acceptodds
Under review as a conference paper at ICLR 2027

Q-LEAP: Quantum-Inspired Localized Evolution and Anchor Projection for Cross-Subject sEMG Recognition

Abstract

Cross-subject sEMG recognition is challenging due to substantial physiological variability, electrode shifts, and non-stationary signal drift across individuals. We propose Q-LEAP, a quantum-inspired framework that integrates localized evolution and anchor projection to learn transferable sEMG representations. Q-LEAP maps multichannel signals into normalized complex-valued states in Hilbert space, models cross-channel interactions through learnable localized evolution, and coherently fuses time- and frequency-domain representations via quantum-inspired interference. It further constructs fidelity-based anchor states and projects learned representations onto these anchors through projective measurement, yielding a stable reference space for cross-subject recognition. Extensive experiments on five NinaPro benchmarks demonstrate that Q-LEAP matches or exceeds twelve representative baselines in cross-session recognition, while providing substantial zero-shot and few-shot generalization advantages on unseen subjects and robust cross-database transfer between intact and amputee populations. These findings establish the effectiveness of quantum-inspired localized dynamics and fidelity geometry for robust physiological signal decoding under cross-subject variability.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
quantum-inspired neural networks, sEMG gesture recognition

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_pmpROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Q-LEAP for sEMG gesture recognition. A 1D-CNN produces magnitude and phase, which form a unit-norm complex vector. A learned unitary matrix, exp(−iH) with Hermitian H, mixes its components. The time-domain vector and its FFT are combined with two learned scalars. A loss pushes the squared amplitudes toward an exponential envelope. After pre-training, anchors are picked from training features by farthest point sampling under fidelity. Each sample is then represented by its fidelity to the anchors.

Experiments use NinaPro DB1 to DB5. The main table reports within-subject accuracy against twelve baselines. Further tables report zero-shot and few-shot results on unseen subjects and transfer between DB2 and DB3, each against one baseline. Ablations, a noise test and an efficiency comparison are included.

Critical strengths

The few-shot protocol is practical. Freezing the backbone and anchors and fitting only a linear layer on a few samples per gesture is efficient to run for a new user, and the model is small.

Critical weaknesses
  • The quantum framing does not correspond to what the model computes, and the physical claims are unsupported.

    • What the components are: The model is a complex-valued network with unit normalization, a learned unitary linear map, a fixed two-scalar combination of a vector with its FFT, and a layer of squared inner products with stored anchors followed by a linear classifier. Each has a well-known non-quantum form: complex-valued networks (Trabelsi et al., ICLR 2018), learned unitary transforms (Arjovsky et al., ICML 2016), and anchor-relative similarity representations (Moschella et al., ICLR 2023). None is cited, and the related-work section is only a single paragraph. The authors could have used appendix for related work if the page limitation was the concern.

    • “Spatial” localization: The localization loss (Eq. 12) and the IPR analysis are defined over the C = 256 latent dimensions produced by the CNN, not over the electrodes. Index distance |c − c₀| between latent dimensions has no spatial meaning, so statements that the model concentrates energy on “core active muscle groups” or “physically confines spatial crosstalk” do not follow. Nothing in the unitary map is disordered or local; the envelope is imposed by a loss on the output.

    • Phase and interference: The phase is the output of a convolutional branch, not a measured property of the signal. The fusion in Eq. 7 uses two global scalars, so it is a fixed weighted sum and cannot cancel noise adaptively. Adding a time-domain vector to its Fourier transform index by index is not motivated.

    • Untested advantages: Table 1 lists benefits such as robustness to electrode displacement and phase drift. No experiment varies electrode position. The only robustness test adds white Gaussian noise.

  • The experiments do not support the cross-subject claim, and the comparison is not controlled.

    • Main result is within-subject: Table 4 trains and tests on different repetitions of the same subjects. It is called “cross-session” with “strictly isolated” chronology, but train and test repetitions are interleaved. Table 2 also lists 10 sessions for DB2 and DB3 while the split uses repetitions up to 6.

    • Margins: In Table 4, Q-LEAP leads on three datasets by 0.06 to 0.41 points, within its own standard deviation, and trails on DB2 and DB4. Only Q-LEAP has standard deviations (three seeds).

    • Baselines are not re-run: The “−” entries are explained as baselines that “did not evaluate on that dataset”, so the numbers appear to be taken from other papers with possibly different windowing, preprocessing and splits. The grid search is said to use a validation set that is not defined.

    • Cross-subject evidence is thin: Tables 6 and 7 compare with a single method (CLDF), without standard deviations in Table 6, and without stating how CLDF’s numbers were obtained or adapted. No domain-adaptation or domain-generalization baseline trained in the same pipeline is included. Zero-shot accuracy on unseen subjects is 12% to 32%, which is better than CLDF but far from usable.

    • Ablation: The gains over the closest classical substitutes are about 2 to 3.5 points averaged over datasets, from single runs, while the seed standard deviation on DB2 alone is 2.4.

Questions for authors

Were any baselines re-implemented and run under your preprocessing, windowing and splits? If so, which, and what are their standard deviations?

How is the validation set constructed, and is it used for early stopping?

In Eq. 12, what gives the 256 latent dimensions a spatial ordering? Can the localization loss and IPR be computed over the physical electrodes instead?

How were the CLDF results in Tables 6 and 7 obtained? What does “zero-shot” mean operationally for each method?

How does a real-valued model with the same backbone, an orthogonal mixing layer, cosine similarity to FPS-selected anchors and the same linear-probe protocol perform on Tables 4 and 6?

Please report an experiment with electrode shift, since robustness to it is a stated advantage.

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

This paper proposes Q-LEAP, a quantum-inspired model for sEMG gesture recognition. It represents CNN features as normalized complex-valued vectors, uses unitary transformations generated by learnable Hermitian matrices for channel mixing, combines time-domain and FFT-based representations through weighted superposition, and adds a localization loss. After training, the backbone is frozen. Anchors are selected by farthest point sampling based on fidelity, samples are projected onto these anchors, and only the linear head is tuned for cross-subject calibration. Experiments on NinaPro DB1–5 cover cross-session evaluation, zero- and few-shot cross-subject transfer, and intact-to-amputee cross-database transfer.

Critical strengths

The experiments cover a broad range of settings, and the model is small, which is relevant to real-time applications such as prosthetics. This does not, however, outweigh the methodological and experimental issues.

Critical weaknesses
  1. The claimed physical mechanisms do not match the implementation. According to Table 10, the encoder maps 20 time points directly to 256 features, leaving no time axis for the FFT to capture the signal spectrum. The localization loss decays with distance between hidden-channel indices, which have no spatial order. Moreover, DB1 provides RMS-rectified envelopes without phase. The quantum advantages claimed in Table 1 are therefore unsupported; the model is essentially a complex-valued CNN with anchor features and a linear head.

  2. The main results are difficult to compare. Table 4 quotes baselines from papers with different windows and experimental protocols, and the appendix applies one-sample t-tests to published means. Even so, Q-LEAP leads slightly on DB1/3/5, loses to AT-Transformer on DB2/4.

Questions for authors
  1. Along which dimensions do the FFT, Hamiltonian, summation over τ in Eq. (8), and L_loc operate? Once the encoder has absorbed the time axis, how should “frequency domain” and “spatial localization” be understood?
  2. Removing Phase Θ also reduces performance on DB1 by 2.53 points, although DB1 signals have no phase. Could this gain simply come from the extra capacity of the complex-valued representation? Could you provide a real-valued version with a matched parameter count?
  3. Could you rerun several of the strongest baselines, such as AT-Transformer, CLDF, and sEMGXCM, under the same protocol, with consistent windows, preprocessing, and seeds?
  4. The paper states that hyperparameters were selected by grid search on a validation set, but Table 2 lists only training and test repetitions. Which repetitions were used for validation?
  5. Electrode displacement is one of the paper’s main motivations, but there is no corresponding experiment. Could you simulate electrode rotation on DB5 (Myo armband) and compare fidelity anchors with cosine anchors?
  6. The cross-subject and cross-database experiments compare against only CLDF. Were the CLDF results rerun or quoted? Could you add fine-tuning, DANN[1], or test-time adaptation methods such as Tent[2]?

[1]: Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-Adversarial Training of Neural Networks. Journal of Machine Learning Research 2016.

[2]: Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-Time Adaptation by Entropy Minimization. ICLR 2021.

Reviewer_EPsgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Q-LEAP, a quantum-inspired framework for cross-subject sEMG gesture recognition. It combines complex-valued amplitude-phase representations, learnable unitary evolution, time-frequency fusion, and fidelity-based anchor projection. The backbone is first trained on source data, after which anchors are selected from the learned representations and the backbone is frozen for lightweight classifier adaptation. The method is evaluated on five NinaPro datasets, with experiments covering cross-session recognition, zero-shot and few-shot cross-subject generalization, cross-database transfer, ablations, noise robustness, and computational efficiency. The results are competitive on most datasets, and the anchor-based adaptation strategy is potentially useful. However, some aspects of the evaluation protocol and the interpretation of the proposed mechanisms need further clarification. In particular, it remains unclear how much of the improvement comes from the specific quantum-inspired formulations rather than from the underlying representation learning and regularization techniques.

Critical strengths
  1. The few-shot results are encouraging. Evaluating both intact and amputee subjects, as well as transfer in both directions between DB2 and DB3, makes the study relevant to the intended application.
  2. Updating only the classifier offers a simple way to reduce adaptation cost. The component replacement experiments and measured CPU/GPU inference times also help assess the design and its computational requirements.
Critical weaknesses
  1. Several details of the cross-subject evaluation remain unclear, including the validation split for each held-out subject, the unit of a shot, and the separation of calibration and test windows. These details are needed to interpret and reproduce Table 6.
  2. Equation (12) regularizes latent coordinates after electrode pooling (Table 10), but their correspondence to physical electrode or muscle locations is not explained. The claimed connection between this regularizer and concentration around active muscle groups therefore needs further support.
  3. Although Section 5.1 states that the splits are consistent with prior work, Table 4 does not identify which baselines were rerun under the same preprocessing. The baseline tests in Table 9 also treat published means as fixed reference values, without accounting for their uncertainty.
  4. Table 5 includes component ablations and replacements, but corresponding results are not reported under the cross-subject zero-shot/few-shot setting of Table 6. The contribution of these components to the gains on unseen subjects remains unclear.
Questions for authors
  1. How is the validation set constructed in each leave-one-subject-out fold? After replacing the learned reference states with fixed anchors, which classifier is used for zero-shot evaluation, and on which data is it trained?
  2. Does one shot in Table 6 mean a single 200 ms window or a complete gesture repetition? How are calibration and test data separated, and can overlapping windows from the same recording interval appear in both sets?
  3. How do distances between the latent coordinates in Equation (12) relate to physical electrode or muscle locations? Which tensor and axis are used to compute IPR in Figure 2(c), and how does this analysis relate to the localization loss?
  4. Which Table 4 baselines were rerun using the reported preprocessing and splits, and does “session” in Table 2 refer to a separate recording or a gesture repetition? For the key baselines in Table 9, are subject-level results under the same protocol available to estimate the performance difference and its uncertainty?
  5. Could you evaluate the localization-loss and anchor ablations under the cross-subject protocol of Table 6, using the same calibration budget and classifier-training procedure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.