acceptodds
Under review as a conference paper at ICLR 2027

Whiten, Then Mix: A Two-Part Adapter for Few-Shot sEMG Personalization

Abstract

Personalizing a gesture decoder to a new sEMG band wearer is conventionally evaluated by the gain from that wearer's own calibration data, but this quantity does not distinguish transferable improvement from wearer-specific benefit. A correction that closes a between-wearer gap may also benefit other wearers. We introduce a cross-wearer substitution diagnostic: each correction is fitted on a different wearer and applied to the recipient. Using this diagnostic, the two axes show different transfer behavior: cross-wearer amplitude corrections transfer at every measured depth, whereas cross-wearer electrode-position corrections widen the gap. Under an invertible linear mixing model with identity source covariance, the population spatial covariance identifies the symmetric polar factor but not the orthogonal one. This distinction motivates our two-part few-shot adapter: Euclidean Alignment performs label-free, wearer-specific whitening, while ElectrodeMix adapts the residual on the electrode axis with learned parameters beyond the classifier head. Meta-pretraining uses a restricted inner loop to support this adaptation. Across the evaluated sEMG archives and both backbones, the method achieves the highest mean accuracy in every 10-shot cross-dataset setting. On EPN612-to-UCI-EMG transfer with ViM-Tiny, it exceeds the frozen probe for all 34 wearers, by 6.9 percentage points on average, using 1,024 adapter parameters versus 6.8M for full fine-tuning. This diagnostic requires only an additional cross-wearer fit and distinguishes transferable improvement from the advantage of recipient-specific fitting.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
surface electromyography (sEMG), few-shot personalization, meta-learning, domain alignment, wearable biosignals, parameter-efficient adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9CY5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose a method to personalize sEMG gesture decoders with few-shot samples. They find that corrections fitted to the target user have the largest advantage over other users' corrections at the tokenizer. Based on this finding, they propose ElectrodeMix, an adapter that mixes features across electrodes using different matrices for different feature groups. The method first uses EA to normalize the input, then ElectrodeMix uses labeled samples to correct the remaining differences while keeping the encoder fixed. The authors also use self-supervised meta-learning to prepare the model for few-shot adaptation.

Critical strengths
  • The design makes good use of sEMG-specific inductive biases. It considers differences in signal amplitude and electrode placement, and uses the depth analysis to motivate placing ElectrodeMix at the tokenizer stage.
  • The two-stage design is useful. In the 10-shot experiments, adding ElectrodeMix to EA consistently improves accuracy across multiple datasets and both backbones. Comparisons with other adapters after the same EA step further support its effectiveness.
Critical weaknesses
  • The need for the proposed personalization metric is unclear. When a model has already been trained on a large amount of data, I'm not sure why it is necessary to compare a correction learned from the user's own data with one learned from others'. For in-distribution data, adaptation uses small additional samples, so I would expect its average benefit to decrease as the amount of pre-training data increases. Although adaptation can move the model in different directions when facing out-of-distribution samples, these updates may depend strongly on the selected user. Given the strong user-specific differences in sEMG, it is unclear whether this variability has been sufficiently considered. While I understand how the proposed metric differs from the usual adaptation gain, I am not fully convinced that it is appropriate in this setting or provides a reliable measure of personalization performance.

  • Connected to the first point, donor sensitivity is partly tested. The main diagnostic uses one donor per recipient. The appendix tests 12 assignments for part of the depth analysis, but does not show that the largest target-versus-donor advantage consistently occurs at the tokenizer. More donor assignments across the full depth analysis and task-level experiments would make the conclusions more convincing.

  • The novelty of the method is unclear. ElectrodeMix appears to be the main new component. Although combining it with EA improves performance, EA is an existing method. The meta-learning setup also seems to be a design choice for effective few-shot adaptation. Previous work has already defined meta-learning tasks for domain adaptation, and choosing which parameters to update or freeze in the inner and outer loops does not seem new. The paper should explain more clearly what is new beyond combining these components.

  • The distinction from prior meta-learning methods is unclear. SelfReplay [1] already combines domain tasks with meta self supervised learning for user adaptation. DARLING [2] also uses domain information in contrastive pre-training to improve generalization to unseen domains. Restricting which parameters are updated in the inner loop is also established in methods such as ANIL. The paper should explain what is gained by adapting ElectrodeMix and normalization parameters, and compare with SelfReplay under matched data and adaptation budgets.

  • Experiments section lacks tuning details. The paper does not clearly explain how the shared hyperparameters were selected or how sensitive the results are to these choices.

[1] Yoon, Hyungjun, et al. "Selfreplay: Adapting self-supervised sensory models via adaptive meta-task replay." Proceedings of the 23rd ACM Conference on Embedded Networked Sensor Systems. 2025.

[2] Zhang, Xingxuan, et al. "Towards unsupervised domain generalization." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.

Questions for authors

I have discussed the points in the weaknesses. Please refer to that section.

Reviewer_VC7LOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a new approach to cross-subject adaptation for hand gesture recognition from EMG. It states the hypothesis that the variation between users can be decomposed into 2 factors: EMG amplitude and electrode rotation. It assess this hypothesis by proposing a meta learning method which aligns samples from different subject along these two axis before performing fine-tuning. Multiple public datasets are used to evaluate the method agains other techniques.

Critical strengths

The hypothesis of two axis decomposition to explain cross-subject variation and attempt to correct it in an unsupervised manner is very interesting and could lead the way to more explainable and robust methods for cross-subject hand gesture recognition.

Critical weaknesses
  • The main comment is that the paper is very difficult to understand, even for someone who is deeply involved in the topic of the paper. Even understanding if the analysis verifies the two-axis decomposition is not trivial.
  • The table 1 reports results on the EPN612 dataset, with the best accuracy being 73.3%. Other papers have performed cross-subject unsupervised adaptation on this dataset, reaching far higher accuracy. This suggests that the proposed method is not really efficient.
Questions for authors

None

Reviewer_5ZFzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. this paper introduces cross-wearer adapter substitution to distinguish transferable improvements from the additional benefit of recipient-specific calibration. It combines Euclidean Alignment with a compact electrode-mixing adapter, supported by restricted meta-pretraining.

  2. evaluation covers three sEMG archives and two backbones. On EPN612-to-UCI transfer with ViM-Tiny, the method reaches 86.93% accuracy versus 80.03% for the linear probe, using 1,024 adapter parameters beyond the classifier.

Critical strengths
  1. useful evaluation diagnostic. Comparing recipient-fitted and donor-fitted adapters, while refitting the classifier on identical recipient data, directly tests the value of personalization. Table 3 demonstrates a recipient-specific advantage on both backbones.

  2. compact adaptation with consistent benefits. The method improves across multiple datasets and architectures. The comparisons with EA alone and EA combined with other adapters in Tables 21–22 strengthen the evidence that residual adaptation contributes beyond whitening.

Critical weaknesses
  1. baseline optimization limits the comparative conclusions. All methods use a checkpoint pretrained specifically to support ElectrodeMix, and gradient-based methods share an untuned learning rate of 0.05 (Appendix A.3). Without method-specific validation tuning and broader pretraining controls, the reported rankings do not establish robust superiority.

  2. the fractional-shift diagnostic is not information-preserving as written. For eight electrodes, taking the real part in Eq. 21 scales the Nyquist component by (cos⁡(πδ\cos(\pi\delta)); a half-electrode shift removes it entirely. This can confound positional correction with amplitude attenuation. The integer-shift control in Table 13 does not test this issue.

Questions for authors
  1. do the gains persist with validation-tuned learning rates and a shared encoder pretrained without favoring ElectrodeMix, particularly against EA + SAL?

  2. how is the Nyquist component handled in Eq. 21, and do the position-transfer conclusions remain after controlling for its attenuation?

  3. what is the exact SAL implementation? The reported 72 parameters differ from the original SAL’s seven spatial-transformation parameters, excluding optional channel biases. Please explain the adaptation to ring inputs.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.