Research map / Architectures and efficiency / Pruning and efficient tuning
SplashNet: Split‑and‑Share Encoders for Accurate and Efficient Typing with Surface Electromyography
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 SplashNet, a set of three cheap modifications (causal rolling time normalization, aggressive channel masking, and weight-shared split-per-hand encoders) plus 33-to-6 spectral band reduction, improving zero-shot and finetuned wrist-sEMG typing on emg2qwerty.
Problem
Wrist sEMG text entry degrades sharply on unseen users (51.8% CER zero-shot for the emg2qwerty baseline), and cross-user signal shift and fragile high-order feature reliance limit generalization.
Approach
Combine per-feature causal rolling z-score normalization, very aggressive electrode/band masking to favor low-order feature combinations, and a Split-and-Share encoder that processes each hand with shared weights, plus reduced spectral resolution, to get a smaller and more accurate sEMG decoder.
Key results
- SplashNet reduces zero-shot test CER from 51.78% (baseline) to 35.67%, and SplashNet-mini reaches 36.41% with 1.38M parameters.
- After finetuning, SplashNet achieves 5.51% test CER versus 6.95% for the baseline (about 21% relative improvement).
- Greedy decoding (no LM) favors the Joint-Hand baseline on zero-shot and finetuned unseen users, a reported limitation of the approach.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the 33-bin spectrogram with six frequency bands (ablation) | emg2qwerty baseline · emg2qwerty zero-shot, Test domain test | CER (%): 51.78 ± 4.61 -> 47.18 ± 5.19 | Helped |
| use Rolling Time Normalization instead of batch-level normalization (ablation) | emg2qwerty baseline with RSG · emg2qwerty zero-shot, Test domain test | CER (%): 47.18 ± 5.19 -> 39.15 ± 6.20 | Helped |
| add Aggressive Channel Masking without RTN (ablation) | emg2qwerty baseline with RSG · emg2qwerty zero-shot, Test domain test | CER (%): 47.18 ± 5.19 -> 42.62 ± 7.10 | Helped |
| add RTN and ACM (ablation) | emg2qwerty baseline with RSG · emg2qwerty zero-shot, Test domain test | CER (%): 47.18 ± 5.19 -> 36.42 ± 7.11 | Helped |
| process each hand separately without sharing encoder weights (ablation) | Joint-Hand baseline with RSG, RTN, and ACM · emg2qwerty zero-shot, Test domain test | CER (%): 36.42 ± 7.11 -> 37.37 ± 7.34 | Hurt |
| use split, weight-shared hand encoders (ablation) | Joint-Hand baseline with RSG, RTN, and ACM · emg2qwerty zero-shot, Test domain test | CER (%): 36.42 ± 7.11 -> 36.41 ± 7.30 | Helped |
| upscale the SplashNet-mini embedding and final convolutional layers (ablation) | SplashNet-mini · emg2qwerty zero-shot, Test domain test | CER (%): 36.41 ± 7.30 -> 35.67 ± 6.79 | Helped |
| add Aggressive Channel Masking without RTN (ablation) | emg2qwerty baseline with RSG · emg2qwerty fine-tuned, Test domain test | CER (%): 6.92 ± 3.79 -> 7.45 ± 3.87 | Hurt |
| add RTN to the RSG model (ablation) | emg2qwerty baseline with RSG · emg2qwerty fine-tuned, Test domain test | CER (%): 6.92 ± 3.79 -> 6.63 ± 3.58 | Helped |
| use split, weight-shared hand encoders (ablation) | Joint-Hand baseline with RSG, RTN, and ACM · emg2qwerty fine-tuned, Test domain test | CER (%): 6.53 ± 3.27 -> 5.87 ± 3.04 | Helped |
| unshare encoder weights during fine-tuning (ablation) | SplashNet-mini · emg2qwerty fine-tuned, Test domain test | CER (%): 5.87 ± 3.04 -> 5.96 ± 3.28 | Hurt |
| upscale the SplashNet-mini embedding and final convolutional layers (ablation) | SplashNet-mini · emg2qwerty fine-tuned, Test domain test | CER (%): 5.87 ± 3.04 -> 5.51 ± 2.81 | Helped |
Surprising findings
- Under greedy decoding the Joint-Hand baseline outperforms the Split-and-Share models, with the gap reversing only once a 6-gram character LM is applied.
- Matching the masking value to the per-sample feature mean with BatchNorm does not reproduce RTN+ACM gains, suggesting a stable masking value matters more.
Limitations
- Under greedy decoding without an external LM, the Joint-Hand baseline performs better than the proposed models.
- Only healthy participants were studied; transfer to motor or limb-difference populations is untested.
- The split architecture still merges hand information at the final linear layer, so fully on-band independent inference is not achieved.
Experimental setup
Datasets: emg2qwerty
Benchmarks: emg2qwerty
Metrics: character error rate GFLOPs parameter count
Baselines: emg2qwerty baseline (Sivakumar et al. 2024) Joint-Hand baseline
Builds on: emg2qwerty SpecAugment TDSConv emg2pose
Closest papers by meaning
- emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (NeurIPS 2024 · similarity 0.91)
- emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation (NeurIPS 2024 · similarity 0.87)
- EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography (NeurIPS 2024 · similarity 0.85)
- SpGesture: Source-Free Domain-adaptive sEMG-based Gesture Recognition with Jaccard Attentive Spiking Neural Network (NeurIPS 2024 · similarity 0.85)
- Revisiting Noise Resilience Strategies in Gesture Recognition: Short-Term Enhancement in sEMG Analysis (ICML 2025 · similarity 0.84)
- Electromyography-Informed Facial Expression Reconstruction for Physiological-Based Synthesis and Analysis (CVPR 2025 · similarity 0.82)
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.