Response to Reviewer QgDP

Thank you for your professional review and valuable time. Your positive assessment is incredibly encouraging to our research team. We sincerely appreciate your thoughtful comments and would like to address each of your concerns as follows:

Broader Scientific Literature

Thank you for bringing this to our attention. We will discuss these two critical works (SCN, SpGesture) in the camera-ready version from the perspective of SNN development in sEMG gesture recognition.

Focal Loss vs Asymmetric Optimization

Thank you very much for bringing this to our attention. Although this was not our core contribution, we appreciate that providing these results would offer readers more valuable insights.

We have conducted additional tests on the Grabmyo:

• STET + Focal Loss: 89.42%
• STET + Asymmetric Optimization: 90.54%

Modeling in Non-Parallel vs Parallel

Thank you for your suggestion. We compared the Focal Transformer (non-parallel) with our method.

In our inference speed experiments, we used a 4-stage Focal Transformer, while STET used two long-term layers plus two short-term layers.

Concerns about Hard Samples

We appreciate your insightful question. Below is our detailed response:

Definition of Hard Samples

Hard Positives:

• Definition: Positive samples (ground-truth $y_{i,j} = 1$) with low predicted probabilities $\widehat{y}_{i,j}$
• Weighting Mechanism: The term $(1 - \widehat{y}_{i,j} )^{\gamma^+}$ assigns higher weights to samples where the model lacks confidence.

Proportions and Dynamics

Classic Mask Method vs Our Method

Thank you for your suggestion! We actually included related experiments in Appendix A.3. Classical mask uses a standard pretraining with a general mask, employing a standard masking approach.

We understand your concern and will highlight this experiment more prominently in the revised paper.

Concerns about Noise Reduction

We appreciate your thoughtful comment. We would like to clarify this relationship with additional insights and reasoning.

The connection between short-term feature modeling and noise resilience is rooted in the fundamental characteristics of sEMG signals and common noise patterns.

1. Electrode-skin interface noise: Movement artifacts and momentary changes in skin-electrode contact typically manifest as brief bursts of interference rather than consistent long-term corruption [1, 2].
2. Environmental electromagnetic interference: Such interference usually occurs in short, sporadic patterns rather than continuously affecting the entire signal [3].

It's worth noting that we're not focusing on constant background noise (such as power line interference at 50/60Hz), as these can be effectively removed through conventional filtering techniques. Our primary concern is with transient, unpredictable noise patterns that are more challenging to address through traditional methods and represent common challenges in real-world applications.

By adopting a sliding window attention mechanism, our STEM concentrates on local temporal windows to isolate noise within affected segments, preventing it from spreading across the entire sequence. This method outperforms the long-term decoder under various noise types (as shown in Figure 3) and offers multiple sampling perspectives: overlapping windows that partially include noisy regions can still capture valuable information, thus reinforcing resilience.

We believe visualizing the noise distribution is a worthwhile idea; however, precisely pinpointing the noise remains challenging, so we cannot present such a visualization yet, but we will continue to work toward it.

References

[1] Surface electromyography: physiology, engineering, and applications. IEEE Press.

[2] Filtering the surface EMG signal: Movement artifact and baseline noise contamination. Journal of Biomechanics

[3] Sampling, noise-reduction and amplitude estimation issues in surface electromyography. Journal of Electromyography and Kinesiology

We sincerely appreciate all your valuable suggestions and will incorporate them into our revised manuscript. Your positive assessment is greatly encouraging to our research team.

Response to Reviewer ybGK

We sincerely appreciate your thorough review and valuable feedback on our manuscript. We are grateful for your recognition of our paper's strengths, including the clear writing style, effective visualizations, strong performance , comprehensive ablation studies, and the novelty of our approach. Your positive assessment is greatly encouraging to our research team.

We have carefully considered your concerns and would like to address them as follows:

Regarding Dataset Evaluation

We apologize for causing this misunderstanding. While we did indeed evaluate our method on both GRABMyo (results presented in Table 1) and Ninapro DB2 (results presented in Table 4), we acknowledge that this was not sufficiently clear in our presentation. We will revise the manuscript to clearly indicate the comprehensive nature of our evaluation.

Furthermore, to provide even more robust evaluation, we have added results on an additional dataset, Ninapro DB5, as shown in the table below:

This enhancement brings our total to three distinct datasets, offering a more comprehensive assessment of our method's performance across varied conditions.

Regarding Comparison Methods

We thank you for highlighting the need for more recent comparison methods. Following your suggestion, we have expanded our comparisons to include two state-of-the-art methods published in late 2024:

These additional comparisons demonstrate that our method maintains its performance advantage even against the most recent approaches in the field.

Our Main Contributions

To summarize, we would like to reiterate the key contributions of our work:

1. Introduction of sEMG Signal Masking: We propose a novel self-supervised pretraining technique using sEMG Signal Masking (Sensor-wise and Contiguous Masked Segments following a Geometric Distribution) to leverage the inherent variability in sEMG data.

2. STEM Module for Enhanced Noise Resilience: From the perspective of improving short-term feature representation, we propose STEM, an adaptive and noise-resistant module. Integrating STEM into various neural networks has demonstrated significant performance gains.

3. Improved Noise Resilience: Extensive experiments confirm that our overall design substantially enhances the noise resilience of models for sEMG data.

We sincerely appreciate your insightful comments, which have helped us improve the quality and comprehensiveness of our manuscript.

Response to Reviewer AUAg

We sincerely thank you for your thorough review and insightful feedback on our paper. We particularly appreciate your recognition of our well-motivated problem statement and the novel approach of using short-term feature extraction. Your professional suggestions have been invaluable in improving our work.

Regarding N-fold Cross-validation

Thank you for this excellent suggestion. While we initially followed standard evaluation methods in the field of sEMG-based gesture recognition (as seen in [1], [2], [3]) and used validation sets to prevent overfitting (detailed in the appendix's dataset section), we also believe that your suggestion is extremely valuable.

As per your recommendation, we have conducted a 6-fold cross-validation comparing our method with the second-best performer (Informer):

Statistical significance was confirmed with a t-test (p < 0.001), further supporting the robustness of our approach.

Concern about error bars

We apologize for any confusion. We would like to clarify that we have indeed included the mean and standard deviation in Table 1 of our original manuscript, as well as in our newly added n-fold experiments. We appreciate your suggestion and recognize the importance of clearly presenting statistical significance when comparing our approach with the baselines. Based on your feedback, we will revise the manuscript to make these error metrics more prominent and ensure they are easily noticeable throughout our experimental results section.

Regarding Inference Latency

We completely agree with your insightful point about the importance of latency evaluation for downstream applications. While this information was included in our appendix under "Inference Performance and Parameter Comparison of Models," we have expanded this section with additional comparative analysis:

Our STEM module adds minimal computational overhead (only 0.1ms on GPU and 2ms on CPU) while delivering superior accuracy with competitive inference speed.

Conclusion

Once again, we are grateful for your expert review and constructive feedback. We hope the additional evaluations address the concerns you raised while further validating the effectiveness of our approach. Your positive assessment is encouraging to our research team.

[1] Multi-attention feature fusion network for accurate estimation of finger kinematics from surface electromyographic signals IEEE Transactions on Human-Machine Systems

[2] Cross-Subject Lifelong Learning for Continuous Estimation from Surface Electromyographic Signal IEEE Transactions on Neural Systems and Rehabilitation Engineering

[3] A CNN-attention network for continuous estimation of finger kinematics from surface electromyography IEEE Robotics and Automation Letters