Pith. sign in

REVIEW 4 major objections 4 minor 30 references

Lightweight test-time adaptation — causal adaptive batch normalization, replay-regularized GMM alignment, and meta-learning — raises inter-session EMG gesture accuracy on NinaPro DB6 from 56.6% to 80.3%, enabling robust wearable decoding wi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 12:04 UTC pith:ZXNU2LDA

load-bearing objection Solid empirical comparison of three TTA strategies for EMG, but the abstract's '82% self-supervised' claim is unsupported—the real unsupervised ceiling is ~70%. the 4 major comments →

arxiv 2601.04181 v2 pith:ZXNU2LDA submitted 2026-01-07 cs.LG cs.HC

Lightweight Test-Time Adaptation for EMG-Based Gesture Recognition

Nia Touko , Matthew O A Ellis , Cristiano Capone , Alessio Burrello , Elisa Donati , Luca Manneschi This is my paper
classification cs.LG cs.HC
keywords test-time adaptationEMG gesture recognitioncausal batch normalizationGaussian mixture model alignmentmeta-learningexperience replayNinaPro DB6inter-session generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the inter-session performance drop in EMG-based gesture recognition, caused by electrode shifts, fatigue, and posture changes, can be substantially repaired at deployment time using lightweight, causal test-time adaptation instead of retraining or large calibration sets. On the NinaPro DB6 multi-session dataset, a temporal convolutional network baseline that reaches 85% intra-session accuracy falls to 56.6% on unseen sessions; the paper's three strategies raise that to 68.9% (adaptive batch normalization), 69.8% (replay-regularized GMM alignment), and 80.3% (meta-learning with a few labeled calibration gestures). The key message is that robustness need not come from exhaustive training-time coverage of all operating conditions; modest on-device statistical alignment can correct session-specific drift. This matters because wearables and prosthetics cannot afford heavy retraining pipelines, yet their real-world utility depends on cross-session reliability.

Core claim

The paper's central claim is that three complementary, deployment-ready adaptation strategies can bridge most of the inter-session accuracy gap in EMG decoding while keeping computational overhead small. First, an online causal variant of batch normalization blends source EMA statistics with Welford-updated target statistics using an adaptive schedule, yielding 68.93% inter-session accuracy with negligible extra compute. Second, a replay-regularized statistical alignment method updates only a low-rank LoRA parameter subset by matching feature moments or a Gaussian Mixture Model to the target stream, with Dark Experience Replay anchoring source exemplars; this reaches 69.78% and, crucially, p

What carries the argument

The load-bearing mechanism is test-time statistical alignment of feature distributions, implemented in three variants sharing a single backbone. The first is causal adaptive batch normalization: instead of static population statistics, the model computes online per-channel mean and variance over the target prefix via Welford's method and blends them with source EMAs using a time-dependent coefficient alpha(t) = min(1, beta*t/N). The second is low-rank adaptation (LoRA) of a single intermediate TCN block, driven by moment matching or sliced-Wasserstein GMM alignment, regularized by a Dark Experience Replay buffer that preserves source logits. The third is gradient-based meta-learning (MAML) t

Load-bearing premise

The target prefix used for adaptation must contain enough gesture variety over a sufficiently long window (tens of seconds) for the estimated statistics to represent the whole session; the paper itself shows that with only 2–4 repetitions, aggressive adaptation can drop accuracy below the 56.6% baseline.

What would settle it

Run the same adaptive-BN and GMM-alignment protocols on NinaPro DB6 but construct the adaptation prefix as 2–4 repetitions of a single gesture (or a mostly-rest stream); if inter-session accuracy on the remaining data does not beat the 56.6% non-adapted baseline, the moment-representativeness assumption fails. Alternatively, apply the methods to a dataset where drift is known to be non-moment-based (e.g., electrode shear causing nonlinear spectral changes); if no method exceeds baseline, the statistical-alignment model of drift is incomplete.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Unsupervised test-time adaptation alone (adaptive BN or GMM alignment) can recover more than half of the inter-session accuracy gap, so prosthetic controllers could recalibrate during normal use without user-labeled calibration.
  • Replay-regularized alignment preserves or improves accuracy on non-adapted sessions, suggesting that continual-learning principles can stabilize any non-stationary physiological decoder at deployment.
  • Meta-learning with only a few labeled repetitions approaches fully supervised accuracy (80.3% vs. the non-adapted 56.6%), implying that brief user calibration phases can be dramatically shortened.
  • The compute overhead is small enough for edge deployment: adaptive BN adds no MMACs, GMM alignment adds roughly 1.2 MMAC per step (about 10 MMAC over a full gesture), and memory stays below a few hundred kilobytes.
  • Because the evaluation is causal and includes transient and rest periods, the reported gains are more likely to transfer to real streaming use than results from segment-filtered benchmarks.

Where Pith is reading between the lines

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

  • If the moment-shift model of EMG drift holds beyond NinaPro DB6, the same three strategies could be applied to other biosignal domains (e.g., EEG or ECG) where session-to-session distribution shift is a known barrier; the paper does not test this.
  • The discrepancy between source EMA statistics and online target estimates could serve as an inexpensive out-of-distribution detector, toggling adaptation strength or triggering a user calibration — a concrete extension the paper only sketches.
  • The paper's proposed hybrid pipeline (continuous BN adaptation plus intermittent replay-regularized updates) is left as future work; a direct test would be to implement it on-device and measure accuracy drift over many days of real prosthesis use.
  • A layer-wise ablation of which statistics matter most (which layers, which moments) would clarify when the GMM objective's extra fidelity is worth its compute over simple moment matching; the paper reports aggregate comparisons only.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a lightweight test-time adaptation framework for EMG-based gesture recognition, built on a TCN backbone and evaluated on the multi-session NinaPro DB6 dataset. Three adaptation strategies are studied: causal adaptive batch normalization (AdaBN), replay-regularized statistical alignment with GMM/moment matching via LoRA, and MAML-based few-shot meta-learning. The authors report that the non-adapted TCN reaches 56.61% inter-session accuracy, while AdaBN reaches 68.93%, GMM alignment 69.78%, and supervised meta-learning 80.30% (Table I). The abstract claims that self-supervised test-time adaptation methods reach up to 82% inter-session accuracy. The paper also reports resource usage in terms of parameters, memory, and MMAC.

Significance. If the reported results hold, the paper makes a useful empirical contribution: it shows that lightweight, test-time adaptation can substantially narrow the inter-session accuracy gap for EMG gesture recognition without retraining or large calibration sets. The strengths of the study are the strict causal temporal split (first five sessions for training, last five for testing), inclusion of transient and rest periods in the evaluation, reporting of active-gesture accuracy excluding the rest class, and the use of an external public benchmark. The authors also provide an explicit resource-usage accounting. However, the central claim as stated in the abstract is not supported by the reported results, and several experimental details needed for reproducibility are missing.

major comments (4)
  1. [Abstract vs. Table I] The abstract states: 'our self-supervised test-time adaptation methods reach up to 82% inter-session accuracy.' This figure does not appear anywhere in the paper. Table I reports AdaBN at 68.93%, GMM at 69.78%, and supervised Meta at 80.30%. The only way to reach ~82% would be an unreported configuration or a per-subject maximum. The self-supervised methods top out at ~70%, and Section III-C explicitly labels meta-learning as supervised adaptation. This is a load-bearing inconsistency: changing 'up to 82%' to the actual maximum materially changes the contribution claimed. The abstract must be corrected to attribute 82% to supervised meta-learning if such a result exists, or to report the true maximum for self-supervised methods (~70%).
  2. [Table I] Table I compares the proposed active-gesture accuracy (rest class excluded) with prior work marked with '*', which 'includes rest state classification in the accuracy.' Because rest-state classification is typically easier and more prevalent, comparing active-gesture accuracy to rest-inclusive accuracy can inflate the apparent improvement. The footnote acknowledges this difference, but the comparison in Table I and the text's 'significantly improving upon prior approaches' claim rest on this mixed basis. The authors should report comparable metrics, either by re-evaluating prior methods on the same protocol or by clearly separating the comparisons and qualifying the claim.
  3. [Section II-D2, Eq. (16) and Section II-D3] Key hyperparameters for the central experiments are not reported. In Eq. (16), the loss weights λ_μ, λ_Σ, α, and β are not specified. The GMM component count, LoRA rank (stated as rank-4 in Section III-B but not tied to a full configuration), and MAML outer-loop hyperparameters (e.g., meta-learning rate, number of outer steps) are missing. Without these, the results are not reproducible. The authors should provide a complete hyperparameter table or a supplement.
  4. [Section II-C and Section IV] The 'unbalanced sequences' setting in Section II-C is a random sequence of gestures, but the paper's own Figure 2a and Section IV show that with only 2–4 gesture repetitions, aggressive BN adaptation can degrade below baseline and recovery requires tens of seconds. In real daily use, the first minutes of a target stream may be dominated by one gesture or rest, which would violate the representativeness assumption. The paper acknowledges this as a limitation, but it is central to the claim of robust 'plug-and-play' deployment. A quantitative evaluation of the proposed methods under class-imbalanced or highly skewed streaming conditions would strengthen the claim; at minimum, the limitation should be stated more prominently in the abstract and conclusions.
minor comments (4)
  1. [Eq. (10)] The schedule α(t)=min(1, β t/N) reaches β at t=N unless β≥1. Clarify whether β is intended to be a fraction or whether the schedule should be α(t)=min(1, β t/N) with β>1 for full adaptation.
  2. [Fig. 4] The x-axis '# Steps' for meta-learning is not clearly defined. The text says 'The same data points are used for the different steps of the adaptation, as in few-shot learning settings,' but the inset in Fig. 4a seems to show first- vs second-order MAML at different steps. Please clarify what is plotted and how adaptation step count relates to the reported accuracy.
  3. [Table I] The MMAC values are reported per time step with sequence length 1 and batch size 32, while adaptation costs are given in parentheses as totals over 4 or 21 gestures. These are different units; please make the reporting consistent or add a clear explanation.
  4. [References] Reference [21] duplicates the Ioffe reference [12] with an incomplete citation. Also, some references are given as arXiv preprints; ensure the final version includes journal/conference details where available.

Circularity Check

0 steps flagged

No circularity: the adaptation results come from held-out external-benchmark experiments, not from fitting/definitional equivalence.

full rationale

The paper's derivation chain is empirical rather than definitional. The TCN baseline is trained on the first five NinaPro DB6 sessions (Sec. II-C) and tested on the held-out later five sessions; Table I reports measured inter-session accuracies (56.61 baseline, 68.93 AdaBN, 69.78 GMM, 80.30 Meta). The adaptation objectives (Eqs. 8-16) are standard test-time alignment losses: AdaBN blends source EMA statistics with target online statistics, and statistical alignment minimizes moment/GMM discrepancies in a frozen-backbone LoRA subspace with DER replay. None of these losses is constructed to force a particular held-out accuracy, and no fitted parameter is later renamed as a prediction. Meta-learning (Sec. II-D3) is explicitly described as supervised few-shot calibration, providing a reference point rather than a self-supervised claim. Self-citations (e.g., [7], [24]) appear only as background/comparison, not as load-bearing justification; the central numbers are from the current experiments on an external public benchmark. The abstract's 'up to 82%' statement is not supported by Table I (closest reported value is 80.30% supervised meta-learning), but that is an internal reporting/consistency issue, not an equation-level circularity. Section IV's limitation about aggressive BN degrading with 2-4 repetitions is an honest caveat, not a circular step. Therefore score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claims rest on the NinaPro DB6 benchmark, the causal split, and on the assumption that cross-session EMG drift manifests as feature-distribution shift. The free parameters are dominated by adaptation hyperparameters (BN schedule, alpha, LoRA rank, loss weights, GMM components, MAML steps), most scanned or unspecified. No new physical or architectural entities are introduced.

free parameters (6)
  • BN adaptation schedule beta = 0.01, 0.02, 0.05, 0.2 (explored)
    Eq. 10 sets alpha(t)=min(1, beta/N*t); no principled rule selects beta, and Fig. 2d,e shows no single beta is best across subjects.
  • alpha blending coefficient = 0 to 0.88 (Fig 2a,b)
    In the fixed-alpha ablation, the optimal alpha depends on the number of gesture repetitions; it is scanned rather than derived.
  • LoRA rank and placement = rank 4, single intermediate block
    Section III-B: 'injecting a rank-4 LoRA at a single intermediate block is sufficient' was chosen after preliminary exploration; only this configuration is reported.
  • Statistical alignment loss weights = not specified
    Eq. 16 uses alpha*L_align + beta*L_er, and Eq. 12-13 use lambda_mu and lambda_Sigma; values are not reported.
  • GMM component count = not specified
    Section II-D2 fits a GMM to source features for Sliced Wasserstein matching; the number of components is not given.
  • MAML inner-loop steps = K = 4
    Section III-C states 'four inner-loop gradient steps during meta-training'; no sensitivity analysis for K is provided.
axioms (4)
  • domain assumption The NinaPro DB6 temporal split (first five sessions source, last five target) accurately models real deployment drift.
    Section II-C; the entire evaluation rests on this split being representative of long-term wearable/prosthetic use.
  • domain assumption Cross-session EMG drift is mainly a feature-distribution shift that can be corrected by aligning BN statistics, low-order moments, or GMMs.
    Section II-D; if the shift also changes the optimal decision boundary in a way not captured by distribution matching, the methods would fail.
  • domain assumption The causally observed target prefix is representative of the remainder of the target session.
    Section II-C and Fig. 2a; with very few or unbalanced repetitions, target statistics are poor and adaptation can degrade below baseline.
  • domain assumption Low-rank LoRA updates and a small replay buffer can express and stabilize the needed session adaptation.
    Sections II-D2 and III-B; the capacity of LoRA is chosen empirically, not guaranteed by construction.

pith-pipeline@v1.3.0-alltime-deepseek · 14069 in / 15265 out tokens · 159009 ms · 2026-08-03T12:04:48.820660+00:00 · methodology

0 comments
read the original abstract

Reliable long-term decoding of gestures from surface electromyography (EMG) is hindered by signal drift caused by electrode displacement, muscle fatigue, and/or posture changes. Although modern models achieve high intra-session accuracy, their performance often degrades substantially across recording sessions. Existing approaches to mitigate this problem typically rely on large training datasets or computationally intensive pipelines that are unsuitable for energy-efficient wearable devices. We propose a lightweight test-time adaptation framework for EMG decoding. The framework includes three complementary adaptation strategies: (i) causal adaptive batch normalization for online statistical alignment, (ii) Gaussian Mixture Model alignment with experience replay to mitigate forgetting, and (iii) meta-learning for rapid few-shot calibration. We evaluate these methods on the multi-session NinaPro DB6 dataset. All approaches substantially improve inter-session robustness relative to a non-adaptive baseline while maintaining low computational overhead. Replay-regularized statistical alignment provides the most stable adaptation under limited data, while meta-learning achieves the highest accuracy when sparse calibration labels are available. Overall, our self-supervised test-time adaptation methods reach up to 82% inter-session accuracy, significantly improving upon prior approaches while maintaining resource-efficient operation. These results demonstrate that lightweight test-time adaptation can enable robust, long-term EMG decoding for wearable or prosthetic applications.

Figures

Figures reproduced from arXiv: 2601.04181 by Alessio Burrello, Cristiano Capone, Elisa Donati, Luca Manneschi, Matthew O A Ellis, Nia Touko.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework for inter-session adaptation in surface [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance of the causal variant of batch normalization. (a,b) Batch-normalization adaptation: evolution of error rate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance of the statistical alignment strategy across sessions. (a,b) GMM statistical loss with DER alignment scheme’s [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance of Meta-learning across sessions. (a,b) illustrate the evolution of the error rate of the current session (a) and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative accuracy improvement of each test-time adaptation approach over the baseline with respect to the data buffer [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

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

30 extracted references · 6 linked inside Pith

  1. [1]

    EMG Acquisition and Processing for Hand Movement Decoding on Embedded Systems: State of the Art and Challenges

    Simone Benatti et al. “EMG Acquisition and Processing for Hand Movement Decoding on Embedded Systems: State of the Art and Challenges”.Proceedings of the IEEE(2025)

    2025

  2. [2]

    Exploring Arm Posture and Temporal Variability in Myoelectric Hand Gesture Recognition

    B. Milosevic, E. Farella, and S. Benatti. “Exploring Arm Posture and Temporal Variability in Myoelectric Hand Gesture Recognition”.2018 7th IEEE International Conference on Biomedical Robotics and Biomechatronics (Biorob). 2018

    2018

  3. [3]

    The effects of electrode size and orientation on the sensitivity of myoelectric pattern recognition systems to electrode shift

    Aaron J Young, Levi J Hargrove, and Todd A Kuiken. “The effects of electrode size and orientation on the sensitivity of myoelectric pattern recognition systems to electrode shift”.IEEE transactions on biomedical engineering58 (2011)

    2011

  4. [4]

    Fluctuating EMG signals: Investi- gating long-term effects of pattern matching algorithms

    P. Kaufmann et al. “Fluctuating EMG signals: Investi- gating long-term effects of pattern matching algorithms”. 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology. IEEE. 2010

    2010

  5. [5]

    Deep Learning for Electromyo- graphic Hand Gesture Signal Classification Using Trans- fer Learning

    U. Côté-Allard et al. “Deep Learning for Electromyo- graphic Hand Gesture Signal Classification Using Trans- fer Learning”.IEEE Transactions on Neural Systems and Rehabilitation Engineering27 (2019)

    2019

  6. [6]

    A Comprehensive Survey on Transfer Learning

    Fuzhen Zhuang et al. “A Comprehensive Survey on Transfer Learning”.Proceedings of the IEEE109 (2021)

    2021

  7. [7]

    Long-term stable electromyography classification using canonical correlation analysis

    Elisa Donati et al. “Long-term stable electromyography classification using canonical correlation analysis”.2023 11th International IEEE/EMBS Conference on Neural Engineering (NER). IEEE. 2023

    2023

  8. [8]

    Ttt++: When does self-supervised test-time training fail or thrive?

    Yuejiang Liu et al. “Ttt++: When does self-supervised test-time training fail or thrive?”Advances in Neural Information Processing Systems34 (2021)

    2021

  9. [9]

    Autoencoder based self-supervised test- time adaptation for medical image analysis

    Yufan He et al. “Autoencoder based self-supervised test- time adaptation for medical image analysis”.Medical image analysis72 (2021)

    2021

  10. [10]

    Test-time training with masked autoencoders

    Yossi Gandelsman et al. “Test-time training with masked autoencoders”.Advances in Neural Information Process- ing Systems35 (2022)

    2022

  11. [11]

    Revisiting batch normalization for practical domain adaptation

    Yanghao Li et al. “Revisiting batch normalization for practical domain adaptation”.arXiv preprint arXiv:1603.04779(2016)

    Pith/arXiv arXiv 2016

  12. [12]

    Batch normaliza- tion: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. “Batch normaliza- tion: Accelerating deep network training by reducing internal covariate shift”.International conference on machine learning. pmlr. 2015

    2015

  13. [13]

    Memory-Latency-Accuracy Trade-Offs for Continual Learning on a RISC-V Extreme- Edge Node

    Leonardo Ravaglia et al. “Memory-Latency-Accuracy Trade-Offs for Continual Learning on a RISC-V Extreme- Edge Node”.2020 IEEE Workshop on Signal Processing Systems (SiPS). 2020

    2020

  14. [14]

    Pietro Buzzega et al.Dark Experience for General Continual Learning: a Strong, Simple Baseline. 2020. arXiv: 2004.07211[stat.ML]

    Pith/arXiv arXiv 2020

  15. [15]

    sEMG-based Regression of Hand Kinematics with Temporal Convolutional Networks on a Low-Power Edge Microcontroller

    M. Zanghieri et al. “sEMG-based Regression of Hand Kinematics with Temporal Convolutional Networks on a Low-Power Edge Microcontroller”.IEEE COINS 2021. 2021

    2021

  16. [16]

    Repeatability of grasp recogni- tion for robotic hand prosthesis control based on sEMG data

    Francesca Palermo et al. “Repeatability of grasp recogni- tion for robotic hand prosthesis control based on sEMG data”.2017 International Conference on Rehabilitation Robotics (ICORR). IEEE. 2017

    2017

  17. [17]

    Meta-learning with memory- augmented neural networks

    Adam Santoro et al. “Meta-learning with memory- augmented neural networks”.International conference on machine learning. PMLR. 2016. 12

    2016

  18. [18]

    A comprehensive survey on test-time adaptation under distribution shifts

    Jian Liang, Ran He, and Tieniu Tan. “A comprehensive survey on test-time adaptation under distribution shifts”. International Journal of Computer Vision133 (2025)

    2025

  19. [19]

    Continual lifelong learning with neural networks: A review

    German I Parisi et al. “Continual lifelong learning with neural networks: A review”.Neural networks113 (2019)

    2019

  20. [20]

    A survey of deep meta-learning

    Mike Huisman, Jan N Van Rijn, and Aske Plaat. “A survey of deep meta-learning”.Artificial Intelligence Review54 (2021)

    2021

  21. [21]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe. “Batch normalization: Accelerating deep network training by reducing internal covariate shift”. arXiv preprint arXiv:1502.03167(2015)

    Pith/arXiv arXiv 2015

  22. [22]

    Robust Real-Time Embedded EMG Recognition Framework using Temporal Convolutional Networks on a Multicore IoT Processor

    M Zanghieri et al. “Robust Real-Time Embedded EMG Recognition Framework using Temporal Convolutional Networks on a Multicore IoT Processor”.IEEE TBIOCAS 14 (2019)

    2019

  23. [23]

    Robust long-term hand grasp recog- nition with raw electromyographic signals using multi- dimensional uncertainty-aware models

    Yuzhou Lin et al. “Robust long-term hand grasp recog- nition with raw electromyographic signals using multi- dimensional uncertainty-aware models”.IEEE Transac- tions on Neural Systems and Rehabilitation Engineering 31 (2023)

    2023

  24. [24]

    Bioformers: Embedding trans- formers for ultra-low power sEMG-based gesture recog- nition

    Alessio Burrello et al. “Bioformers: Embedding trans- formers for ultra-low power sEMG-based gesture recog- nition”.2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE. 2022

    2022

  25. [25]

    WaveFormer: A Lightweight Trans- former Model for sEMG-based Gesture Recognition

    Yanlong Chen et al. “WaveFormer: A Lightweight Trans- former Model for sEMG-based Gesture Recognition”. arXiv preprint arXiv:2506.11168(2025)

    arXiv 2025

  26. [26]

    Towards generalisable time se- ries understanding across domains

    Özgün Turgut et al. “Towards generalisable time se- ries understanding across domains”.arXiv preprint arXiv:2410.07299(2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Moment: A family of open time-series foundation models

    Mononito Goswami et al. “Moment: A family of open time-series foundation models”.arXiv preprint arXiv:2402.03885(2024)

    Pith/arXiv arXiv 2024

  28. [28]

    EdwardJ.Huetal.LoRA:Low-RankAdaptationofLarge Language Models. 2021. arXiv: 2106.09685[cs.CL]

    Pith/arXiv arXiv 2021

  29. [29]

    Model- agnostic meta-learning for fast adaptation of deep net- works

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. “Model- agnostic meta-learning for fast adaptation of deep net- works”.International conference on machine learning. PMLR. 2017

    2017

  30. [30]

    sEMG-based regression of hand kinematics with temporal convolutional networks on a low-power edge microcontroller

    Marcello Zanghieri et al. “sEMG-based regression of hand kinematics with temporal convolutional networks on a low-power edge microcontroller”.2021 IEEE International Conference on Omni-Layer Intelligent Systems (COINS). IEEE. 2021

    2021