methods.
- Energy: 0.0779 mJ vs 0.4526 mJ for DBPNet (reported 5.8× reduction) at 2s cross-trial window.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ANN operations with spike-driven components (ablation) | S^2M-Former on DTU · within-trial, 2-second | accuracy: 80.94 ± 5.66 -> 85.28 ± 6.01 | Helped |
| replace standard LIF neurons with channel-wise CPLIF neurons (ablation) | S^2M-Former on DTU · within-trial, 2-second | accuracy: 84.13 ± 5.96 -> 85.28 ± 6.01 | Helped |
| add SGCM and MPTM for complementary fusion (ablation) | S^2M-Former on DTU · within-trial, 2-second | accuracy: 82.98 ± 6.25 -> 85.28 ± 6.01 | Helped |
| add the frequency branch to the spatial branch (ablation) | S^2M-Former on DTU · within-trial, 2-second | accuracy: 82.28 ± 7.13 -> 85.28 ± 6.01 | Helped |
| add the spatial branch to the frequency branch (ablation) | S^2M-Former on DTU · within-trial, 2-second | accuracy: 70.48 ± 7.78 -> 85.28 ± 6.01 | Helped |
| combine SCSA global modeling with SMSC local modeling (ablation) | S^2M-Former on DTU · within-trial, 0.1-second | accuracy: 60.79 ± 4.23 -> 75.84 ± 5.46 | Helped |
| combine SCSA global modeling with SMSC local modeling (ablation) | S^2M-Former on DTU · within-trial, 1-second | accuracy: 78.51 ± 6.64 -> 82.87 ± 6.92 | Helped |
| combine SCSA global modeling with SMSC local modeling (ablation) | S^2M-Former on DTU · within-trial, 2-second | accuracy: 81.54 ± 6.66 -> 85.28 ± 5.61 | Helped |
| increase spiking time steps beyond 4 (ablation) | S^2M-Former · cross-trial, 2-second, KUL, DTU, and AV-GC | accuracy | No clear effect |
| replace the full S^2M-Former with the spatial branch alone (ablation) | S^2M-Former on KUL · cross-trial | accuracy: 70.18% vs. 69.87% | Mixed |
| replace the full S^2M-Former with the frequency branch alone (ablation) | S^2M-Former on AV-GC · within-trial and cross-trial | accuracy | Hurt |
| use ANN counterpart SM-Former instead of S^2M-Former (ablation) | S^2M-Former on AV-GC · cross-trial, 1-second | accuracy: 65.77 ± 15.58 -> 66.52 ± 13.75 | Hurt |
Surprising findings
- The ANN counterpart SM-Former slightly outperforms S²M-Former on AV-GC at the 1s cross-trial window.
- Single-branch SNN frequency branch alone (0.01M) matches or exceeds larger SNN transformers such as QKFormer and Spikformer on KUL and AV-GC.
Limitations
- Performance gap remains between within-trial and cross-trial settings, with high standard deviations on KUL and AV-GC.
- No statistical significance tests or confidence intervals are reported.
- Energy consumption is higher than single-branch SNNs due to dual-branch input features.
Experimental setup
Datasets: KUL DTU AV-GC-AAD
Benchmarks: KUL DTU AV-GC-AAD
Metrics: accuracy standard deviation parameter count FLOPs SOPs energy (mJ)
Baselines: SSF-CNN MBSSFCC DARNet DBPNet M-DBPNet QKFormer Spike-driven Transformer Spikformer SM-Former (ANN)
Builds on: DBPNet M-DBPNet DARNet Spikformer QKFormer Spike-driven Transformer SpikingJelly
Closest papers by meaning
- Spike-based Neuromorphic Model for Sound Source Localization (NeurIPS 2024 · similarity 0.87)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.87)
- Brain-Inspired Spiking Neural Networks for Energy-Efficient Object Detection (CVPR 2025 · similarity 0.86)
- DARNet: Dual Attention Refinement Network with Spatiotemporal Construction for Auditory Attention Detection (NeurIPS 2024 · similarity 0.85)
- Spiking Neural Networks Need High-Frequency Information (NeurIPS 2025 · similarity 0.85)
- FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping (NeurIPS 2026 · similarity 0.85)
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.