pith. sign in

arxiv: 2406.19581 · v2 · submitted 2024-06-28 · 💻 cs.HC · cs.LG

Quasi-Linear ICA for Motor Unit Decomposition during Dynamic Contractions

Alexander Kenneth Clarke , Dimitrios Halatsis , Agnese Grison , Irene Mendez Guerra , Noura Ezaz-Nikpay , Pranav Mamidanna , Shihan Ma , Silvia Muceli
show 1 more author
Dario Farina
This is my paper

Pith reviewed 2026-05-24 00:06 UTC · model grok-4.3

classification 💻 cs.HC cs.LG
keywords electromyographymotor unit decompositionindependent component analysisdynamic contractionsquasi-linear ICAspike train recoveryhigh-density EMGnon-stationary mixing
0
0 comments X

The pith

Quasi-linear ICA recovers more motor neuron spike trains from EMG during dynamic contractions by placing a learned low-rank time-varying transformation before a static linear separator.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard ICA for EMG decomposition assumes the mixing from motor neurons to electrodes remains fixed over time, but movement deforms the volume conductor and makes the mixing time-varying. The paper places a trainable low-rank invertible transformation in front of a conventional linear ICA separator. The transformation is optimized with a stationarity loss on the recovered source while the separator uses an independence loss on the raw projection, with no gradients shared between the two steps. This keeps the source extraction step identical to classical linear ICA and therefore identifiable, while the transformation absorbs the non-stationarity. On a public benchmark of dynamic high-density EMG recordings that include ground-truth spike trains, the method recovers more units at higher accuracy than four existing adaptive ICA algorithms at every recall threshold.

Core claim

By preceding a static linear ICA separator with a learned low-rank time-varying invertible transformation and training the two modules with decoupled losses (independence on the uncompensated projection, stationarity on the recovered source), the formulation absorbs non-stationary volume-conductor distortion while preserving the identifiability guarantee of linear ICA. The closed-form inverse of the transformation further supports sequential peel-off with time-varying templates.

What carries the argument

Quasi-linear ICA: a low-rank time-varying invertible transformation followed by a static linear separator, trained with separate independence and stationarity losses and no gradient sharing.

If this is right

  • The method outperforms four adaptive ICA baselines at every recall threshold on the public dynamic high-density EMG benchmark.
  • More motor units are recovered at higher accuracy during dynamic contractions than with prior methods.
  • The source-extraction step reduces exactly to classical linear ICA and therefore inherits its identifiability guarantee.
  • The closed-form inverse of the transformation enables per-spike subtraction with a time-varying template in sequential peel-off decomposition.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could enable motor-neuron-driven control of prosthetics and exoskeletons during natural, non-isometric movements.
  • The same decoupled quasi-linear structure might be useful for other blind-source-separation problems that exhibit non-stationary mixing.
  • Real-time implementations would need to test whether the low-rank transformation can be updated fast enough for closed-loop neural interfaces.

Load-bearing premise

The non-stationary volume-conductor distortion admits a low-rank invertible time-varying transformation whose effect can be fully absorbed without destroying the statistical independence visible to the downstream linear ICA separator.

What would settle it

On the public benchmark of dynamic high-density EMG with ground-truth spike trains, the method fails to recover more units at higher accuracy than the four adaptive ICA baselines at one or more recall thresholds.

Figures

Figures reproduced from arXiv: 2406.19581 by Agnese Grison, Alexander Kenneth Clarke, Dario Farina, Dimitrios Halatsis, Irene Mendez Guerra, Noura Ezaz-Nikpay, Pranav Mamidanna, Shihan Ma, Silvia Muceli.

Figure 1
Figure 1. Figure 1: HarmonICA in operation. a After an initial pretraining phase, the linear separation vector is converged on a single source. However it lacks the capacity to account for the non-stationary latents in its source prediction. b Using alternating backpropagation, HarmonICA cycles between improving the separation vector and training a neural network to adapt the vector such that it accounts for non-stationaritie… view at source ↗
read the original abstract

Decomposing surface electromyography (EMG) into the spike trains of individual motor neurons is a long-standing inverse problem and a key step toward motor-neuron-driven neural interfaces such as prosthetics and exoskeletons. The standard approach, independent component analysis (ICA) of the multichannel signal, assumes that the mixing from neurons to electrodes is stationary in time. This assumption fails during movement, when volume-conductor deformation makes the mixing time-varying, and current decomposition algorithms are correspondingly restricted to isometric contractions. We introduce a quasi-linear ICA formulation in which a static linear separator is preceded by a learned, low-rank, time-varying invertible transformation. The separator is trained with an independence loss on the uncompensated projection, and the transformation with a stationarity loss on the recovered source. Gradients are not shared between the two, so the source-extraction step reduces to classical linear ICA and inherits its identifiability guarantee, while non-stationary distortion is absorbed by the transformation. The closed-form inverse of the transformation enables per-spike subtraction with a time-varying template during sequential peel-off. On a public benchmark of dynamic high-density EMG with ground-truth spike trains, the method outperforms four adaptive ICA baselines at every recall threshold, recovering more units at a higher accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces a quasi-linear ICA approach for decomposing surface EMG signals into motor unit spike trains during dynamic contractions, where the mixing is time-varying due to volume conductor changes. A low-rank time-varying invertible transformation compensates for non-stationarity, while a static separator is trained separately with an independence loss on the raw projection. The method claims to inherit classical ICA identifiability due to decoupled training and reports superior performance over four adaptive ICA baselines on a public dynamic high-density EMG benchmark with ground-truth spike trains.

Significance. If the empirical gains hold under scrutiny, the approach could extend reliable motor-unit decomposition beyond isometric conditions, supporting motor-neuron-driven interfaces during movement. The evaluation on a public benchmark with ground-truth spike trains is a clear strength, enabling reproducible comparison. The significance is tempered by the need to verify that the reported outperformance is robust and that the claimed inheritance of identifiability is correctly justified.

major comments (1)
  1. [Abstract (paragraph describing the quasi-linear formulation and the two decoupled losses)] Abstract (paragraph describing the quasi-linear formulation and the two decoupled losses): The claim that 'the source-extraction step reduces to classical linear ICA and inherits its identifiability guarantee' because 'gradients are not shared' is not supported by the construction. The independence loss is applied only to the uncompensated projection W x(t), yet x(t) = A(t) s(t) where the mixing matrix A(t) remains explicitly time-varying. Classical ICA identifiability theorems require stationary mixing; the effective mixing seen by the separator is W A(t), which is non-stationary. Decoupling the losses does not alter the data supplied to the independence objective or restore the theorem hypotheses.
minor comments (2)
  1. The abstract states outperformance 'at every recall threshold' but does not report error bars, number of independent runs, or statistical tests supporting the cross-method comparison.
  2. No ablation or sensitivity analysis is described for the rank of the time-varying transformation, which is the sole free parameter listed in the method.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying an overstatement in our description of the method's theoretical properties. We address the comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: Abstract (paragraph describing the quasi-linear formulation and the two decoupled losses): The claim that 'the source-extraction step reduces to classical linear ICA and inherits its identifiability guarantee' because 'gradients are not shared' is not supported by the construction. The independence loss is applied only to the uncompensated projection W x(t), yet x(t) = A(t) s(t) where the mixing matrix A(t) remains explicitly time-varying. Classical ICA identifiability theorems require stationary mixing; the effective mixing seen by the separator is W A(t), which is non-stationary. Decoupling the losses does not alter the data supplied to the independence objective or restore the theorem hypotheses.

    Authors: We agree with the referee's analysis. The decoupling of gradients does not restore the stationarity assumption required by classical ICA identifiability results, since the separator still receives the non-stationary effective mixing W A(t). The original claim therefore overstates the theoretical inheritance. We will revise the abstract and the corresponding paragraph in the methods section to remove the assertion that the source-extraction step inherits classical ICA identifiability guarantees. The revised text will instead describe the decoupled training as a practical design choice that lets the separator optimize an independence objective on the raw signals while the low-rank transformation absorbs time-varying distortion, with performance validated empirically on the ground-truth benchmark. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external classical ICA without reducing claims to self-defined inputs

full rationale

The paper's central construction decouples the independence loss on the static separator W from the stationarity loss on the time-varying transformation T(t), with the explicit statement that non-shared gradients cause the separator to reduce to classical linear ICA. This is an architectural assertion about training dynamics rather than a self-definitional loop (no quantity is defined in terms of itself) or a fitted input renamed as prediction. No self-citations are invoked as load-bearing uniqueness theorems, no ansatz is smuggled, and no known result is merely renamed. Performance is assessed via comparison to four external adaptive ICA baselines on a public benchmark containing independent ground-truth spike trains, rendering the derivation self-contained against external validation. Any question of whether the uncompensated projection satisfies classical ICA stationarity hypotheses is a matter of correctness, not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the modeling assumption that non-stationary mixing admits a low-rank invertible correction and on the empirical performance on one public benchmark; no free parameters or invented entities are explicitly quantified in the abstract.

free parameters (1)
  • rank of time-varying transformation
    The low-rank constraint on the learned transformation is a modeling choice whose specific value must be selected or tuned.
axioms (1)
  • domain assumption Volume-conductor deformation during dynamic contraction can be represented by a low-rank invertible time-varying transformation that leaves the independence structure of the sources intact for the downstream linear separator.
    Invoked to justify why the transformation can be learned separately without breaking the identifiability guarantee of classical ICA.

pith-pipeline@v0.9.0 · 5791 in / 1413 out tokens · 27576 ms · 2026-05-24T00:06:26.829590+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Blind identification of the spinal cord output in humans with high-density electrode ar rays implanted in muscles,

    S. Muceli, W. Poppendieck, A. Holobar, S. Gandevia, D. Li ebetanz, and D. Farina, “Blind identification of the spinal cord output in humans with high-density electrode ar rays implanted in muscles,” Science advances, vol. 8, no. 46, p. 5040, 2022

    work page 2022

  2. [2]

    A generic noninvasive neuromotor interface for human-computer interaction,

    CTRL Labs at Reality Labs, D. Sussillo, P. Kaifosh, and T. Reardon, “A generic noninvasive neuromotor interface for human-computer interaction,” bioRxiv, pp. 2024–02, 2024

    work page 2024

  3. [3]

    Neuro- musculoskeletal mapping for man-machine interfacing,

    T. Kapelner, M. Sartori, F. Negro, and D. Farina, “Neuro- musculoskeletal mapping for man-machine interfacing,” Scientific Reports, vol. 10, no. 1, p. 5834, 2020

    work page 2020

  4. [4]

    Principles of human movement augmentation and the challen ges in making it a reality,

    J. Eden, M. Bräcklein, J. Ibáñez, D. Y. Barsakcioglu, G. D i Pino, D. Farina, E. Burdet, and C. Mehring, “Principles of human movement augmentation and the challen ges in making it a reality,” Nature Communications, vol. 13, no. 1, p. 1345, 2022

    work page 2022

  5. [5]

    The neural resource allocation problem when enhancing hu man bodies with extra robotic limbs,

    G. Dominijanni, S. Shokur, G. Salvietti, S. Buehler, E. P almerini, S. Rossi, F. De Vignemont, A. d’A vella, T. R. Makin, D. Prattichizzo, et al. , “The neural resource allocation problem when enhancing hu man bodies with extra robotic limbs,” Nature Machine Intelligence , vol. 3, no. 10, pp. 850–860, 2021

    work page 2021

  6. [6]

    Spike sorting with kilosort4,

    M. Pachitariu, S. Sridhar, J. Pennington, and C. Stringe r, “Spike sorting with kilosort4,” Nature Methods, pp. 1–8, 2024

    work page 2024

  7. [7]

    A modular implementation to handle and benchmark drift correction for high-density ext racellular recordings,

    S. Garcia, C. Windolf, J. Boussard, B. Dichter, A. P. Bucc ino, and P. Yger, “A modular implementation to handle and benchmark drift correction for high-density ext racellular recordings,” Eneuro, vol. 11, no. 2, 2024

    work page 2024

  8. [8]

    Adaptive r eal-time identification of motor unit discharges from non-stationary high-density surface electromyographic s ignals,

    C. Chen, S. Ma, X. Sheng, D. Farina, and X. Zhu, “Adaptive r eal-time identification of motor unit discharges from non-stationary high-density surface electromyographic s ignals,” IEEE Transactions on Biomedical Engineering , vol. 67, no. 12, pp. 3501–3509, 2020

    work page 2020

  9. [9]

    On the prediction of motor unit filter changes in blind source separation of high- density surface electromyograms during dynamic muscle con tractions,

    M. Kramberger and A. Holobar, “On the prediction of motor unit filter changes in blind source separation of high- density surface electromyograms during dynamic muscle con tractions,” IEEE Access, vol. 9, pp. 103533–103540, 2021

    work page 2021

  10. [10]

    Dredge: robust motion correction for high-density extra cellular recordings across species,

    C. Windolf, H. Yu, A. C. Paulk, D. Meszéna, W. Muñoz, J. Bo ussard, R. Hardstone, I. Caprara, M. Ja- mali, Y. Kfir, et al. , “Dredge: robust motion correction for high-density extra cellular recordings across species,” BioRxiv, 2023

    work page 2023

  11. [11]

    Solving th e spike sorting problem with kilosort,

    M. Pachitariu, S. Sridhar, and C. Stringer, “Solving th e spike sorting problem with kilosort,” BioRxiv, 2023. 12

    work page 2023

  12. [12]

    Independent component analys is: algorithms and applications,

    A. Hyvärinen and E. Oja, “Independent component analys is: algorithms and applications,” Neural networks , vol. 13, no. 4-5, pp. 411–430, 2000

    work page 2000

  13. [13]

    Va riational autoencoders and nonlinear ica: A unifying framework,

    I. Khemakhem, D. Kingma, R. Monti, and A. Hyvarinen, “Va riational autoencoders and nonlinear ica: A unifying framework,” in International Conference on Artificial Intelligence and St atistics, pp. 2207–2217, PMLR, 2020

    work page 2020

  14. [14]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron , R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” Communications of the ACM , vol. 65, no. 1, pp. 99–106, 2021

    work page 2021

  15. [15]

    Multichannel blind source se paration using convolution kernel compensation,

    A. Holobar and D. Zazula, “Multichannel blind source se paration using convolution kernel compensation,” IEEE Transactions on Signal Processing , vol. 55, no. 9, pp. 4487–4496, 2007

    work page 2007

  16. [16]

    Accurate and representative decoding of the neural drive t o muscles in humans with multi-channel intramuscular thin-film electrodes,

    S. Muceli, W. Poppendieck, F. Negro, K. Yoshida, K. P. Ho ffmann, J. E. Butler, S. C. Gandevia, and D. Farina, “Accurate and representative decoding of the neural drive t o muscles in humans with multi-channel intramuscular thin-film electrodes,” The Journal of physiology , vol. 593, no. 17, pp. 3789–3804, 2015

    work page 2015

  17. [17]

    Emglab: an interactive emg decomposition program,

    K. C. McGill, Z. C. Lateva, and H. R. Marateb, “Emglab: an interactive emg decomposition program,” Journal of neuroscience methods , vol. 149, no. 2, pp. 121–133, 2005

    work page 2005

  18. [18]

    NeuroMotion: Open-source simulator with neuromechanical and deep network models to generate surface EMG signals during voluntary movement,

    S. Ma, I. Mendez Guerra, A. H. Caillet, J. Zhao, A. K. Clar ke, K. Maksymenko, S. Deslauriers-Gauthier, X. Sheng, X. Zhu, and D. Farina, “NeuroMotion: Open-source simulator with neuromechanical and deep network models to generate surface EMG signals during voluntary movement, ” BioRxiv, 2023

    work page 2023

  19. [19]

    A myoelectric digital twin for fast and realistic modelling in deep learning,

    K. Maksymenko, A. K. Clarke, I. Mendez Guerra, S. Deslau riers-Gauthier, and D. Farina, “A myoelectric digital twin for fast and realistic modelling in deep learning,” Nat. Commun. , vol. 14, no. 1, p. 1600, 2023

    work page 2023

  20. [20]

    A musculoskeletal model of the hand and wrist capable of simul ating functional tasks,

    D. C. McFarland, B. I. Binder-Markey, J. A. Nichols, S. J . Wohlman, M. De Bruin, and W. M. Murray, “A musculoskeletal model of the hand and wrist capable of simul ating functional tasks,” IEEE Trans. Biomed. Eng. , vol. 70, no. 5, pp. 1424–1435, 2023

    work page 2023

  21. [21]

    We arable neural interfaces: Real-time identification of motor neuron discharges in dynamic motor tasks,

    I. Mendez Guerra, D. Y. Barsakcioglu, and D. Farina, “We arable neural interfaces: Real-time identification of motor neuron discharges in dynamic motor tasks,” BioRxiv, 2024

    work page 2024

  22. [22]

    Multi-channel intramuscular and surface emg decomposition by convolutive blind source separation,

    F. Negro, S. Muceli, A. M. Castronovo, A. Holobar, and D. Farina, “Multi-channel intramuscular and surface emg decomposition by convolutive blind source separation,” Journal of neural engineering , vol. 13, no. 2, p. 026027, 2016. 13

    work page 2016