Pith. sign in

REVIEW 3 major objections 4 minor 18 references

QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUs

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read QScheduler matches hand-tuned gradient-sample counts on INT8 NPUs, no sweep required.

desk verdict Useful adaptive-q scheduler for ZO on-device training, but the 'no q tuning' headline claim is overstated because Qmax is set from the fixed-q sweep; still deserves a real referee. read the letter →

arxiv 2607.18802 v1 pith:API47333 submitted 2026-07-21 cs.LG

Victor Felipe Domingues Do Amaral (GeePs) , Pierre Demaj , Erwan Libessart (GeePs) , Laurent Folliot , Anthony Kolar (GeePs) , Philippe Bénabès (GeePs) This is my paper · ORCID
classification cs.LG
keywords zeroth-orderoptimizationon-devicelearningINT8quantizationadaptivegradientsamplingneuralprocessingunittransferhyperparameter-freetrainingmicrocontroller
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Zeroth-order optimization can fine-tune a model on an NPU that only runs inference, because gradients come from extra forward passes instead of a backward pass. The catch is the number of gradient samples q: too few and the gradient noise stalls training, too many and each update costs excessive compute, and the best q changes with model and task. QScheduler removes the q choice by starting small and doubling q whenever validation progress stalls for five epochs. In the paper's experiments on EuroSAT and STL-10, this adaptive schedule matches the best fixed-q configurations for both ResNet18 and MobileNetV2 in floating-point and INT8 settings, and it reports the first demonstration of on-device learning on the STM32N6 NPU. If this holds, on-device fine-tuning on inference-only microcontrollers can drop a costly hyperparameter search while keeping accuracy.

What carries the argument

QScheduler is the load-bearing mechanism: a patience-based controller that tracks the best smoothed validation metric and counts epochs without an improvement beyond tolerance τ. After p=5 stagnant epochs it raises q by α=2 on a discrete ladder (8, 16, 32, ...) up to Qmax. It sits on top of a zeroth-order gradient estimator that averages q Rademacher perturbations, each requiring one extra quantized forward pass on the NPU, with momentum accumulation on the CPU; in the INT8 setting each ±1 perturbation in integer space imposes the constraint that the effective perturbation radius must be at least the weight quantization scale.

What would settle it

Run QScheduler on the same EuroSAT ResNet18/MobileNetV2 INT8 setup with Qmax set to a value outside the sweep-informed range (e.g., 256 and 16 respectively, or a compute-budget cap), and check whether final accuracy stays within the fixed-q envelope; a large drop when Qmax changes would show the cap, not the scheduler, carried the tuning.

Watch

Extended reading notes

Core claim

The central claim is that the optimal number of zeroth-order gradient samples need not be chosen in advance. Small initial q gives cheap early updates; when validation progress stalls for a patience window, QScheduler doubles q up to a cap, spending extra forward passes only when gradient noise is the bottleneck. On EuroSAT this schedule lands within the spread of the best fixed-q runs for ResNet18 and MobileNetV2 in float and INT8, and it runs on the STM32N6 with MobileNetV2 on a 2000-image STL-10 subset. The INT8 gap relative to float is attributed to the discrete integer perturbation step and fixed quantization scales while weights drift.

Load-bearing premise

The claim that QScheduler needs no q tuning assumes the cap Qmax is not itself doing the tuning: the paper sets Qmax to the best q from the fixed-q sweep (1024 for ResNet18, 64 for MobileNetV2), so if success depended on that sweep-informed cap, the 'no hyperparameter search' conclusion would not follow.

Editorial extensions

If this is right

  • On inference-only INT8 NPUs, fine-tuning is possible without any backward-pass primitive: the NPU only runs forwards, and the CPU estimates and accumulates gradients.
  • If QScheduler works as reported, q no longer needs a per-task hyperparameter sweep; starting at q=8 and doubling on plateaus lands near the tuned fixed-q accuracy.
  • The scheduler naturally stops at the saturation point where extra samples give marginal gains, so compute is spent on gradient quality only when it changes accuracy.
  • The paper's analysis attributes the float-to-INT8 accuracy drop to the fixed quantization scale and discrete perturbation step, implying that dynamic scale recalibration is a concrete way to close the gap.
  • The integer-space ZO constraint, that a perturbation must be at least one quantization step, is a general design rule for quantized zeroth-order training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'no hyperparameter search' framing is only as strong as the cap Qmax: the paper sets Qmax to the sweep-found optimum (1024/64), so a truly hyperparameter-free recipe would need to show the schedule still matches when Qmax is chosen by compute budget rather than by looking at the baselines.
  • Patience and multiplier are new knobs in place of q; the paper reports α=2 and p=5 worked, but does not characterize how sensitive the match is to those choices, so a natural test is sweeping patience over a wider range.
  • Nothing in the mechanism is specific to the last fully-connected layer or to the two datasets tested, so the same plateau-triggered sampling could apply to fine-tuning more layers or other forward-only accelerators; that is a testable extension the paper does not run.
  • A dynamic recalibration of quantization scales at plateaus, suggested by the paper's own error analysis, would let QScheduler jointly manage gradient noise and quantization drift; this could reduce the INT8 gap beyond what fixed scales allow.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes QScheduler, an adaptive algorithm for zeroth-order (ZO) optimization that increases the number of gradient samples q when validation accuracy plateaus, aiming to avoid manual tuning of q. It is evaluated in simulation on EuroSAT with ResNet18 and MobileNetV2 under both floating-point and INT8 fake quantization, and on real hardware with MobileNetV2 INT8 on the STM32N6's Neural-ART NPU using a subset of STL-10. The central claim is that QScheduler matches well-tuned fixed-q configurations without requiring prior q hyperparameter optimization, and that this is the first proof-of-concept of INT8 on-device training on the STM32N6.

Significance. If substantiated, the paper addresses a practical pain point in ZO-based on-device learning: the need to choose q for a given model/dataset. The idea of adaptively increasing q on plateaus is simple and plausible, and the STM32N6 hardware demonstration is valuable for the tinyML community. However, the current evidence does not establish the headline claim: the scheduler's maximum sample count Qmax is set to the best q found in the same fixed-q sweep that defines the baselines, and the scheduler's own patience/multiplier parameters were tuned on the same tasks. The on-device result is presented without numerical tables, error bars, or multiple seeds. These issues are load-bearing for the 'no q tuning' claim, but they are addressable with additional experiments and a more careful framing.

major comments (3)
  1. [§IV-A, Tables II–IV] Qmax is set to 1024 for ResNet18 and 64 for MobileNetV2, exactly the highest q values in the fixed-q sweeps and the best-performing fixed-q baselines (e.g., Table III INT8 q=1024: 86.05% vs QScheduler 86.21%; Table IV INT8 q=64: 85.40% vs QScheduler 85.09%). Thus Qmax is not an innocuous cap; it encodes the outcome of the same q sweep that defines the 'well-tuned fixed-q' baselines. The claim that QScheduler matches well-tuned fixed-q 'without requiring prior q hyperparameter optimization' is therefore circular as presented. Please provide sensitivity experiments with Qmax values not informed by the sweep (e.g., a single budget for both models, or several values), and report the q trajectory over training.
  2. [§IV-A and Algorithm 1] The scheduler's own parameters—α (set to 2), patience (set to 5), q0 (set to 8), Qmax, and the tolerance τ used in the preference relation M≻M_best—are either tuned on the same tasks or left unspecified. The text states that α∈{1.5,2.0} and patience∈{5,10} were tested and the best selected, so the method is not free of hyperparameter optimization. Moreover, τ appears in the algorithm definition but is not listed in Table II or the Require clause. Please report τ, provide sensitivity/ablations, or restrict the claim to 'no q-sweep needed' once the other parameters are fixed.
  3. [§IV-B, Figure 4] The on-device experiment is the paper's hardware proof-of-concept, but it reports only a single training curve without numerical final accuracies, standard deviations, or numbers of seeds. It also does not report wall-clock time or energy, despite the cost argument in Eq. (3). As presented, Figure 4 cannot independently support the claim that QScheduler matches fixed-q baselines on the STM32N6. Please provide a table with mean±std over at least 3 runs (or state run-to-run variability) and timing/energy measurements.
minor comments (4)
  1. [Eq. (1)] The ZO estimator is not typeset clearly: it should be \hat∇L(θ;z) = (L(θ+μz)−L(θ))/μ · z. The current notation, with z in the denominator, is ambiguous.
  2. [Algorithm 1] The tolerance τ is used in the definition of M≻M_best but is missing from the Require list and from Table II. Please specify its value.
  3. [Table II] The Qmax column lists {64,1024}, but the value is model-dependent (1024 for ResNet18, 64 for MobileNetV2). Clarify this in the caption.
  4. [§III-B] The relationship between the Rademacher perturbation w_INT8±1 and the real-space step θ±S_w is stated but not derived; a one-line derivation would help readers understand why μ≥S_w is necessary.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'no q tuning' claim is undercut: Qmax is set to the best q from the same fixed-q sweep, so QScheduler's final q is effectively the tuned q.

  1. fitted input called prediction [Abstract; Algorithm 1; Section IV-A, Tables II-IV]
    "Q max caps the maximum samples to avoid excessive computation in late training. ... TABLE III: ResNet18 validation accuracy (%) on EuroSAT (Qmax = 1024). TABLE IV: MobileNetV2 validation accuracy (%) on EuroSAT (Qmax = 64). ... QScheduler matches well-tuned fixed-qconfigurations for both ResNet18 and MobileNetV2, without requiring priorqhyperparameter optimization."

    In Tables III-IV, Qmax is exactly the largest q in the fixed-q baseline grid (ResNet18: 1024; MobileNetV2: 64), which is also the best/near-best configuration. Algorithm 1 monotonically increases q on plateaus via q←min(α·q, q max), so whenever training stalls QScheduler is driven toward the sweep-optimal q. The claimed 'match' to well-tuned fixed-q is therefore arranged by feeding the tuned q into the scheduler as Qmax; the tuning burden is renamed from 'q' to 'Qmax'. If Qmax had been set independently, e.g., 32 for MobileNetV2, the scheduler could not reach q=64 and the match would disappear (float q=32: 89.38% vs q=64: 91.73%). Thus the headline no-prior-q-tuning result is partly forced by construction.

full rationale

The internal controller (Algorithm 1) is not circular: it is a plateau-triggered multiplier and no theoretical result is imported from the authors' prior work. However, the central experimental claim is circular in an important respect: the scheduler is capped by Qmax, and Qmax is taken from the fixed-q sweep that defines the 'well-tuned' baselines. This transfers the q-selection problem into Qmax-selection while claiming no q hyperparameter optimization is needed. The floating-point/INT8 comparisons otherwise use standard external benchmarks and baselines, and no self-citation chain or uniqueness argument is load-bearing. Because the evaluation is partly self-confirming but the method itself still has independent content (adaptive scheduling could be evaluated with an externally chosen Qmax), the appropriate score is 6 rather than 8.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim's success depends on scheduler constants (Qmax in particular) chosen from the same sweep used to define the baseline. Beyond that, the method relies on the unproven heuristic that plateau-triggered q increases yield convergence and on standard ZO variance results. No new physical entities are introduced.

free parameters (5)
  • Qmax (maximum samples) = 1024 (ResNet18), 64 (MobileNetV2)
    Set to the best-performing q from the fixed-q sweep; caps QScheduler's reach and is central to the matching result.
  • patience p = 5 epochs
    Chosen from {5,10} after testing on EuroSAT (Section IV-A).
  • multiplier alpha = 2.0
    Chosen from {1.5,2.0} after testing on EuroSAT.
  • initial samples q0 = 8
    Chosen as a small start; not swept but a hand-picked constant.
  • training hyperparameters (lr, momentum, mu, batch size) = values from Bayesian search (Table I)
    Optimized on the same EuroSAT tasks; shared across methods, so they don't by themselves bias the QScheduler comparison, but they are fitted quantities.
assumptions (4)
  • standard math ZO gradient estimate variance scales as O(1/q), so increasing q improves gradient quality.
    Established result from Nesterov & Spokoiny [11], cited in Section II-B; the scheduler's premise that more samples reduce noise rests on it.
  • ad hoc to paper A plateau in validation accuracy indicates that gradient quality is the bottleneck and increasing q will improve convergence.
    This is the central heuristic of QScheduler (Algorithm 1) and is not derived or validated independently; it is an assumption about the training dynamics.
  • domain assumption Validation accuracy on a 20% split is a reliable progress metric for scheduling.
    The scheduler uses Mbest = validation accuracy (Section III-A); assumes this reflects generalization and is stable enough for plateau detection, especially with small validation sets.
  • domain assumption Fixed INT8 quantization scales, calibrated after warmup, remain adequate throughout quantized ZO training.
    The framework does not recalibrate during training (Section III-C, V); the authors themselves note in Discussion that scale drift contributes to the INT8 accuracy gap, so this assumption is acknowledged as imperfect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUs." pith.science (2026). https://pith.science/paper/API47333

@misc{pith2026260718802,
  author       = {Pith},
  title        = {Pith review of: QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/API47333}},
  note         = {Machine review of arXiv:2607.18802}
}
read the original abstract

Zeroth-Order (ZO) optimization enables On-Device Learning (ODL) on NPU-equipped microcontrollers by estimating gradients through forward passes alone, bypassing the need for backpropagation primitives and reducing memory requirements. The number of gradient samples q critically affects training: insufficient samples produce noisy gradients that plateau early, while excessive samples consume more computational resources. However, finding an optimal q typically requires costly hyperparameter searches. This work introduces QScheduler, an adaptive algorithm that adjusts q based on training progress, and provides the first proof-of-concept of INT8 quantized on-device training on the STM32N6's Neural-ART NPU. Experiments on EuroSAT and STL-10 show that QScheduler matches well-tuned fixed-q configurations for both ResNet18 and MobileNetV2, without requiring prior q hyperparameter optimization.

Figures

Figures reproduced from arXiv: 2607.18802 by the authors.

Figure 1
Figure 1. Backpropagation memory dependency: the backward [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. ZO training framework showing the computational [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Training curves for MobileNetV2 on EuroSAT: [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: QScheduler vs. fixed-q baselines across training epochs on STM32N6. Dots indicate q transitions. V. DISCUSSION The experiments reveal an important trade-off in ZO opti￾mization: while higher q values consistently improve accuracy, the marginal gains diminish beyond a c…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 2 linked inside Pith

  1. [1]

    MCUNet: Tiny deep learning on IoT devices,

    J. Lin, W.-M. Chen, Y . Lin, j. cohn john, C. Gan, and S. Han, “MCUNet: Tiny deep learning on IoT devices,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33, Curran Associates, Inc., 2020, pp. 11 711–11 722

    2020

  2. [2]

    Gholami, S

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer,A Survey of Quantization Methods for Efficient Neural Network Inference, arXiv:2103.13630 [cs], Jun. 2021. Accessed: May 9, 2023. [3]Stm32 model zoo, https : / / github . com / STMicroelectronics / stm32ai - modelzoo, Accessed: 2023-10-09

    arXiv 2021

  3. [4]

    TinyTL: Reduce memory, not parameters for efficient on-device learn- ing,

    H. Cai, C. Gan, L. Zhu, and S. Han, “TinyTL: Reduce memory, not parameters for efficient on-device learn- ing,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33, Curran Associates, Inc., 2020, pp. 11 285–11 297

    2020

  4. [5]

    Memory-efficient patch-based inference for tiny deep learning,

    J. Lin, W.-M. Chen, H. Cai, C. Gan, and S. Han, “Memory-efficient patch-based inference for tiny deep learning,” inAdvances in Neural Information Process- ing Systems, M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, Eds., vol. 34, Curran Associates, Inc., 2021, pp. 2346–2358

    2021

  5. [6]

    TinyOL: TinyML with online-learning on microcontrollers,

    H. Ren, D. Anicic, and T. A. Runkler, “TinyOL: TinyML with online-learning on microcontrollers,” in International Joint Conference on Neural Networks, IJCNN 2021, Shenzhen, China, July 18-22, 2021, IEEE, 2021, pp. 1–8

    2021

  6. [7]

    Learning representations by back-propagating errors,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533–536, 1986

    1986

  7. [8]

    Stepping forward on the last mile,

    C. Feng, S. Zhuo, X. Zhang, R. K. Ramakrishnan, Z. Yuan, and A. Z. Li, “Stepping forward on the last mile,” inAdvances in Neural Information Processing Systems, A. Globerson et al., Eds., vol. 37, Curran Associates, Inc., 2024, pp. 94 851–94 870

    2024

  8. [9]

    T. Chen, B. Xu, C. Zhang, and C. Guestrin, Training Deep Nets with Sublinear Memory Cost, arXiv:1604.06174 [cs], Apr. 2016. Accessed: May 26, 2023

    arXiv 2016

Show all 18 references
  1. [10]

    POET: Training neural networks on tiny devices with integrated rematerialization and paging,

    S. G. Patil, P. Jain, P. Dutta, I. Stoica, and J. Gonzalez, “POET: Training neural networks on tiny devices with integrated rematerialization and paging,” inProceedings of the 39th International Conference on Machine Learn- ing, K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvari...

    2022

  2. [11]

    Random gradient-free minimization of convex functions,

    Y . Nesterov and V . Spokoiny, “Random gradient-free minimization of convex functions,”Foundations of Computational Mathematics, vol. 17, no. 2, pp. 527– 566, 2017

    2017

  3. [12]

    Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,

    J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,” IEEE Transactions on Automatic Control, vol. 37, no. 3, pp. 332–341, 1992. [13]Stm32n657x0 datasheet: Arm cortex-m55 with neural- art accelerator, DS14555 Rev 3, STMi...

    1992

  4. [14]

    On-device training under 256KB memory,

    J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, and S. Han, “On-device training under 256KB memory,” in Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35, Curran Associates, Inc., 2022, pp. 22 941–22 954

    2022

  5. [15]

    Fine-tuning language models with just forward passes,

    S. Malladi et al., “Fine-tuning language models with just forward passes,” inAdvances in Neural Information Processing Systems, vol. 36, 2023. [16]Pytorch reducelronplateau, https : / / pytorch . org / docs / stable / generated / torch . optim . lr scheduler . ReduceLROnPlatea...

    2023

  6. [17]

    Adam: A method for stochas- tic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochas- tic optimization,” inInternational Conference on Learn- ing Representations (ICLR), 2015

    2015

  7. [18]

    Eu- rosat: A novel dataset and deep learning benchmark for land use and land cover classification,

    P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eu- rosat: A novel dataset and deep learning benchmark for land use and land cover classification,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 12, no. 7, pp. 2217–2226, 2019

    2019

  8. [19]

    An analysis of single- layer networks in unsupervised feature learning,

    A. Coates, A. Ng, and H. Lee, “An analysis of single- layer networks in unsupervised feature learning,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, vol. 15, PMLR, 2011, pp. 215–223

    2011

  9. [20]

    Practical bayesian optimization of machine learning hyperparam- eters,

    J. Snoek, H. Larochelle, and R. P. Adams, “Practical bayesian optimization of machine learning hyperparam- eters,” inAdvances in Neural Information Processing Systems, vol. 25, Curran Associates, Inc., 2012. [21]Stm32cube.ai (x-cube-ai), https : / / www . st . com / en / embed...

    2012

  10. [23]

    Anthropic,Claude, https://www.anthropic.com/claude, Large language model, 2024

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.