Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Quantizing Small-Scale State-Space Models for Edge AI

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Quantization-aware training in recurrent mode lifts a small state-space model from 40% post-training-quantization accuracy back to 96%, with about 6x parameter-memory savings.

desk verdict Solid empirical ablation of S4D quantization with a real QAT recovery result, undercut by a nominal 6x memory claim that omits metadata and state overhead. read the letter →

arxiv 2506.12480 v1 pith:HLUGNDXY submitted 2025-06-14 cs.LG eess.SP

Leo Zhao , Tristan Torchet , Melika Payvand , Laura Kriener , Filippo Moro This is my paper
classification cs.LGeess.SP
keywords state-spacemodelsquantization-awaretrainingpost-trainingquantizationS4DedgeAIheterogeneoussequentialMNISTmemorycompression
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

State-space models process sequences one step at a time with a fixed-size hidden state, which makes them attractive for small, low-power edge devices, but their recurrent state is fragile under the low-precision arithmetic those devices use. This paper studies that fragility in a small diagonal state-space model called S4D on the sequential-MNIST benchmark. It shows that simple post-training quantization is not enough: when the state-transition matrix and the internal state are quantized to 8 bits, accuracy collapses to around 40%. It then shows that quantization-aware training, run in recurrent mode, restores accuracy to roughly 96% with an 8-bit state, and that a heterogeneous assignment of bit-widths across components cuts the model's parameter memory by about 6x. The practical point is a concrete recipe for compressing small sequence models for edge hardware without losing the task.

What carries the argument

The machinery is the recurrent-mode update $x_{n+1} = \bar{A} x_n + \bar{B} u_n$ of a diagonal state-space layer, where the complex state-transition matrix $\bar{A}$ is trained and then quantized after discretization. The argument's load-bearing mechanism is the unit-circle condition: stability requires all eigenvalues (here the diagonal entries) of $\bar{A}$ to land inside the unit circle, and low-precision quantization can push them outside, so errors grow like $\bar{A}^n$; the paper counters this with per-step clipping of $x$ to $[-50,50]$ and with quantization-aware training. Three parameterizations carry the QAT part: training the discrete matrices directly, training the continuous-time parameterization, and freezing $\bar{A}$ while training the rest, with frozen $\bar{A}$ performing best at the low precisions most relevant to hardware.

What would settle it

Measure the actual deployed memory of the 64-head S4D with the W4A6, $\bar{A}$8, $x$8 scheme on an edge microcontroller or a cycle-accurate memory simulator, counting all scales, offsets, state buffers, and alignment operations; if the measured saving is far below 6x, the headline compression claim fails. Alternatively, run the same PTQ-then-QAT protocol on a second sequence task, such as keyword spotting, and check whether the 8-bit boundary for $\bar{A}$ and $x$ still holds.

Watch

Extended reading notes

Core claim

The paper's central claim is that quantization-induced failure in small state-space models is localized to two quantities—the discretized state-transition matrix $\bar{A}$ and the recurrent state $x$—and that this failure can be repaired by quantization-aware training rather than by increasing model width. Because $\bar{A}$ is multiplied into $x$ at every time step, quantization errors in $\bar{A}$ are amplified by powers of the time step, so even one eigenvalue landing outside the unit circle makes the state explode; clipping $x$ to $[-50,50]$ keeps the recurrence bounded. Retraining the quantized model for ten epochs in recurrent mode with a straight-through gradient estimator lifts accuracy from roughly 40% after post-training quantization to 95.99% on the 64-head model, with weights and non-state activations at 4 and 6 bits and $\bar{A}$ and $x$ at 8 bits. This configuration reduces the full-precision model's parameter memory by 84.05%, stated as a 6x compression.

Load-bearing premise

The headline 6x memory saving is counted from the nominal bit-widths of weights and activations and assumes those bit counts dominate the deployed footprint; if the extra bookkeeping values that quantized arithmetic requires (per-head scales, offsets, complex state buffers, and range-alignment overhead) are counted, the true saving could be materially smaller.

Editorial extensions

If this is right

  • Post-training quantization alone cannot push the state-transition matrix and state below 8 bits on this task; even wider models fail below 8 bits, so recurrent-mode quantization-aware training is the practical path for small SSMs.
  • Quantizing most matrices to 4 bits, non-state activations to 6 bits, and only $\bar{A}$ and $x$ to 8 bits is enough to stay near full-precision accuracy, so memory savings come from the insensitive components rather than from the sensitive ones.
  • Freezing $\bar{A}$ during quantization-aware training is the best choice at low bit precision, which means a deployed chip does not need fine-grained updates to the unstable part of the model.
  • At 8 bits for $\bar{A}$ and $x$, the heterogeneous scheme attains 95.99% accuracy on the 64-head model while reducing the full-precision parameter memory by 84.05%, i.e., a 6x compression.

Reading between the lines

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

  • The unit-circle diagnosis suggests a targeted training constraint the paper does not explore: forcing the quantized eigenvalues of $\bar{A}$ to lie strictly inside the unit circle could enable sub-8-bit $\bar{A}$, something the paper observes is otherwise unstable.
  • Because the failure mode is recurrence-specific, the results likely carry over to other diagonal state-space models run recurrently, but selective state-space models with input-dependent transitions may need a separate analysis.
  • The recipe is validated only on sequential MNIST; speech, ECG, or sensor benchmarks would reveal whether the 8-bit boundary and the 6x saving generalize to real edge workloads.
  • The paper's memory accounting counts nominal bit-widths, so the true on-device saving depends on overheads such as per-head scales, offsets, and alignment operations that the paper leaves for future work.
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 / 5 minor

Summary. The paper studies quantization of small S4D state-space models for edge inference. Using the sequential MNIST benchmark, it first applies post-training quantization (PTQ) in recurrent mode and identifies the state transition matrix A and the recurrent state x as the most sensitive components, requiring at least 8 bits, while B, C, D and non-state activations tolerate 4-6 bits. It then applies quantization-aware training (QAT) with three parameterizations of A (discrete, continuous/original, and frozen) and proposes a heterogeneous scheme (4-bit weights, 6-bit non-state activations, 8-bit A and x) that recovers roughly 96% test accuracy from roughly 40% under PTQ on the 64-head model, with an advertised 6x memory reduction. The paper also provides a mechanistic stability analysis: quantization can move eigenvalues of the discrete transition matrix outside the unit circle, causing errors to grow as powers of the time step, and clipping the state is proposed as a mitigation.

Significance. If the memory figure were accurate, the result would be a useful practical demonstration that recurrent-mode QAT plus a heterogeneous bit-width assignment can make small S4D models edge-viable while retaining near-baseline accuracy. The empirical core is carefully executed: the PTQ ablations use five seeds with error bars, the unit-circle instability explanation is mechanistically plausible, and the benchmark is external, so the reported accuracies are direct measurements rather than quantities that reduce by construction. The main limitations are that the 6x memory-reduction claim rests on an incomplete memory model, the design choices are not protected by a described validation protocol, and the 'without sacrificing performance' wording is stronger than the reported accuracy differences support.

major comments (3)
  1. [V-A (Table III) and Conclusion] The '6x memory reduction' is computed from the nominal bit widths of weights and activations, but no full memory model is given. For per-head quantization with H=64, each quantized tensor needs at least one scale and zero-point per head; the complex tensors A, B, C require storage for real and imaginary parts even if quantized with a shared scheme; the recurrent state buffers must physically persist during streaming; and the Outlook itself acknowledges 'computational overhead to align the integer ranges'. These terms are omitted from the 'Mem Savings' percentages in Table III. For a 21,514-parameter model, 64 heads times 4 SSM tensors times (scale + zero-point) is already on the order of 1-2 KB against a roughly 13 KB 5-bit payload, so the true deployed footprint reduction may be materially smaller than 6x. Please provide the memory formula used for Table III and, ideally, an actual on-device memory measurement.
  2. [II and V-A (Fig. 5c, Table III)] The paper does not describe a validation split or a model-selection protocol. The state clipping bound [-50, 50], the gradient clipping bound [-1000, 1000], the 99.999% calibration percentile, and the heterogeneous bit assignment W4A6/A8x8 are all presented as fixed choices, but they appear to be selected after inspecting test-set behavior. If any of these choices were tuned on the sMNIST test set, the reported 95.99 +/- 0.56 accuracy is optimistic. At minimum, report validation accuracy and provide a sensitivity analysis for the state clipping bound and the gradient clipping bound; these are the central free parameters of the proposed method.
  3. [Abstract and Conclusion; Table III] The claim that the scheme reduces memory by 6x 'without sacrificing performance' is not supported by the reported numbers. The best heterogeneous QAT result for S4D-64 (95.99 +/- 0.56) is 2.8 percentage points below the full-precision baseline (98.79 +/- 0.12), and for S4D-16 the same scheme gives 88.39 +/- 1.40 against a 97.20 +/- 0.25 baseline. The paper reports no significance test; on S4D-64 the error bars do not overlap, so the accuracy difference is likely meaningful. Please either report a statistical comparison of the quantized and full-precision models or temper the claim to 'with a limited accuracy loss'.
minor comments (5)
  1. [IV, Eq. (2)] Equation (2) uses A both for the continuous-time matrix and inside the discrete denominator: the expression Q(Disc(\bar{A})) = Q(e^{A\Delta}) is notationally unclear and should distinguish the continuous A from the discrete \bar{A} explicitly.
  2. [II] The text 'qGELUq(x)' appears to be a typographical duplicate; it should read qGELU(x) or a similarly defined quantized GELU once.
  3. [Conclusion] The Conclusion cites the Loihi 2 implementation [20] when referring to the S4D architecture; the architectural description should cite the original S4D paper [4] instead.
  4. [Table III] The notation 'A16x16' is ambiguous because the caption says it denotes precisions for A and x, but the table body does not explicitly separate the A bit-width from the x bit-width; please clarify the notation in the table header or caption.
  5. [Fig. 2b] The legend entries 'Dynamic', 'Static', 'State', and 'Non-state' are difficult to map to the four curves; please label each curve directly or use a clearer legend that distinguishes the quantized tensor (state versus non-state) and the calibration mode.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports direct empirical measurements and an explicit bit-count arithmetic for memory savings; no prediction reduces to its inputs by construction.

full rationale

The paper does not contain a predictive derivation chain. Its main claims are empirical measurements on the sMNIST benchmark: PTQ sensitivity ablations (Fig. 2, Fig. 3), QAT recovery curves (Fig. 5), and final accuracies in Table III. The 40%-to-96% accuracy recovery is a direct measured comparison between PTQ and QAT runs, not a quantity implied by the paper's own equations. The 6x memory-footprint claim is an explicit arithmetic ratio: the reported savings percentages (76.55%, 84.05%, etc.) equal the reduction obtained by replacing 32-bit elements with the chosen 4-, 6-, or 8-bit elements, as stated in Table III. That is a definitional accounting of nominal bit-widths, not a hidden fit or a self-referential prediction. The paper also discloses in the Outlook that the heterogeneous scheme 'incurs some computational overhead to align the integer ranges,' which is a limitation on the deployed footprint rather than a circular step. There are no load-bearing self-citations: references to S4D, HiPPO, Q-S5, and quantization tooling are external prior work, and the paper's authors do not invoke their own prior results to forbid alternatives or justify the central claim. The skeptical concern about omitted per-head metadata and real/imaginary storage is a correctness or completeness risk about whether the nominal 6x survives on-device, but it does not make any derivation circular. Accordingly, no circularity steps are identified, and the score is 0.

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

The central claims rest mainly on the empirical setup rather than on derived quantities. The hand-chosen clipping, calibration, gradient, and precision-mix parameters shift the measured outcomes, and the stability explanation relies on the standard eigenvalue criterion plus the assumption that recurrent mode is the right deployment model. No new entities are postulated.

free parameters (4)
  • State clipping bound = [-50, +50]
    Heuristic clip on real and imaginary parts of the recurrent state at each time step to prevent divergence (Section IV). The bound is not swept or justified, and it affects all PTQ and QAT results.
  • Calibration percentile = 99.999%
    Used to set quantization ranges (Section II). Chosen to remove outliers; no sensitivity analysis is reported.
  • QAT gradient clipping bound = [-1000, +1000]
    Applied to straight-through estimator gradients during recurrent QAT (Section II) to prevent exploding gradients; chosen by hand.
  • Heterogeneous bit-width assignment = 4-bit weights, 6-bit activations, 8-bit A and state x
    Selected to balance the observed sensitivity of each component and then evaluated on the same benchmark (Section V-A, Table III), so the reported tradeoff is partly a fit to the test set.
assumptions (4)
  • standard math A discrete LTI system is stable if and only if its eigenvalues lie inside the unit circle.
    Used in Section IV to explain why quantized A values outside the unit circle cause state explosion.
  • domain assumption Recurrent-mode inference accurately models how quantization error propagates at the edge.
    Motivates the switch from convolutional to recurrent mode before quantization in Section II.
  • domain assumption The straight-through estimator provides usable gradients through the quantization function.
    Standard QAT practice, invoked in Section II without additional justification.
  • ad hoc to paper The state-clipping bound of [-50, +50] is a sufficient and general solution to instability.
    Introduced in Section IV; the paper presents no sensitivity analysis or formal justification for this exact range.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantizing Small-Scale State-Space Models for Edge AI." pith.science (2026). https://pith.science/paper/HLUGNDXY

@misc{pith2026250612480,
  author       = {Pith},
  title        = {Pith review of: Quantizing Small-Scale State-Space Models for Edge AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLUGNDXY}},
  note         = {Machine review of arXiv:2506.12480}
}
read the original abstract

State-space models (SSMs) have recently gained attention in deep learning for their ability to efficiently model long-range dependencies, making them promising candidates for edge-AI applications. In this paper, we analyze the effects of quantization on small-scale SSMs with a focus on reducing memory and computational costs while maintaining task performance. Using the S4D architecture, we first investigate post-training quantization (PTQ) and show that the state matrix A and internal state x are particularly sensitive to quantization. Furthermore, we analyze the impact of different quantization techniques applied to the parameters and activations in the S4D architecture. To address the observed performance drop after Post-training Quantization (PTQ), we apply Quantization-aware Training (QAT), significantly improving performance from 40% (PTQ) to 96% on the sequential MNIST benchmark at 8-bit precision. We further demonstrate the potential of QAT in enabling sub-8-bit precisions and evaluate different parameterization schemes for QAT stability. Additionally, we propose a heterogeneous quantization strategy that assigns different precision levels to model components, reducing the overall memory footprint by a factor of 6x without sacrificing performance. Our results provide actionable insights for deploying quantized SSMs in resource-constrained environments.

Figures

Figures reproduced from arXiv: 2506.12480 by the authors.

Figure 1
Figure 1. Quantization of small-scale SSM networks for edge Artificial Intelligence (AI). Edge AI processors impose constraints on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Accuracy on sMNIST using PTQ of different S4D model components with 16 heads and different quantization modes. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Effect of scaling the model width on PTQ performance. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Quantization of complex-valued parameters. (a) Real and Imaginary parts are quantized independently, here with 3-bit [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: QAT on S4D model with 16 (a) and 64 (b) heads, with all parameters and activations quantized at 8 bits, using [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. QS4D: Quantization-aware training for efficient hardware deployment of structured state-space sequential models

    cs.LG 2025-07 conditional novelty 4.0 of 10

    Quantization-aware training allows S4D sequence models to run at much lower precision, cutting estimated hardware costs by up to two orders of magnitude while keeping accuracy.

Reference graph

Works this paper leans on

29 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hippo: Recurrent memory with optimal polynomial projections,

    A. Gu, T. Dao, S. Ermon, A. Rudra, and C. R ´e, “Hippo: Recurrent memory with optimal polynomial projections,” Advances in neural information processing systems , vol. 33, pp. 1474–1487, 2020

    2020

  2. [2]

    Efficiently modeling long sequences with structured state spaces. arxiv 2021,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces. arxiv 2021,” arXiv preprint arXiv:2111.00396

    arXiv 2021

  3. [3]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers,

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, and C. R ´e, “Combining recurrent, convolutional, and continuous-time models with linear state space layers,” Advances in neural information processing systems, vol. 34, pp. 572–585, 2021

    2021

  4. [4]

    On the parameterization and initialization of diagonal state space models,

    A. Gu, K. Goel, A. Gupta, and C. R ´e, “On the parameterization and initialization of diagonal state space models,” Advances in Neural Information Processing Systems , vol. 35, pp. 35971–35983, 2022

    work page 2022

  5. [5]

    Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models,

    M. Sch ¨one, N. M. Sushma, J. Zhuge, C. Mayr, A. Subramoney, and D. Kappel, “Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models,” in 2024 International Conference on Neuromorphic Systems (ICONS) , pp. 124–131, IEEE, 2024

    work page 2024

  6. [6]

    Exploring the capability of mamba in speech applications,

    K. Miyazaki, Y . Masuyama, and M. Murata, “Exploring the capability of mamba in speech applications,” arXiv preprint arXiv:2406.16808, 2024

    arXiv 2024

  7. [7]

    Spiking structured state space model for monaural speech enhancement,

    Y . Du, X. Liu, and Y . Chua, “Spiking structured state space model for monaural speech enhancement,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 766–770, 2024

    work page 2024

  8. [8]

    Eeg-ssm: Leveraging state-space model for dementia detection,

    X.-T. Tran, L. Le, Q. T. Nguyen, T. Do, and C.-T. Lin, “Eeg-ssm: Leveraging state-space model for dementia detection,” arXiv preprint arXiv:2407.17801, 2024

    arXiv 2024

Show all 29 references
  1. [9]

    Harmamba: Effi- cient wearable sensor human activity recognition based on bidirectional selective ssm,

    S. Li, T. Zhu, F. Duan, L. Chen, H. Ning, and Y . Wan, “Harmamba: Effi- cient wearable sensor human activity recognition based on bidirectional selective ssm,” arXiv e-prints, pp. arXiv–2403, 2024

    2024

  2. [10]

    Optimis- ing tinyml with quantization and distillation of transformer and mamba models for indoor localisation on edge devices,

    T. Suwannaphong, F. Jovan, I. Craddock, and R. McConville, “Optimis- ing tinyml with quantization and distillation of transformer and mamba models for indoor localisation on edge devices,” Scientific Reports , vol. 15, no. 1, p. 10081, 2025

    2025

  3. [11]

    Learning long sequences in spiking neural networks,

    M.-I. Stan and O. Rhodes, “Learning long sequences in spiking neural networks,” Scientific Reports, vol. 14, no. 1, p. 21957, 2024

    2024

  4. [12]

    Zero-shot temporal resolution domain adaptation for spiking neural networks,

    S. Karilanova, M. Fabre, E. Neftci, and A. ¨Ozc ¸elikkale, “Zero-shot temporal resolution domain adaptation for spiking neural networks,” arXiv preprint arXiv:2411.04760 , 2024

    2024

  5. [13]

    Quantization-guided training for compact tinyml models,

    S. Ghamari, K. Ozcan, T. Dinh, A. Melnikov, J. Carvajal, J. Ernst, and S. Chai, “Quantization-guided training for compact tinyml models,” arXiv preprint arXiv:2103.06231 , 2021

    2021 arXiv

  6. [14]

    Quantization and deployment of deep neural networks on microcontrollers,

    P.-E. Novac, G. Boukli Hacene, A. Pegatoquet, B. Miramond, and V . Gripon, “Quantization and deployment of deep neural networks on microcontrollers,” Sensors, vol. 21, no. 9, p. 2984, 2021

    2021

  7. [15]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 2704–2...

    2018

  8. [16]

    Q-s5: Towards quantized state space models,

    S. Abreu, J. E. Pedersen, K. M. Heckel, and A. Pierro, “Q-s5: Towards quantized state space models,” arXiv preprint arXiv:2406.09477 , 2024

    2024 arXiv

  9. [17]

    Quamba: A post-training quantization recipe for selective state space models,

    H.-Y . Chiang, C.-C. Chang, N. Frumkin, K.-C. Wu, and D. Marculescu, “Quamba: A post-training quantization recipe for selective state space models,” arXiv preprint arXiv:2410.13229 , 2024

    2024 arXiv

  10. [18]

    Qmamba: Post- training quantization for vision state space models,

    Y . Li, X. Liu, J. Li, R. Xu, Y . Chen, and Z. Xiong, “Qmamba: Post- training quantization for vision state space models,” arXiv preprint arXiv:2501.13624, 2025

    2025 arXiv

  11. [19]

    Mamba-ptq: Outlier channels in recurrent large language models,

    A. Pierro and S. Abreu, “Mamba-ptq: Outlier channels in recurrent large language models,” arXiv preprint arXiv:2407.12397 , 2024

    2024 arXiv

  12. [20]

    A diagonal structured state space model on loihi 2 for efficient streaming sequence processing,

    S. M. Meyer, P. Weidel, P. Plank, L. Campos-Macias, S. B. Shrestha, P. Stratmann, and M. Richter, “A diagonal structured state space model on loihi 2 for efficient streaming sequence processing,” arXiv preprint arXiv:2409.15022, 2024

    2024 arXiv

  13. [21]

    Efficient neuromorphic signal processing with loihi 2,

    G. Orchard, E. P. Frady, D. B. D. Rubin, S. Sanborn, S. B. Shrestha, F. T. Sommer, and M. Davies, “Efficient neuromorphic signal processing with loihi 2,” in 2021 IEEE Workshop on Signal Processing Systems (SiPS) , pp. 254–259, IEEE, 2021

    2021

  14. [22]

    How to train your hippo: State space models with generalized orthogonal basis projections,

    A. Gu, I. Johnson, A. Timalsina, A. Rudra, and C. R ´e, “How to train your hippo: State space models with generalized orthogonal basis projections,” arXiv preprint arXiv:2206.12037 , 2022

    2022 arXiv

  15. [23]

    Estimating or propagating gradients through stochastic neurons for conditional computation,

    Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,” arXiv preprint arXiv:1308.3432, 2013

    2013 arXiv

  16. [24]

    Simplified state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. W. Linderman, “Simplified state space layers for sequence modeling,” arXiv preprint arXiv:2208.04933 , 2022

    2022 arXiv

  17. [25]

    S7: Selective and simplified state space layers for sequence modeling,

    T. Soydan, N. Zubi ´c, N. Messikommer, S. Mishra, and D. Scaramuzza, “S7: Selective and simplified state space layers for sequence modeling,” arXiv preprint arXiv:2410.03464 , 2024

    2024 arXiv

  18. [26]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  19. [27]

    Speech commands: A dataset for limited-vocabulary speech recognition,

    P. Warden, “Speech commands: A dataset for limited-vocabulary speech recognition,” arXiv preprint arXiv:1804.03209 , 2018

    2018 arXiv

  20. [28]

    Tinyml-based classifi- cation in an ecg monitoring embedded system,

    E. Kim, J. Kim, J. Park, H. Ko, and Y . Kyung, “Tinyml-based classifi- cation in an ecg monitoring embedded system,” Computers, Materials and Continua, vol. 75, no. 1, pp. 1751–1764, 2023

    2023

  21. [29]

    Long-term stable electromyography classification using canonical correlation analysis,

    E. Donati, S. Benatti, E. Ceolini, and G. Indiveri, “Long-term stable electromyography classification using canonical correlation analysis,” in 2023 11th International IEEE/EMBS Conference on Neural Engineering (NER), pp. 1–4, IEEE, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.