Research map / Label-efficient and robust learning / Domain adaptation and generalization
SpGesture: Source-Free Domain-adaptive sEMG-based Gesture Recognition with Jaccard Attentive Spiking Neural Network
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 SpGesture, a spiking neural network for sEMG gesture recognition with a Spiking Jaccard Attention module and a source-free domain adaptation method using membrane-potential memory, plus a new multi-posture sEMG dataset.
Problem
sEMG gesture recognition with SNNs suffers from accuracy loss, distribution shift across forearm postures, and attention being hard to apply to sparse binary spike features; source data may also be unavailable for privacy reasons.
Approach
Computes attention directly on spike trains using a Jaccard similarity built from element-wise min/max (comparison/addition-friendly), and performs source-free adaptation by generating pseudo-labels from k-nearest neighbors in a membrane-potential memory list (Pearson similarity) with random exploration.
Key results
- JASNN reaches 89.26% top-1 accuracy on the self-collected dataset, versus 88.32% for the best DNN baseline (Informer) and 86.24% for LSNN.
- Spiking Jaccard Attention raises SCNN accuracy from 84.12% to 87.44%, while raw attention on spikes drops to 11.31%.
- Source-free adaptation (PLG + SNLL/KL) improves accuracy by 4.10% over the unadapted model; the paper reports Spiking Jaccard Attention inference 36.37x faster than traditional attention on CPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SJA attention instead of no attention (ablation) | SCNN · Collected sEMG gesture dataset | Accuracy (%): 84.12% -> 87.44% | Helped |
| Add Raw Attention directly to spike sequences (ablation) | SCNN · Collected sEMG gesture dataset | Accuracy (%): 84.12% -> 11.31% | Hurt |
| Use MA-SNN attention instead of no attention (ablation) | SCNN · Collected sEMG gesture dataset | Accuracy (%): 84.12% -> 85.67% | Helped |
| Replace NLL with SNLL+KLL (ablation) | JASNN · Collected sEMG gesture dataset | Accuracy (%): 87.44% -> 89.72% | Helped |
| Use PL pseudo-label generation (ablation) | JASNN · Source-free adaptation | Improved accuracy: +1.87% | Helped |
| Use NPL pseudo-label generation (ablation) | JASNN · Source-free adaptation | Improved accuracy: +2.33% | Helped |
| Use PLG pseudo-label generation (ablation) | JASNN · Source-free adaptation | Improved accuracy: +4.1% | Helped |
| Use PLG pseudo-label generation (ablation) | MA-SNN · Source-free adaptation | Improved accuracy: +3.81% | Helped |
| Add SJA to LSNN | LSNN · Collected sEMG gesture dataset | Top-1 accuracy (%): 86.24 -> 88.10 | Helped |
| Apply Spiking Jaccard Attention rather than conventional attention | SpGesture · CPU attention inference | Speedup: 36.37x | Helped |
Surprising findings
- Raw attention applied directly to spike trains collapses accuracy to 11.31%, far below the 84.12% SCNN baseline.
Limitations
- Evaluated only on a Spiking CNN backbone and one self-collected dataset
- Efficiency measured on CPU and GPU, not on neuromorphic hardware
- Element-wise Jaccard weighting not validated
Experimental setup
Datasets: self-collected 15-subject sEMG gesture dataset with three forearm postures
Benchmarks: self-collected multi-posture sEMG gesture dataset
Metrics: Top-1 accuracy standard deviation inference time RAM usage
Baselines: CNN (Asif et al. 2020) TCN (Tsinganos et al. 2020) Transformer (Rahimian et al. 2021) GRU (Chen et al. 2021) Informer (Zhou et al. 2021) LSNN (Bellec et al. 2018) SIB+SNN (Zhang et al. 2022) SCNN (Xu et al. 2023)
Builds on: SuperSpike surrogate gradient LSNN MA-SNN Norse
Closest papers by meaning
- emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (NeurIPS 2024 · similarity 0.89)
- Revisiting Noise Resilience Strategies in Gesture Recognition: Short-Term Enhancement in sEMG Analysis (ICML 2025 · similarity 0.88)
- emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation (NeurIPS 2024 · similarity 0.88)
- SplashNet: Split‑and‑Share Encoders for Accurate and Efficient Typing with Surface Electromyography (NeurIPS 2025 · similarity 0.85)
- Posture-Informed Muscular Force Learning for Robust Hand Pressure Estimation (NeurIPS 2024 · similarity 0.84)
- EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography (NeurIPS 2024 · similarity 0.84)
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.