PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-User Adaptation for sEMG Gesture Recognition via Continual Learning

Aran Ramakrishnan, Shubhi Mishra, Tyler Ham, Assel Kembay, Jason Eshraghian

arXiv:2610.34011v1cs.LGeess.SP

TL;DR

Person-specific sEMG signals make gesture recognition models need adaptation to new wearers and gestures, but continual-learning strategies for recurrent sEMG models have limited evaluation under realistic drift. The paper compares strategies across domain- and class-incremental settings and finds replay most robust, while showing that shuffled, independently windowed training and rehearsal matter for stable learning and feature retention.

  • Problem

    Continual-learning strategies have been evaluated largely outside the recurrent models common in sEMG, leaving limited evidence about which strategies withstand realistic cross-user drift.

    Source: §I · p. 1

  • Method

    The study compares naive training, EWC, and experience replay across domain- and class-incremental drift using LSTM and HGRN models on NinaPro DB1.

    Source: §I · p. 1, §II · p. 1

  • Results

    83.3% final accuracy was reached by CIL-LSTM with stateless replay, which maintained 1.8%–5.8% accuracy drop in forgetting across replay methods.

    Source: A. CL Method Comparison · p. 3

  • Takeaways & Limitations

    Shuffled, stateless training stabilizes learning on independently windowed segments, and under class-incremental drift rehearsal is needed to retain time-domain feature gains.

    Source: §I · p. 1, B. Feature Representation · p. 3

Abstract

from arXiv · show

Prosthetic and assistive interfaces increasingly depend on decoding hand gestures from surface electromyography (sEMG). Because the signal is shaped by person-specific factors, the recurrent neural networks (RNNs) commonly used to model these temporal streams must adapt to new wearers and new gesture classes after deployment. Continual learning (CL) enables this adaptation without retraining from scratch, yet applying it to recurrent sEMG models is not straightforward. Standard recurrent training practices can destabilize learning, and which CL strategy holds up under realistic cross-user drift is unestablished. We evaluate three representative CL strategies under domain- and class-incremental drift on NinaPro DB1, and find experience replay to outperform elastic weight consolidation (EWC) and naive fine-tuning. It reaches 83.3% final accuracy for class-incremental drift while limiting forgetting to 1.8-5.8% accuracy drop, as against 16.2-28.5% for EWC and naive training. We further show these models must be trained on shuffled, independently-windowed segments. Streaming windows in temporal order collapses the network into predicting a single dominant gesture, while a shuffled, stateless regime trains stably. Finally, hand-crafted time-domain features raise per-task accuracy by up to about 40% over raw signals, but only replay preserves that gain under class-incremental drift.

I. INTRODUCTION

sEMG gesture recognition is person-specific, motivating continual adaptation to new wearers and gestures without retraining from scratch; recurrent sEMG models remain underexplored in continual-learning research. This paper compares strategies across domain- and class-incremental drift and recurrent model families, finding replay most robust and identifying training-order and feature-retention effects.

I. INTRODUCTION

  • Experience replay is the most robust strategy across domain- and class-incremental drift and both recurrent architectures, with the continual-learning method outweighing architecture choice.The comparison spans NinaPro DB1 and builds on prior work that largely used a single continual-learning scenario or non-recurrent decoders –.Source: §I · p. 1, §II · p. 1
  • Carrying hidden state across tasks is ill-posed for these independently windowed segments, and temporal-order training collapses prediction to a majority class.The paper traces the collapse primarily to batch ordering rather than recurrent memory and reports that purging statefulness between tasks and shuffling samples corrects the regime.Source: §I · p. 1
  • Time-domain features improve per-task learning, but only rehearsal preserves that improvement under class-incremental drift.This is the paper’s third finding after the corrected shuffled, stateless training regime.Source: §I · p. 1

II. PRIOR WORK

  • Prior sEMG continual-learning studies mostly used a single scenario and feedforward or generative decoders, leaving recurrent models largely untested.Chen et al. grouped subjects rather than treating each as a sequential task; Kanoga et al. and Yoneda et al. [6] examined memory mechanisms or model families.Source: §II · p. 1

A. Dataset and Task Formulation

The evaluation uses NinaPro DB1 and tests continual adaptation across subjects and an expanding gesture set. It also specifies repetition-level splits, signal preprocessing, and a time-domain feature representation.

A. Dataset and Task Formulation

  • NinaPro DB1 supplies data from 27 subjects and three exercises containing isotonic and isometric hand and wrist gestures.The exercises contain 12, 17, and 23 gestures, respectively.Source: A. Dataset and Task Formulation · pp. 1–2
  • The DIL and CIL formulations follow van de Ven & Tolias [10], with DIL testing cross-subject adaptation and CIL testing an expanding class set.The CIL label indices are offset sequentially for new gestures.Source: A. Dataset and Task Formulation · p. 2
  • Domain-incremental learning keeps Exercise 1 gestures fixed while treating each of the 27 subjects as a sequential task.This setup tests generalization across anatomical variation and sensor-placement discrepancies between individuals.Source: A. Dataset and Task Formulation · p. 2
  • Class-incremental learning presents Exercises 1, 2, and 3 sequentially, expanding the gesture labels while requiring retention of earlier performance.For HGRN, the classification head is sized at initialization for all classes, and the rest state remains shared across tasks.Source: A. Dataset and Task Formulation · p. 2

B. Signal Preprocessing and Feature Extraction

  • Gesture executions are split at the repetition level into 70% training and 30% test sets before normalization and sliding-window segmentation.The window label is the active stimulus at the final sample; normalization parameters are fitted on training data and applied to both sets.Source: B. Signal Preprocessing and Feature… · p. 2
  • Each 200-sample window becomes 10 sub-windows of 20 samples, represented by Hudgins features MAV, RMS, WL, ZC, and SSC.ZC and SSC use a noise-rejection threshold of 0.01σ, followed by per-feature standardization fitted on training windows.Source: B. Signal Preprocessing and Feature… · p. 2

C. Base Network Architectures

The study compares LSTM and HGRN recurrent models under a shared four-layer configuration and classification-head design. HGRN uses a depth-scaled lower bound in its gated recurrence.

C. Base Network Architectures

1) Long Short-Term Memory:

  • The LSTM uses explicit memory cells to mitigate vanishing gradients and serves as the standard sequential baseline.The paper identifies it as a standard baseline for this experiment.Source: 1) Long Short-Term Memory · p. 2

2) Hierarchical Gated Recurrent Network:

D. Continual Learning Methods

The study establishes stateless and stateful naive baselines, then describes experience replay as a rehearsal strategy. The baselines differ in whether hidden state resets or carries across batches and tasks.

1) Naive Stateless and Stateful Baselines:

  • The stateless baseline resets recurrent hidden state to zero at each batch, while the stateful baseline carries it across batches within and between tasks.These are lower-bound regimes before any memory-retention mechanism is applied; statelessness matches independently windowed EMG segments.Source: 1) Naive Stateless and Stateful… · p. 2

2) Standard Experience Replay:

3) Online Elastic Weight Consolidation (EWC):

Online EWC estimates and accumulates empirical Fisher information across completed tasks to regularize changes to the latest task’s weights. The evaluation reports final accuracy, forgetting, and plasticity.

3) Online Elastic Weight Consolidation (EWC):

  • Online EWC accumulates empirical Fisher information across prior tasks and adds a quadratic penalty for deviations from the most recently completed task’s weights.The Fisher is estimated from predicted class probabilities on the completed task’s training data; λ_ewc is set to 30, the largest value preserving plasticity across the full sequence.Source: 3) Online Elastic Weight Consolidation… · p. 2

E. Evaluation Metrics

  • Final accuracy measures accuracy on all tasks after the final task.The evaluation metrics are computed from R, where R_i,j is accuracy on Task j after training on Task i.Source: E. Evaluation Metrics · pp. 2–3
  • Forgetting measures the average per-task accuracy drop after all subsequent tasks have been trained.Source: E. Evaluation Metrics · p. 3
  • Plasticity is the mean final-epoch training accuracy over the last min(10, T) tasks, distinguishing learning from weight freezing.Those tasks are indexed by S = {T − min(10, T) + 1, . . . , T}.Source: E. Evaluation Metrics · p. 3

A. CL Method Comparison

In the stateless, time-domain-feature comparison, replay achieves the strongest resistance to forgetting across the evaluated scenarios and architectures. CIL replay has lower forgetting than DIL replay, consistent with the shorter CIL task stream.

A. CL Method Comparison

  • 83.3% final accuracy for CIL-LSTM and 81.6% for CIL-HGRN; replay’s forgetting was 1.8%–5.8% accuracy drop, compared with EWC’s 16.2%–28.5%.Table I compares continual-learning methods using stateless time-domain features.Source: Table I · p. 3, A. CL Method Comparison · p. 3
  • Replay forgetting was 1.8%–2.3% accuracy drop under CIL versus 4.3%–5.8% under DIL; the shorter three-task CIL stream versus 27 DIL tasks helps explain the difference.Forgetting accumulates with sequence length, so it is expected to be smaller under CIL on this basis alone.Source: A. CL Method Comparison · p. 3
  • Under replay, immediate and final task-level accuracies remain closely aligned in DIL-HGRN and CIL-LSTM, indicating reduced forgetting.Shaded regions show ±1 standard deviation over multiple runs (seeds or subjects).Source: Figure 1 · p. 3

B. Feature Representation

Time-domain features improve DIL accuracy and stability, but their CIL benefit depends on replay. For independently windowed data, shuffling is essential for plasticity, while carried state offers no useful continuity.

B. Feature Representation

  • ∼40% higher DIL accuracy with time-domain features comes with tighter convergence and removal of raw-window training instability.Every TD DIL configuration has std ≤0.03, compared with high seed variance in weaker raw-feature configurations.Source: B. Feature Representation · p. 3
  • In CIL, time-domain features help replay but barely change naive and EWC results; increased forgetting can erase their initial accuracy gain.For LSTM EWC, forgetting rises from 12.3% to 28.5% accuracy drop.Source: B. Feature Representation · p. 3

C. Carrying State

  • Because windows are independent segments, recurrent models should reset hidden state per window; temporal-order training lets rest-dominated gradients drive collapse.Carried state across shuffled windows instead injects irrelevant noise into batch initial states.Source: C. Carrying State · p. 4
  • 94%–95% plasticity with stateless shuffling contrasts with 27%–45% when shuffling is off, below the 60% frozen threshold.Shuffled stateful training reaches 68%–76%, still below stateless training.Source: C. Carrying State · p. 4, Figure 2 · p. 4

D. EWC Regularization Quality

EWC is difficult to calibrate across the full task sequence and, when it remains plastic, still shows substantial forgetting. LSTM and HGRN perform comparably, while HGRN achieves similar accuracy at lower cost.

D. EWC Regularization Quality

  • λewc tuning does not transfer from a 6-subject proxy to the 27-subject sweep, and an excessively large value freezes the network.For TD features, λewc = 30 was the largest value preserving plasticity across the full sequence.Source: D. EWC Regularization Quality · p. 4
  • 0.162–0.285 accuracy drop persists in EWC configurations that avoid freezing, indicating limited constraint on parameter drift under strong DIL and CIL shifts.The authors identify plasticity below 0.60 on final tasks as freezing.Source: D. EWC Regularization Quality · p. 4

E. Effect of Recurrent Architecture

V. CONCLUSION

Across 27 subjects and three exercises, the study evaluates continual learning for sEMG gesture classification under domain- and class-incremental drift. Replay performs best, while experiments remain limited to offline NinaPro DB1 data.

V. CONCLUSION

  • 83.3% final accuracy with forgetting as low as 1.8% drop in accuracy is reported for experience replay, which outperforms EWC and naive baselines.The conclusion attributes long-term retention more to the continual-learning method than to recurrent architecture.Source: §V · p. 4
  • A 2×2 ablation attributes stateful collapse to batch ordering rather than recurrent state, consistent with windowing into independent sEMG segments.Source: §V · p. 4
  • Time-domain features improve per-task learning, but rehearsal is needed to retain that gain under class-incremental drift.Source: §V · p. 4
  • The experiments were conducted offline and only on NinaPro DB1; transfer to other datasets and experimental settings remains for future work.Source: §V · p. 4

Related papers

Loading 2610.34011v1…