Pith. sign in

REVIEW 3 major objections 5 minor 57 references

SneakPeek: Data-Aware Model Selection and Scheduling for Inference Serving on the Edge

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read SneakPeek claims that estimating model accuracy from live label frequencies beats static profiling for edge inference scheduling.

desk verdict A solid engineering paper with a real, modest idea—online class-frequency estimation for accuracy-aware edge scheduling—whose evaluation is good enough to publish but short of airtight; the stable-recall assumption is the one soft spot worth probing. read the letter →

arxiv 2505.06641 v1 pith:ZOZWF3TQ submitted 2025-05-10 cs.DC

Joel Wolfrath , Daniel Frink , Abhishek Chandra This is my paper · ORCID
classification cs.DC
keywords inferenceservingedgecomputingmodelselectionschedulingaccuracyscalingdata-awareestimationclassfrequencybatching
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

The paper tries to establish that on a single GPU serving multiple applications, the right way to pick a model for each request is to estimate the current data's class frequencies and compute each model's expected accuracy from them, rather than trusting one profiled accuracy number. It argues that static profiled accuracy is biased toward the test set's label distribution and can mislead even an optimal scheduler. If correct, accuracy scaling at the edge becomes robust to label distribution shifts without retraining or modifying models. The authors also claim that grouping requests by application and batching them, plus a 'short-circuit' fast model, yields higher-utility schedules with few deadline violations.

What carries the argument

The load-bearing identity is Eq. (9): $\mathrm{Accuracy}(m) = \sum_i \theta_i \, \mathrm{recall}_i$, which decomposes profiled accuracy into class frequencies times per-class recall. Since recall is treated as a stable property of the trained model, the only quantity that changes out-of-sample is $\theta$, so the paper estimates $\theta$ in real time with a SneakPeek model: a k-nearest-neighbor lookup that produces multinomial evidence $y$, combined with a Dirichlet prior to obtain posterior SneakPeek probabilities. Those probabilities sharpen every model's accuracy estimate in one inference, and they also enable short-circuit inference and data-aware subgroup splitting. Scheduling uses the request priority $(1 + \mathrm{Var}[\mathrm{Accuracy}(\mathcal{M}_{a_i})])\,e^{-d_i}$, locally optimal model choice, and greedy application-level grouping to exploit batching.

What would settle it

Run SneakPeek on a stream whose class frequencies match the profiling set but whose inputs are modified within each class, such as different camera angles or added sensor noise; if observed per-class recall drops measurably, then Eq. (9)'s estimated accuracy diverges from observed accuracy, demonstrating bias under covariate shift.

Watch

Extended reading notes

Core claim

SneakPeek's central claim is that model accuracy for a classification model is not a static property but a function of the class frequencies $\theta$ of the data being served, via $\mathrm{Accuracy}(m) = \sum_i \theta_i \, \mathrm{recall}_i(m)$. The paper proposes estimating $\theta$ from live data using a lightweight SneakPeek model, built from k-nearest-neighbor evidence combined with a Dirichlet-Multinomial posterior, then using these sharpened accuracy estimates inside a scheduler that orders requests by a priority combining deadline urgency and accuracy variance, selects locally optimal models, groups requests by application for batching, and optionally uses the SneakPeek model itself to answer low-priority requests. On three healthcare-style applications, the combined approach achieves the highest utility, with a reported 2x utility increase over LO-EDF and almost zero deadline violations.

Load-bearing premise

The method assumes each model's per-class recall rates are stable and known from profiling, so only class frequencies change with the data; if the live data shifts within a class, the recall values themselves drift and the estimates are biased even with a perfect frequency estimate.

Editorial extensions

If this is right

  • Existing schedulers that rely on a single profiled accuracy number are systematically biased when live class frequencies differ from the test set, and they can be improved by swapping in per-class recall profiles plus a theta estimator.
  • Because theta is shared across all models of an application, one cheap SneakPeek inference updates accuracy estimates for every candidate model at once.
  • Batching requests by application and choosing one model per group reduces model-swap overhead, which matters more than per-request optimality under tight deadlines.
  • When deadlines are too tight for any profiled model, using the SneakPeek model itself as a zero-latency fallback salvages utility instead of returning zero.

Reading between the lines

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

  • The same theta decomposition should extend to other metrics such as F1 and quadratic scores, which the paper's appendix sketches, so the mechanism is not accuracy-specific.
  • A testable extension would be to calibrate theta estimates online with a small labeled stream and measure how much scheduling utility improves as SneakPeek estimation error falls, giving an operational target for estimator quality.
  • The approach may transfer to multi-GPU settings where the benefit of batching shrinks but the benefit of data-awareness grows, as the paper's preliminary multi-worker results suggest.
  • Since strongly informative priors can suppress the data signal, production deployments should prefer weak or uninformative priors unless the label distribution is known to be stationary.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 proposes SneakPeek, a data-aware model selection and scheduling framework for inference serving on resource-constrained edge systems. It formulates joint model selection and scheduling as a utility-maximization problem, then estimates the label distribution theta online using a k-nearest-neighbor-based SneakPeek model and a Dirichlet-Multinomial posterior. These theta estimates are plugged into a class-frequency decomposition of model accuracy (Eq. 9) to obtain sharper, data-dependent accuracy estimates for each candidate model. The scheduler combines a priority-based ordering, greedy group-level batching, and an optional short-circuit inference mechanism that uses the SneakPeek model itself when deadlines are tight. The evaluation uses three healthcare-style applications (fall detection, speech commands, ECG monitoring) on a single GPU and reports that SneakPeek achieves up to a 2x improvement in scheduling utility over a locally-optimal EDF baseline.

Significance. The analytical starting point is correct and clearly presented: Eq. (9) is a standard identity that decomposes profiled accuracy into class frequencies times per-class recalls, and the Dirichlet-Multinomial update is mathematically sound. If the stable-recall assumption holds, SneakPeek is an appealing, model-agnostic mechanism for adapting to label distribution shift without modifying the candidate DNNs. The paper also makes a useful pragmatic contribution by integrating batching into the scheduling decision rather than treating it as an afterthought. However, the empirical claims are not yet fully supported: the core invariance assumption is untested, the evaluation lacks repeated trials and error bars, and the reported utility may be computed from the same accuracy estimates that the method is designed to improve. These issues are fixable within the scope of the manuscript, but they are load-bearing for the paper's central claim.

major comments (3)
  1. [Section IV-A, Eq. (9) and Section VI] The central assumption that per-class recalls are stable between profiling time and online operation is load-bearing but untested. The text in Section IV-A states that recall depends only on the trained model, but this holds only if the conditional input distribution within each class remains unchanged. Realistic edge deployments exhibit within-class covariate shift (camera angle and lighting for fall detection, microphone placement for speech commands, electrode configuration for ECG), which changes the confusion matrix and hence the recall terms. Figure 6 evaluates only the error in estimating theta under the same test distribution; it never measures recall stability. Without an experiment that perturbs the within-class conditional distributions and reports both recall drift and end-to-end schedule utility, the central contribution is conditional on an unverified invariance. Please add such an experiment (e.g., synthetic covariate shifts such as blur, noise, or sensor offset) and report whether SneakPeek's utility gains persist.
  2. [Section VI-B and VI-C] The empirical evaluation does not report repeated trials or error bars, and it is unclear whether the reported utility is computed from realized inference outcomes or from the model-accuracy estimates that SneakPeek produces. Equation (2) defines utility in terms of an expected accuracy term, so if Figure 5's utility is computed using SneakPeek's own estimated accuracies, then the improvement over baselines is partly by construction. Please specify exactly how each utility number is computed in Figures 5, 7, 10, and 12, and report mean and standard deviation over multiple random request sets. In addition, report the realized (actual) accuracy of the selected models, not only the estimated accuracy, to demonstrate that the scheduling choices translate into better outcomes.
  3. [Section V-B, Algorithm 1, and Section VI-D] The grouped scheduling algorithm has two free hyperparameters that are never subjected to sensitivity analysis: the brute-force threshold tau and the subgroup-splitting threshold theta_i > 0.5. The reported benefits of grouped scheduling in Figures 10-12 may depend on tau, and the splitting rule is justified only by intuition. Please report the tau values used in each experiment and test alternative split thresholds (or state that the results are insensitive to these choices). This is important because the paper's practical recommendation to adopt group-level scheduling depends on these parameters.
minor comments (5)
  1. [Equation (2)] The utility definition uses a scalar Accuracy(mj), while Eq. (9) defines accuracy as a sum over classes parameterized by theta; the notation should clarify that the scalar in Eq. (2) is the expected accuracy under the current theta estimate, otherwise the two notions of accuracy are inconsistent.
  2. [Section VI-C1] The phrase "true model accuracy" is potentially misleading: setting theta_i = 1 for the true class label gives the model's per-request conditional accuracy, not a long-run frequency. Please state explicitly that this is a per-request oracle estimate.
  3. [Figure 6 caption] The caption says "SneakPeek KNN" but the text explains that approximate nearest neighbors (Faiss) is used; please make the caption consistent and specify that k=1 and k=5 are the number of neighbors in the SneakPeek model, not in the candidate classifiers.
  4. [Section VI-B] Typo: "relativly" should be "relatively" in the discussion of the grouped scheduler's utility.
  5. [Algorithm 1] In the brute-force branch, the algorithm returns a solution for grouped requests, but the text does not define group-level utility for a group containing multiple requests; please state whether it is the average of the request utilities or another aggregation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Eq. 9 is a standard confusion-matrix identity, and SneakPeek's theta estimate is externally validated against a one-hot oracle; the only self-citation is contextual.

full rationale

The paper's central analytical step, Eq. 9, is an algebraic identity rewriting classification accuracy as a class-frequency-weighted sum of per-class recalls. It is not derived from the scheduling result, and it does not assume the conclusion; it simply exposes the dependence of profiled accuracy on the test class distribution. Theta is estimated independently using a k-nearest-neighbor evidence model with a Dirichlet-Multinomial posterior, and Figure 6 evaluates the resulting accuracy estimates against a per-request oracle that sets theta to the true class label. This provides an external check on the estimation mechanism rather than defining the prediction in terms of the fit. The scheduling utility comparison is likewise not forced: utility (Eq. 2) is a stated function of model accuracy and deadline penalty, and the paper reports utility, realized accuracy, and deadline violations as separate metrics. The paper cites the authors' prior work [7] for the observation that model accuracy varies across target classes and for multi-modal inference context, but this claim is also established analytically in the paper itself through Eq. 7-9, so the self-citation is not load-bearing. The main limitation, that per-class recall values are assumed stable under within-class covariate shift, is a robustness concern rather than a circularity: it may bias accuracy estimates in shifted environments, but it does not make the derivation equivalent to its inputs. No uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in solely through citation. Overall, the derivation chain is self-contained and the central predictions have independent empirical support.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the standard confusion-matrix identity and on two domain assumptions: per-class recalls transfer from profiling to online data, and kNN evidence from training data is informative about current class frequencies. The tunable prior and k give flexibility but do not constitute fitting the target result. No new physical or architectural entity is introduced.

free parameters (3)
  • k (kNN neighbors) = 5 (default; k=1 also evaluated)
    The number of nearest neighbors used to form the multinomial evidence y affects the posterior concentration. The paper defaults to k=5 without tuning and evaluates k=1 in Figure 6.
  • Dirichlet prior hyperparameters alpha_i = 0.5 for all i (Jeffreys uninformative prior) by default
    Hyperparameters control the prior on class frequencies. The paper evaluates weakly and strongly informative priors in Section VI-C3, but uses the uninformative prior for the main results.
  • Brute-force threshold tau = not specified in the paper
    Algorithm 1 uses tau to decide when to compute an exact group-level solution instead of the greedy fallback. No value or sensitivity analysis is given.
assumptions (4)
  • domain assumption Per-class recall values of each model are stable and can be profiled on a test set; online data differs only in class frequencies.
    Eq 8-9 rewrite model accuracy as a weighted sum of recalls with weights theta. If recall values change under distribution shift, the accuracy estimate is biased even with perfect theta estimation.
  • domain assumption kNN labels from the training set provide statistically useful evidence about the true class of a new query, and hence about theta.
    SneakPeek uses k nearest neighbors to construct the multinomial evidence y. This assumes the query lies on the same feature manifold as the training data within each class.
  • standard math Dirichlet-Multinomial conjugacy and the use of posterior means as probability estimates are valid for the application.
    Eq 10-11 use standard conjugate Bayesian updating. The math is correct, though the paper does not calibrate the posterior against actual long-run frequencies.
  • domain assumption Profiled inference latency and GPU memory usage are accurate and deterministic on the target hardware.
    The scheduler's deadline feasibility and utility computation rely on profiled latencies in Eq 1-2. If latencies vary, schedules become suboptimal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SneakPeek: Data-Aware Model Selection and Scheduling for Inference Serving on the Edge." pith.science (2026). https://pith.science/paper/ZOZWF3TQ

@misc{pith2026250506641,
  author       = {Pith},
  title        = {Pith review of: SneakPeek: Data-Aware Model Selection and Scheduling for Inference Serving on the Edge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZOZWF3TQ}},
  note         = {Machine review of arXiv:2505.06641}
}
read the original abstract

Modern applications increasingly rely on inference serving systems to provide low-latency insights with a diverse set of machine learning models. Existing systems often utilize resource elasticity to scale with demand. However, many applications cannot rely on hardware scaling when deployed at the edge or other resource-constrained environments. In this work, we propose a model selection and scheduling algorithm that implements accuracy scaling to increase efficiency for these more constrained deployments. We show that existing schedulers that make decisions using profiled model accuracy are biased toward the label distribution present in the test dataset. To address this problem, we propose using ML models -- which we call SneakPeek models -- to dynamically adjust estimates of model accuracy, based on the underlying data. Furthermore, we greedily incorporate inference batching into scheduling decisions to improve throughput and avoid the overhead of swapping models in and out of GPU memory. Our approach employs a new notion of request priority, which navigates the trade-off between attaining high accuracy and satisfying deadlines. Using data and models from three real-world applications, we show that our proposed approaches result in higher-utility schedules and higher accuracy inferences in these hardware-constrained environments.

Figures

Figures reproduced from arXiv: 2505.06641 by the authors.

Figure 1
Figure 1. SneakPeek system model for inference serving [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. SneakPeek model updating for fall detection [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of request priority. Therefore, the two hyperparameters in our system are the choice of prior distribution and the number of neighboring points (k) to include in the evidence. The value of k can be optimized via standard supervised learning procedures. An important insight is that existing scheduling algorithms can directly incorporate our SneakPeek modeling to improve accuracy estimation. The only chan… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Groups are split based on the computed SneakPeek probabilities [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of schedule utility across approaches [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Accuracy estimation error [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison of schedule utility across approaches [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Accuracy estimation error with different priors. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Impact of deadlines Figure 9a shows the resulting effects on accuracy estimation when the prior captures the true distribution in the data. We observe that our accuracy estimation improves for both uninformative priors and weakly informative priors. However, the stron…
Figure 12
Figure 12. Figure 12: Effect of request arrival rate. the order specified by the deadlines, as the tighter deadline requests become comparatively more urgent. 2) Number of Applications: We now examine how our approaches behave as the number of applications increases. Note that increasing t…
Figure 13
Figure 13. Figure 13: Choice of penalty function. across a 100ms scheduling window. We now fix the deadline at 200ms and vary the number of requests per window to observe the effects on schedule utility. We can think of this as scaling the number of patients that must be monitored by the e…
Figure 15
Figure 15. Figure 15: Comparisons with multiple workers. same hardware used in the single-worker experiments. We ensure that each worker operates independently to minimize resource contention. In our first experiment (figure 15a), we compare schedule utilities with two workers and vary the…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages

  1. [1]

    Applied machine learning at facebook: A datacenter infrastructure perspective,

    K. Hazelwood, S. Bird, D. Brooks, S. Chintala, U. Diril, D. Dzhulgakov, M. Fawzy, B. Jia, Y . Jia, A. Kalro, J. Law, K. Lee, J. Lu, P. Noordhuis, M. Smelyanskiy, L. Xiong, and X. Wang, “Applied machine learning at facebook: A datacenter infrastructure perspective,” in 2018 IEEE International Symposium on High Performance Computer Architecture (HPCA), 2018...

    work page 2018

  2. [2]

    INFaaS: Automated model-less inference serving,

    F. Romero, Q. Li, N. J. Yadwadkar, and C. Kozyrakis, “INFaaS: Automated model-less inference serving,” in 2021 USENIX Annual Technical Conference (USENIX ATC 21) . USENIX Association, Jul. 2021, pp. 397–411

    work page 2021

  3. [3]

    Holmes: Health online model ensemble serving for deep learning models in intensive care units,

    S. Hong, Y . Xu, A. Khare, S. Priambada, K. Maher, A. Aljiffry, J. Sun, and A. Tumanov, “Holmes: Health online model ensemble serving for deep learning models in intensive care units,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , ser. KDD ’20. New York, NY , USA: Association for Computing Machinery,...

    work page 2020

  4. [4]

    Hercules: Heterogeneity-aware inference serving for at-scale personal- ized recommendation,

    L. Ke, U. Gupta, M. Hempstead, C.-J. Wu, H.-H. S. Lee, and X. Zhang, “Hercules: Heterogeneity-aware inference serving for at-scale personal- ized recommendation,” in 2022 IEEE International Symposium on High- Performance Computer Architecture (HPCA) , 2022, pp. 141–154

    work page 2022

  5. [5]

    Serving deep neural networks at the cloud edge for vision applications on mobile platforms,

    Z. Fang, D. Hong, and R. K. Gupta, “Serving deep neural networks at the cloud edge for vision applications on mobile platforms,” in Proceedings of the 10th ACM Multimedia Systems Conference, ser. MMSys ’19. New York, NY , USA: Association for Computing Machinery, 2019, p. 36–47

    work page 2019

  6. [6]

    Octopus: Slo-aware progressive in- ference serving via deep reinforcement learning in multi-tenant edge cluster,

    Z. Zhang, Y . Zhao, and J. Liu, “Octopus: Slo-aware progressive in- ference serving via deep reinforcement learning in multi-tenant edge cluster,” in Service-Oriented Computing , F. Monti, S. Rinderle-Ma, A. Ruiz Cort ´es, Z. Zheng, and M. Mecella, Eds. Cham: Springer Nature Switzerland, 2023, pp. 242–258

    work page 2023

  7. [7]

    Leveraging multi-modal data for efficient edge inference serving,

    J. Wolfrath, A. Achanta, and A. Chandra, “Leveraging multi-modal data for efficient edge inference serving,” in 2024 IEEE/ACM 24th International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2024, pp. 408–417

    work page 2024

  8. [8]

    Sponge: Inference serving with dynamic slos using in-place vertical scaling,

    K. Razavi, S. Ghafouri, M. M ¨uhlh¨auser, P. Jamshidi, and L. Wang, “Sponge: Inference serving with dynamic slos using in-place vertical scaling,” in Proceedings of the 4th Workshop on Machine Learning and Systems, ser. EuroSys ’24. ACM, Apr. 2024, p. 184–191

    work page 2024

Show all 57 references
  1. [10]

    One size does not fit all: Quantifying and exposing the accuracy- latency trade-off in machine learning cloud service apis via tolerance tiers,

    M. Halpern, B. Boroujerdian, T. Mummert, E. Duesterwald, and V . J. Reddi, “One size does not fit all: Quantifying and exposing the accuracy- latency trade-off in machine learning cloud service apis via tolerance tiers,” in 2019 IEEE International Symposium on Performance Anal...

    2019

  2. [11]

    Qos-aware schedul- ing of heterogeneous servers for inference in deep neural networks,

    Z. Fang, T. Yu, O. J. Mengshoel, and R. K. Gupta, “Qos-aware schedul- ing of heterogeneous servers for inference in deep neural networks,” in Proceedings of the 2017 ACM on Conference on Information and Knowledge Management , ser. CIKM ’17. New York, NY , USA: Association for ...

    2017

  3. [12]

    Scheduling with fully compressible tasks: Appli- cation to deep learning inference with neural network compression,

    T. D. S. Barros et al., “Scheduling with fully compressible tasks: Appli- cation to deep learning inference with neural network compression,” in 2024 IEEE/ACM 24th International Symposium on Cluster, Cloud and Internet Computing (CCGrid) , 2024, pp. 327–336

    2024

  4. [13]

    MArk: Exploiting cloud services for Cost-Effective, SLO-Aware machine learning inference serving,

    C. Zhang, M. Yu, W. Wang, and F. Yan, “MArk: Exploiting cloud services for Cost-Effective, SLO-Aware machine learning inference serving,” in 2019 USENIX Annual Technical Conference (USENIX ATC 19). Renton, W A: USENIX Association, Jul. 2019, pp. 1049–1062

    2019

  5. [14]

    Proteus: A high-throughput inference-serving system with accuracy scaling,

    S. Ahmad, H. Guan, B. D. Friedman, T. Williams, R. K. Sitaraman, and T. Woo, “Proteus: A high-throughput inference-serving system with accuracy scaling,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Sy...

    2024

  6. [15]

    Loki: A system for serving ml inference pipelines with hardware and accuracy scaling,

    S. Ahmad, H. Guan, and R. K. Sitaraman, “Loki: A system for serving ml inference pipelines with hardware and accuracy scaling,” in Proceedings of the 33rd International Symposium on High-Performance Parallel and Distributed Computing , ser. HPDC ’24. New York, NY , USA: Associ...

    2024

  7. [16]

    Jellyfish: Timely inference serving for dynamic edge networks,

    V . Nigade, P. Bauszat, H. Bal, and L. Wang, “Jellyfish: Timely inference serving for dynamic edge networks,” in 2022 IEEE Real-Time Systems Symposium (RTSS), 2022, pp. 277–290

    2022

  8. [17]

    Real-time video analytics: The killer app for edge computing,

    G. Ananthanarayanan et al., “Real-time video analytics: The killer app for edge computing,” Computer, 2017

    2017

  9. [18]

    I. E. Olatunji and C.-H. Cheng, Video Analytics for Visual Surveillance and Applications: An Overview and Survey . Springer, 2019

    2019

  10. [19]

    Audio- and video-based human activity recognition systems in healthcare,

    S. Cristina, V . Despotovic, R. P ´erez-Rodr´ıguez, and S. Aleksic, “Audio- and video-based human activity recognition systems in healthcare,” IEEE Access, vol. 12, pp. 8230–8245, 2024

    2024

  11. [20]

    M. K. Hassan, A. I. El Desouky, S. M. Elghamrawy, and A. M. Sarhan, Big Data Challenges and Opportunities in Healthcare Informatics and Smart Hospitals. Cham: Springer International Publishing, 2019

    2019

  12. [21]

    Real-Time Arrhythmia Detection Using Hybrid Convolu- tional Neural Networks,

    S. C. Bollepalli, R. K. Sevakula, W. M. Au-Yeung, M. B. Kassab, F. M. Merchant, G. Bazoukis, R. Boyer, E. M. Isselbacher, and A. A. Armoundas, “Real-Time Arrhythmia Detection Using Hybrid Convolu- tional Neural Networks,” J Am Heart Assoc, vol. 10, no. 23, p. e023222, Dec 2021

    2021

  13. [22]

    Machine learning for predicting epileptic seizures using eeg signals: A review,

    K. Rasheed et al. , “Machine learning for predicting epileptic seizures using eeg signals: A review,” IEEE Reviews in Biomedical Engineering , vol. 14, pp. 139–155, 2021

    2021

  14. [23]

    A review of epileptic seizure detection using machine learning classifiers,

    M. K. Siddiqui et al. , “A review of epileptic seizure detection using machine learning classifiers,” Brain Informatics, vol. 7, no. 1, p. 5, May 2020

    2020

  15. [24]

    Prediction of respiratory decom- pensation in covid-19 patients using machine learning: The ready trial,

    H. Burdick, C. Lam, S. Mataraso, A. Siefkas, G. Braden, R. P. Dellinger, A. McCoy, J.-L. Vincent, A. Green-Saxena, G. Barnes, J. Hoffman, J. Calvert, E. Pellegrini, and R. Das, “Prediction of respiratory decom- pensation in covid-19 patients using machine learning: The ready t...

    2020

  16. [25]

    Identifying and monitoring respiratory compromise: Report from the rules and algorithms working group,

    B. Friedman, D. Fuckert, M. Jahrsdoerfer, R. Magness, E. S. Patterson, R. Syed, and J. R. Zaleski, “Identifying and monitoring respiratory compromise: Report from the rules and algorithms working group,” Biomedical Instrumentation & Technology, vol. 53, no. 2, pp. 110– 123, 2019

    2019

  17. [26]

    B. Chen, C. N. Potts, and G. J. Woeginger, A Review of Machine Scheduling: Complexity, Algorithms and Approximability. Boston, MA: Springer US, 1998, pp. 1493–1641

    1998

  18. [27]

    Complexity of task sequencing with deadlines, set-up times and changeover costs,

    J. Bruno and P. Downey, “Complexity of task sequencing with deadlines, set-up times and changeover costs,” SIAM Journal on Computing, vol. 7, no. 4, pp. 393–404, 1978

    1978

  19. [28]

    Metrics for multi-class classification: an overview,

    M. Grandini, E. Bagli, and G. Visani, “Metrics for multi-class classification: an overview,” 2020. [Online]. Available: https://arxiv.org/ abs/2008.05756

    2020 arXiv

  20. [29]

    Layercake: Efficient inference serving with cloud and mobile resources,

    S. Ogden and T. Guo, “Layercake: Efficient inference serving with cloud and mobile resources,” in 2023 23nd IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid) , 2023

    2023

  21. [30]

    Model selection for latency- critical inference serving,

    D. Mendoza, F. Romero, and C. Trippel, “Model selection for latency- critical inference serving,” in Proceedings of the Nineteenth European Conference on Computer Systems , ser. EuroSys ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 1016–1038

    2024

  22. [31]

    Clipper: A Low-Latency online prediction serving system,

    D. Crankshaw, X. Wang, G. Zhou, M. J. Franklin, J. E. Gonzalez, and I. Stoica, “Clipper: A Low-Latency online prediction serving system,” in 14th USENIX Symposium on Networked Systems Design and Imple- mentation (NSDI 17) . Boston, MA: USENIX Association, Mar. 2017, pp. 613–627

    2017

  23. [32]

    Mmact: A large-scale dataset for cross modal human action under- standing,

    Q. Kong, Z. Wu, Z. Deng, M. Klinkigt, B. Tong, and T. Murakami, “Mmact: A large-scale dataset for cross modal human action under- standing,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2019. 12

    2019

  24. [33]

    X3d: Expanding architectures for efficient video recognition,

    C. Feichtenhofer, “X3d: Expanding architectures for efficient video recognition,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA, USA: IEEE Computer Society, jun 2020, pp. 200–210

    2020

  25. [34]

    Minirocket: A very fast (almost) deterministic transform for time series classification,

    A. Dempster, D. F. Schmidt, and G. I. Webb, “Minirocket: A very fast (almost) deterministic transform for time series classification,” ser. KDD ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 248–257

    2021

  26. [35]

    Multi-stage based feature fusion of multi-modal data for human activity recognition,

    H. Choi, A. Beedu, H. Haresamudram, and I. Essa, “Multi-stage based feature fusion of multi-modal data for human activity recognition,” 2022

    2022

  27. [36]

    Speech commands: A public dataset for single-word speech recognition

    P. Warden, “Speech commands: A public dataset for single-word speech recognition.” 2017

    2017

  28. [37]

    Howl: A deployed, open-source wake word detection system,

    R. Tang, J. Lee, A. Razi, J. Cambre, I. Bicking, J. Kaye, and J. Lin, “Howl: A deployed, open-source wake word detection system,” in Proceedings of Second Workshop for NLP Open Source Software (NLP- OSS). Association for Computational Linguistics, Nov. 2020, pp. 61–65

    2020

  29. [38]

    PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals,

    A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. K. Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals,” Circulation, vol. 1...

    2000

  30. [39]

    ECG arrhythmia classification using a 2-d convolutional neural network,

    T. J. Jun, H. M. Nguyen, D. Kang, D. Kim, D. Kim, and Y . Kim, “ECG arrhythmia classification using a 2-d convolutional neural network,” CoRR, vol. abs/1804.06812, 2018. [Online]. Available: http://arxiv.org/abs/1804.06812

    2018 arXiv

  31. [40]

    Slo-aware inference scheduler for heterogeneous processors in edge platforms,

    W. Seo, S. Cha, Y . Kim, J. Huh, and J. Park, “Slo-aware inference scheduler for heterogeneous processors in edge platforms,” ACM Trans. Archit. Code Optim. , vol. 18, no. 4, jul 2021

    2021

  32. [41]

    The faiss library,

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazar ´e, M. Lomeli, L. Hosseini, and H. J ´egou, “The faiss library,” 2024

    2024

  33. [42]

    Kairos: Building cost- efficient machine learning inference systems with heterogeneous cloud resources,

    B. Li, S. Samsi, V . Gadepally, and D. Tiwari, “Kairos: Building cost- efficient machine learning inference systems with heterogeneous cloud resources,” in Proceedings of the 32nd International Symposium on High-Performance Parallel and Distributed Computing, ser. HPDC ’23. As...

    2023

  34. [43]

    Interference-aware scheduling for inference serving,

    D. Mendoza, F. Romero, Q. Li, N. J. Yadwadkar, and C. Kozyrakis, “Interference-aware scheduling for inference serving,” in Proceedings of the 1st Workshop on Machine Learning and Systems , ser. EuroMLSys ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 80–88

    2021

  35. [44]

    Inss: An intelligent scheduling orchestrator for multi-gpu inference with spatio-temporal sharing,

    Z. Han, R. Zhou, C. Xu, Y . Zeng, and R. Zhang, “Inss: An intelligent scheduling orchestrator for multi-gpu inference with spatio-temporal sharing,” IEEE Transactions on Parallel and Distributed Systems , vol. 35, no. 10, pp. 1735–1748, 2024

    2024

  36. [45]

    Elastic- room: Multi-tenant dnn inference engine via co-design with resource- constrained compilation and strong priority scheduling,

    L. Ma, H. Chen, E. Shao, L. Wang, Q. Chen, and G. Tan, “Elastic- room: Multi-tenant dnn inference engine via co-design with resource- constrained compilation and strong priority scheduling,” in Proceedings of the 33rd International Symposium on High-Performance Parallel and Di...

    2024

  37. [46]

    Mosel: Inference serving using dynamic modality selection,

    B. Hu, L. Xu, J. Moon, N. J. Yadwadkar, and A. Akella, “Mosel: Inference serving using dynamic modality selection,” 2023

    2023

  38. [47]

    Splitrpc: A Control + Data path splitting rpc stack for ml inference serving,

    A. Kumar, A. Sivasubramaniam, and T. Zhu, “Splitrpc: A Control + Data path splitting rpc stack for ml inference serving,”Proc. ACM Meas. Anal. Comput. Syst., vol. 7, no. 2, May 2023

    2023

  39. [48]

    Sommelier: Curating dnn models for the masses,

    P. Guo, B. Hu, and W. Hu, “Sommelier: Curating dnn models for the masses,” in Proceedings of the 2022 International Conference on Management of Data , ser. SIGMOD ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 1876–1890

    2022

  40. [49]

    Hawkvision: Low-latency modeless edge ai serving,

    C. Lao, J. Gao, G. Ananthanarayanan, A. Akella, and M. Yu, “Hawkvision: Low-latency modeless edge ai serving,” 2024. [Online]. Available: https://arxiv.org/abs/2405.19213

    2024 arXiv

  41. [50]

    Edgeadaptor: Online configuration adaption, model selection and resource provisioning for edge dnn inference serving at scale,

    K. Zhao et al. , “Edgeadaptor: Online configuration adaption, model selection and resource provisioning for edge dnn inference serving at scale,” IEEE Transactions on Mobile Computing , vol. 22, no. 10, pp. 5870–5886, 2023

    2023

  42. [51]

    On-demand edge inference scheduling with accuracy and deadline guarantee,

    Y . She et al., “On-demand edge inference scheduling with accuracy and deadline guarantee,” in 2023 IEEE/ACM 31st International Symposium on Quality of Service (IWQoS) , 2023, pp. 1–10

    2023

  43. [52]

    Seamless hw-accelerated ai serving in het- erogeneous mec systems with ai@edge,

    A. Tzenetopoulos et al. , “Seamless hw-accelerated ai serving in het- erogeneous mec systems with ai@edge,” in Proceedings of the 33rd International Symposium on High-Performance Parallel and Distributed Computing, ser. HPDC ’24. New York, NY , USA: Association for Computing M...

    2024

  44. [53]

    Offloading algorithms for maximizing inference accuracy on edge device under a time constraint,

    A. Fresa and J. P. Champati, “Offloading algorithms for maximizing inference accuracy on edge device under a time constraint,” CoRR, vol. abs/2112.11413, 2021. [Online]. Available: https://arxiv.org/abs/2112. 11413

    2021 arXiv

  45. [54]

    Optimizing deep learning inference on embedded systems through adaptive model selec- tion,

    V . S. Marco, B. Taylor, Z. Wang, and Y . Elkhatib, “Optimizing deep learning inference on embedded systems through adaptive model selec- tion,” ACM Trans. Embed. Comput. Syst. , vol. 19, no. 1, feb 2020

    2020

  46. [55]

    Dynamic dnn model selection and inference off loading for video analytics with edge-cloud collaboration,

    X. Wang et al., “Dynamic dnn model selection and inference off loading for video analytics with edge-cloud collaboration,” in Proceedings of the 32nd Workshop on Network and Operating Systems Support for Digital Audio and Video . New York, NY , USA: Association for Computing M...

    2022

  47. [56]

    Neurosurgeon: Collaborative intelligence between the cloud and mobile edge,

    Y . Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and L. Tang, “Neurosurgeon: Collaborative intelligence between the cloud and mobile edge,” in Proceedings of the Twenty-Second International Conference on Architectural Support for Programming Languages and Operati...

    2017

  48. [57]

    Chameleon: Scalable adaptation of video analytics,

    J. Jiang et al., “Chameleon: Scalable adaptation of video analytics,” in Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication , ser. SIGCOMM ’18. New York, NY , USA: Association for Computing Machinery, 2018, p. 253–266

    2018

  49. [58]

    Ekya: Continuous learning of video analytics models on edge compute servers,

    R. Bhardwaj et al. , “Ekya: Continuous learning of video analytics models on edge compute servers,” in 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22) . Renton, W A: USENIX Association, Apr. 2022, pp. 119–135. XI. A PPENDIX A. Scheduling Hardness...

    2022

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.