- 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.