Pith. sign in

REVIEW 4 major objections 5 minor 37 references

MARCO: Hardware-Aware Neural Architecture Search for Edge Devices with Multi-Agent Reinforcement Learning and Conformal Prediction Filtering

T0 review · 4 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 MARCO cuts hardware-aware NAS search time 3–4× on edge chips with accuracy within 0.3% of a once-for-all supernet baseline.

desk verdict A useful HW-NAS framework with a credible 3-4x speedup claim, but its accuracy-preservation headline is contradicted by its own CIFAR-100 numbers and the conformal guarantee is looser than stated. read the letter →

arxiv 2506.13755 v1 pith:SAISBTRW submitted 2025-06-16 cs.LG

Arya Fayyazi , Mehdi Kamal , Massoud Pedram This is my paper · ORCID
classification cs.LG
keywords hardware-awareNASmulti-agentreinforcementlearningconformalpredictionedgeAIquantizationTinyMLneuralarchitecturesearchmicrocontroller
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

MARCO is a hardware-aware neural architecture search (NAS) framework for microcontrollers with tight memory and latency budgets. Its central claim is that splitting the search into two cooperating reinforcement-learning agents—one choosing the macro-architecture and one choosing per-layer bit-widths—and filtering candidates with a conformal-prediction surrogate before expensive evaluation makes NAS 3–4 times faster than a once-for-all supernet baseline while staying within 0.3% of its accuracy. This matters because it removes the multi-day supernet pretraining step from edge-AI design and replaces it with a search that can be retargeted to a new chip by changing budget constants. The paper reports these numbers on MNIST, CIFAR-10, and CIFAR-100, and confirms that simulator latency estimates track measurements on a real microcontroller evaluation board to within 5%.

What carries the argument

The load-bearing mechanism is the conformal filter in front of a two-agent PPO loop. A surrogate $g(x(a))$ predicts the reward $R(a)$ from architecture features, and on a calibration set of evaluated designs one computes residuals $\varepsilon_i = |R(a_i) - g(x(a_i))|$ and takes the $(1-\delta)$-quantile as the offset $\alpha_{1-\delta}$; a candidate is discarded when $g(x(a)) + \alpha_{1-\delta} < \tau$. The reward itself is $R(a) = A(a) - \lambda\, T(a)/T_{\text{budget}} - \mu\,\mathbb{1}\{\text{Mem}(a) > M_{\text{budget}}\}$, where $A(a)$ comes from 5 epochs of partial training on 10% of the data and $T(a)$, $\text{Mem}(a)$ come from a hardware simulator. The centralized critic observes the whole partial architecture so both agents can learn to satisfy shared memory and latency budgets while each acts on its own local view.

What would settle it

Take a batch of architectures sampled by MARCO, rank them twice—once by the 5-epoch/10%-data partial-training proxy and once by full retraining on the complete dataset—and compare the orderings; if the proxy's top architectures are not near the top under full training, or if the Spearman correlation between the two scores is near zero, the central accuracy-maintenance claim fails.

Watch

Extended reading notes

Core claim

MARCO claims that architecture search for memory-constrained edge devices can be decomposed into two cooperative agents trained under a centralized-critic, decentralized-execution scheme: a hardware configuration agent decides layer count, kernel sizes, channel widths, and skip/pooling choices, while a quantization agent assigns each layer either 4-bit or 8-bit precision. A conformal-prediction surrogate then attaches an upper confidence bound to the predicted reward of each merged candidate, and any candidate whose bound falls below a threshold is discarded with a negative reward before partial training or hardware simulation. Over three image-classification datasets, this recipe cuts total search time to roughly 1.2–2.9 days versus 7–10 days for the once-for-all supernet baseline, keeps final accuracy within 0.3% of the baseline (e.g., 87.2% vs 87.5% on CIFAR-10), reduces inference latency slightly, and the best simulated architectures reproduce on-chip latency within 2.4–4.0%.

Load-bearing premise

The whole speedup rests on the assumption that a network's score after 5 epochs of training on 10% of the data ranks candidate networks the same way full training would; if that proxy is noisy, the search is optimizing noise and the 0.3% accuracy-maintenance claim would not generalize.

Editorial extensions

If this is right

  • A designer targeting a sub-megabyte microcontroller can go from budget specification to a deployed network in roughly 1.6 days on CIFAR-10 instead of 7 days for the once-for-all supernet pipeline, at 0.3% accuracy cost.
  • The conformal filter prunes about 25–30% of candidates before any training or simulation, and this accounts for an extra 15–20% search-time reduction over the same multi-agent search without the filter.
  • Simulator latency tracks measured on-chip latency within 2.4–4.0% on the reported workloads, so the search's hardware objective is a reliable stand-in for real silicon.
  • Because MARCO only requires black-box latency and memory oracles, porting to a new edge platform consists of changing the design-knob ranges and the budget constants in the reward rather than re-running supernet pretraining.
  • The final discovered architectures, after full-dataset retraining, stay within 0.3% of the once-for-all baseline on MNIST, CIFAR-10, and CIFAR-100.

Reading between the lines

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

  • The same CP-filtered multi-agent recipe should transfer to untested NAS axes such as pruning ratios, operator types, or more than two quantization levels, since neither the reward decomposition nor the conformal bound depends on the specific 4-bit/8-bit choice.
  • The formal $1-\delta$ coverage guarantee assumes exchangeability between calibration and test candidates, which an iteratively updated policy can violate; the paper's periodic recalibration makes the guarantee empirical rather than exact, so a deployment needing the formal bound should run the search as an offline batch or re-verify coverage on the final candidate distribution.
  • If the 5-epoch, 10%-data partial-training score correlates strongly with full-training accuracy, MARCO's speedup could be combined with full retraining of the discovered architectures to close the remaining 0.3% gap; if it does not correlate, the conformal filter is pruning on a noisy signal and the accuracy-maintenance claim would be benchmark-specific.
  • The reported 3–4× speedup is relative to a once-for-all baseline that includes supernet pretraining days; a fair comparison for a practitioner who already has a trained supernet would isolate only the specialization/search phase, which this paper does not report.
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

4 major / 5 minor

Summary. The paper presents MARCO, a multi-agent reinforcement learning (MARL) framework for hardware-aware neural architecture search on edge devices, in which a hardware configuration agent and a quantization agent jointly propose architectures under memory and latency constraints. A conformal prediction (CP) surrogate is used to prune low-reward candidates before partial training and simulation. The authors report a 3–4x reduction in search time relative to an OFA-based baseline, accuracy within 0.3% of the baseline, and hardware validation on a MAX78000 board with simulator errors below 5%. The experiments cover MNIST, CIFAR-10, and CIFAR-100, with ablations on the CP threshold, miscoverage rate, and partial-training schedule.

Significance. If the claims hold, MARCO would be a practical contribution to TinyML NAS: it avoids expensive supernet pretraining, jointly optimizes architecture and mixed-precision quantization, and provides a principled early-filtering mechanism with statistical coverage. The paper includes several strengths that should be credited: real-hardware validation of latency estimates (Table X), ablations of CP coverage and partial-training epochs, a transparent hyperparameter table, and reproducibility details such as five seeds and paired t-tests. However, the central technical claims contain load-bearing weaknesses that must be addressed before the results can be accepted as stated.

major comments (4)
  1. [Sec. III.C.2 and Algorithm 2] The CP filter's feature vector in Eq. (12) includes "partial accuracy," but partial accuracy is defined in Sec. III.B.c as the result of 5-epoch, 10%-data partial training. Algorithm 2 (lines 10–14) applies the CP filter before partial training or simulation. This is contradictory: the surrogate cannot compute g(x(a)) for a new candidate without first running the very partial training the filter is supposed to avoid. Please remove partial accuracy from x(a), or explain how it can be obtained at negligible cost before the filter.
  2. [Sec. IV.A.4.a and Algorithm 1] The conformal calibration uses the same 100-architecture pool to train the surrogate g(·), compute residuals, and select the threshold τ. Algorithm 1 trains g on the calibration set and then evaluates residuals on that same set. Standard split-conformal guarantees (Eq. (10)) require that calibration residuals be independent of the model training data. As written, α1−δ is an in-sample quantile of training residuals and the stated coverage guarantee does not follow. Please use a proper train/calibration split (or a cross-conformal variant) and report empirical coverage on a held-out set of architectures.
  3. [Sec. IV.A.5, Table VI, and Abstract] The headline claim that MARCO maintains accuracy "within 0.3%" of the OFA baseline is contradicted by the CIFAR-100 row of Table VI: MARCO achieves 64.0±0.4% versus OFA's 64.8±0.3%, a gap of 0.8 percentage points. The text in Sec. IV.A.5 only cites the CIFAR-10 example (87.2 vs. 87.5) and does not explain the CIFAR-100 discrepancy. This is an internal inconsistency in the central accuracy-preservation claim. Revise the claim to per-dataset accuracy differences or provide an explicit explanation for CIFAR-100.
  4. [Sec. IV.A.3.a and Table VI] The comparison of search time between MARCO and the OFA baseline needs clarification. For OFA, the reported 7.0 days (MNIST, CIFAR-10) and 10.0 days (CIFAR-100) appear to include supernet pretraining, whereas MARCO's reported time is the end-to-end search with partial training of candidates. If OFA's post-pretraining specialization is much faster, the 3–4x speedup may be an apples-to-oranges comparison. Please define precisely what the baseline search time includes and confirm that the speedup holds under the same definition of search task.
minor comments (5)
  1. [Eq. (7) and Fig. 5] The reward in Eq. (7) is on the accuracy scale (e.g., about 87 for CIFAR-10), whereas Fig. 5 shows top-5 average reward around 0.65. Clarify whether the plotted reward is normalized or transformed, and specify the transformation.
  2. [Tables V and IX] Tables V and IX appear to be duplicates, both showing the effect of 5 vs. 10 partial-training epochs on CIFAR-10. Please remove one of them and reconcile the references in the text.
  3. [Algorithm 2 and Fig. 1 caption] The text contains several formatting artifacts: "Eg. 7" should be "Eq. (7)", and the Fig. 1 caption contains corrupted glyphs such as "♂¶icrochip", "/cubes", and "/d⌢llar-sign". These should be corrected.
  4. [Sec. I] The claim of being "the first" multi-agent HW-NAS formulation is too strong without a more thorough comparison to existing MARL-based NAS work; please temper the novelty claim or provide a more complete related-work discussion.
  5. [Sec. IV.A.5] The paired t-test p-values in Table VI are reported against multiple baselines, but the text does not specify exactly which quantity (accuracy, latency, or search time) is being compared. State the null hypothesis and the tested quantity for each p-value.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: MARCO's headline speed/accuracy results are empirical ablations, and the CP guarantee is a standard conformal theorem rather than a self-referential reduction.

full rationale

The central claims rest on direct measurements: Table VI compares measured search time and retrained accuracy, and the CP contribution is isolated by the MARL-no-CP ablation in Table VII. The conformal filter uses a surrogate g fit to calibration rewards and an offset alpha from residual quantiles, but Eq. (10) is a standard distribution-free coverage statement; choosing tau on the same calibration pool (Sec. IV-A-4) is hyperparameter selection, not an equation-level reduction, and it does not make the search-time speedup tautological. The two self-citations ([4], [15]) are not load-bearing: CP is also cited to external references [5], [6], and [4] appears only in a related-work list. Two non-circular concerns should be flagged: (1) Sec. IV.A.5 claims accuracy 'within 0.3% of OFA' citing only CIFAR-10 (87.2 vs 87.5), but Table VI shows a 0.8 pp gap on CIFAR-100 (64.0 vs 64.8), an internal inconsistency that does not change the circularity verdict; (2) Eq. (12) places 'partial accuracy' in the CP feature vector even though Algorithm 2 applies the filter before partial training, an implementation inconsistency worth correcting. Neither is a circular derivation.

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

No new physical entities are introduced. The framework's novelty rests on combining existing methods; the key unproven inputs are the partial-training proxy and the CP exchangeability assumption.

free parameters (4)
  • lambda (latency penalty) = 0.2
    Set empirically in Eq. (7) to balance accuracy against latency in the reward signal.
  • mu (memory overflow penalty) = 10
    Chosen to strongly discourage memory overflows in Eq. (7).
  • tau (CP filter threshold) = 5.5
    Selected by sweeping {5.0, 5.5, 6.0} on the calibration pool, then kept across datasets (Sec IV-A-4).
  • PPO hyperparameters = epsilon 0.2, learning rate 0.0005, gamma 0.99
    Tuned via 200-trial Bayesian optimization, though they converge to recommended defaults.
assumptions (3)
  • domain assumption Exchangeability of calibration and test architectures for conformal prediction
    Stated in Sec III-C-a as required for the Eq. (10) guarantee, but the sequential RL updates and online recalibration violate it.
  • domain assumption 5-epoch/10%-data partial training accuracy A(a) is a sufficient proxy for final accuracy
    Assumed throughout the reward function (Eq. 7) and only checked indirectly by comparing to a 10-epoch schedule.
  • domain assumption The vendor cycle-accurate simulator faithfully reproduces hardware latency and memory
    Validated on only three architectures with gaps 2.4-4.0% (Table X), but assumed across the full search space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MARCO: Hardware-Aware Neural Architecture Search for Edge Devices with Multi-Agent Reinforcement Learning and Conformal Prediction Filtering." pith.science (2026). https://pith.science/paper/SAISBTRW

@misc{pith2026250613755,
  author       = {Pith},
  title        = {Pith review of: MARCO: Hardware-Aware Neural Architecture Search for Edge Devices with Multi-Agent Reinforcement Learning and Conformal Prediction Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SAISBTRW}},
  note         = {Machine review of arXiv:2506.13755}
}
read the original abstract

This paper introduces MARCO (Multi-Agent Reinforcement learning with Conformal Optimization), a novel hardware-aware framework for efficient neural architecture search (NAS) targeting resource-constrained edge devices. By significantly reducing search time and maintaining accuracy under strict hardware constraints, MARCO bridges the gap between automated DNN design and CAD for edge AI deployment. MARCO's core technical contribution lies in its unique combination of multi-agent reinforcement learning (MARL) with Conformal Prediction (CP) to accelerate the hardware/software co-design process for deploying deep neural networks. Unlike conventional once-for-all (OFA) supernet approaches that require extensive pretraining, MARCO decomposes the NAS task into a hardware configuration agent (HCA) and a Quantization Agent (QA). The HCA optimizes high-level design parameters, while the QA determines per-layer bit-widths under strict memory and latency budgets using a shared reward signal within a centralized-critic, decentralized-execution (CTDE) paradigm. A key innovation is the integration of a calibrated CP surrogate model that provides statistical guarantees (with a user-defined miscoverage rate) to prune unpromising candidate architectures before incurring the high costs of partial training or hardware simulation. This early filtering drastically reduces the search space while ensuring that high-quality designs are retained with a high probability. Extensive experiments on MNIST, CIFAR-10, and CIFAR-100 demonstrate that MARCO achieves a 3-4x reduction in total search time compared to an OFA baseline while maintaining near-baseline accuracy (within 0.3%). Furthermore, MARCO also reduces inference latency. Validation on a MAX78000 evaluation board confirms that simulator trends hold in practice, with simulator estimates deviating from measured values by less than 5%.

Figures

Figures reproduced from arXiv: 2506.13755 by the authors.

Figure 1
Figure 1. MARCO workflow. At each time step, the agents (> HCA and ó QA) propose a candidate, which is merged and sent to a Z CP filter. If the upper confidence bound is below threshold τ, it is q discarded with r = −1. Otherwise, partial training estimates accuracy, ø a hardware simulator returns latency and memory, then reward computation and 8 centralized critic updates converge policies via L actor–critic steps. Our workf… view at source ↗
Figure 2
Figure 2. illustrates the CP filtering mechanism. D. Algorithmic Integration of MARL and CP Algorithm 2 provides the complete pseudocode for MARCO, integrating multi-agent RL with CP filtering. The surrogate model predicts the reward for each candidate gener￾ated by the HCA and QA. If the CP condition g(x(at)) + α1−δ < τ holds, the candidate is immediately assigned rt = −1 (discarded) and no further evaluation is performed. P… view at source ↗
Figure 3
Figure 3. Effect of CP Filtering in MARCO. Left: Before CP filtering, 20 candidates (blue dots) populate the search space based on their predicted reward g(x(a)). Right: After CP filtering, only candidates with g(x(a)) + 0.8 ≥ 5.5 (green dots) remain, significantly reducing the evaluation burden. B. Validation on MAX78000 Evaluation Board To verify that simulator gains translate to silicon, we flash the top-reward architectur… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Top-5 average reward progression vs. search time on CIFAR-100. MARCO converges to high reward within 2–3 days, in contrast with the nearly 10 days required by the OFA-based pipeline. TABLE X. Simulator vs. on-device latency for the best MARCO architecture on each datas…
Figure 4
Figure 4. Figure 4: Accuracy–latency Pareto front for CIFAR-10 on MAX78000. MARCO solutions (star markers) dominate the optimal trade-off region. TABLE VIII. Effect of varying CP miscoverage δ for CIFAR-100. δ Accuracy (%) Time (days) % Discard Miscoverage (%) 0.05 64.1 ± 0.3 3.2 ± 0.1 20…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [1]

    Once for all: Train one network and specialize it for efficient deployment,

    H. Cai, L. Gan, T. Wang, Z. Zhang, and S. Han, “Once for all: Train one network and specialize it for efficient deployment,” in ICLR, 2020

    work page 2020

  2. [2]

    MCUNet: Tiny deep learning on IoT devices,

    J. Lin, W.-M. Chen, Y . Lin, C. Gan, and S. Han, “MCUNet: Tiny deep learning on IoT devices,” in NeurIPS, 2020

    work page 2020

  3. [3]

    MnasNet: Platform-aware neural architecture search for mobile,

    M. Tan, B. Chen, R. Pang, et al., “MnasNet: Platform-aware neural architecture search for mobile,” in CVPR, 2019

    work page 2019

  4. [4]

    Dynamic Co- Optimization Compiler: Leveraging Multi-Agent Reinforcement Learn- ing for Enhanced DNN Accelerator Performance

    Arya Fayyazi, Mehdi Kamal, and Massoud Pedram. Dynamic Co- Optimization Compiler: Leveraging Multi-Agent Reinforcement Learn- ing for Enhanced DNN Accelerator Performance. In Proceedings of the 30th Asia and South Pacific Design Automation Conference (ASPDAC ’25), pages 16–22, Tokyo, Japan, 2025. Association for Computing Machinery. https://doi.org/10.114...

    arXiv 2025

  5. [5]

    A gentle introduction to confor- mal prediction and distribution-free uncertainty quantification,

    A. N. Angelopoulos and S. Bates, “A gentle introduction to confor- mal prediction and distribution-free uncertainty quantification,” arXiv preprint arXiv:2107.07511, 2021

    arXiv 2021

  6. [6]

    A tutorial on conformal prediction,

    G. Shafer and V . V ovk, “A tutorial on conformal prediction,” Journal of Machine Learning Research , 9, 2008

    work page 2008

  7. [7]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint, 2017

    work page 2017

  8. [8]

    Random search for hyper-parameter opti- mization,

    J. Bergstra and Y . Bengio, “Random search for hyper-parameter opti- mization,” Journal of Machine Learning Research, vol. 13, pp. 281–305, 2012

    work page 2012

Show all 37 references
  1. [9]

    Practical Bayesian opti- mization of machine learning algorithms,

    J. Snoek, H. Larochelle, and R. P. Adams, “Practical Bayesian opti- mization of machine learning algorithms,” in NeurIPS, 2012

    2012

  2. [10]

    BOHB: Robust and efficient hyperparameter optimization at scale,

    S. Falkner, A. Klein, and F. Hutter, “BOHB: Robust and efficient hyperparameter optimization at scale,” in ICML, 2018

    2018

  3. [11]

    What matters for on-policy reinforcement learning? A large-scale empirical study,

    M. Andrychowicz et al. , “What matters for on-policy reinforcement learning? A large-scale empirical study,” in ICLR, 2021

    2021

  4. [12]

    MAX78000: Ultra-low-power CNN microcon- troller with integrated accelerator—Datasheet,

    Analog Devices Inc., “MAX78000: Ultra-low-power CNN microcon- troller with integrated accelerator—Datasheet,” Analog Devices, 2020

    2020

  5. [13]

    Designing with the MAX78000 CNN Microcon- troller—Application Note,

    Analog Devices Inc., “Designing with the MAX78000 CNN Microcon- troller—Application Note,” AN-1234, 2021

    2021

  6. [14]

    MaximAI Development Toolkit User Guide,

    Analog Devices Inc., “MaximAI Development Toolkit User Guide,” Version 2.0, 2021

    2021

  7. [15]

    FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems

    Arya Fayyazi, Mehdi Kamal, and Massoud Pedram. FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems. arXiv preprint arXiv:2502.02966, 2025

    2025 arXiv

  8. [16]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proc. IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  9. [17]

    DeepArchitect: Automatically designing and training deep architectures,

    R. Negrinho and G. Gordon, “DeepArchitect: Automatically designing and training deep architectures,” in ICLR, 2017

    2017

  10. [18]

    HAWQ-V2: Hessian aware trace-weighted quantization of neural networks,

    Z. Dong, Z. Y . Yan, Q. Yang, and H. Jin, “HAWQ-V2: Hessian aware trace-weighted quantization of neural networks,” in NeurIPS, 2020

    2020

  11. [19]

    NAS-Bench-Suite: NAS evaluation is (now) surprisingly easy,

    Y . Lu, J. Jiang, and R. Gomez, “NAS-Bench-Suite: NAS evaluation is (now) surprisingly easy,” in ECCV, 2022

    2022

  12. [20]

    FBNetV2: Differentiable neural architecture search for spatial and channel dimensions,

    A. Wan, X. Dai, P. Zhang, et al., “FBNetV2: Differentiable neural architecture search for spatial and channel dimensions,” in CVPR, 2020

    2020

  13. [21]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” Technical report, University of Toronto, 2009

    2009

  14. [22]

    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

  15. [23]

    SqueezeNet: AlexNet- level accuracy with 50x fewer parameters and ¡0.5MB model size,

    F. N. Iandola, S. Han, M. W. Moskewicz, et al., “SqueezeNet: AlexNet- level accuracy with 50x fewer parameters and ¡0.5MB model size,”arXiv preprint arXiv:1602.07360, 2016

    2016 arXiv

  16. [24]

    MobileNets: Efficient convolutional neural networks for mobile vision applications,

    A. G. Howard, M. Zhu, B. Chen, et al., “MobileNets: Efficient convolutional neural networks for mobile vision applications,” arXiv preprint arXiv:1704.04861, 2017

    2017 arXiv

  17. [25]

    ShuffleNet: An extremely efficient convolutional neural network for mobile devices,

    X. Zhang, X. Zhou, M. Lin, and J. Sun, “ShuffleNet: An extremely efficient convolutional neural network for mobile devices,” in CVPR, 2018

    2018

  18. [26]

    R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. MIT press, 2018

    2018

  19. [27]

    Multi-agent deep reinforcement learning: A survey,

    S. Gronauer and K. Diepold, “Multi-agent deep reinforcement learning: A survey,” Artificial Intelligence Review , vol. 55, no. 2, pp. 895–943, 2022

    2022

  20. [28]

    An overview of multi-agent reinforce- ment learning from game theoretical perspective,

    Y . Yang, R. Zhang, et al., “An overview of multi-agent reinforce- ment learning from game theoretical perspective,” arXiv preprint arXiv:2008.03903, 2020

    2008 arXiv

  21. [29]

    Counterfactual multi-agent policy gradients,

    J. Foerster, G. Farquhar, T. Afouras, et al., “Counterfactual multi-agent policy gradients,” in AAAI, 2018

    2018

  22. [30]

    The surprising effectiveness of centralized training for multi-agent reinforcement learning,

    C. Yu, Y . Zhou, et al., “The surprising effectiveness of centralized training for multi-agent reinforcement learning,” in ICLR, 2022

    2022

  23. [31]

    A review of centralized training with decentralized execution for multi-agent reinforcement learning,

    L. Carmack and H. Zha, “A review of centralized training with decentralized execution for multi-agent reinforcement learning,” arXiv preprint arXiv:2105.00626, 2021

    2021 arXiv

  24. [32]

    FBNet: Hardware-aware efficient convnet design via differentiable neural architecture search,

    B. Wu, X. Dai, P. Zhang, et al., “FBNet: Hardware-aware efficient convnet design via differentiable neural architecture search,” in CVPR, 2019

    2019

  25. [33]

    ProxylessNAS: Direct neural architecture search on target task and hardware,

    H. Cai, L. Zhu, and S. Han, “ProxylessNAS: Direct neural architecture search on target task and hardware,” in ICLR, 2019

    2019

  26. [34]

    TVM: An automated end-to-end optimizing compiler for deep learning,

    T. Chen, T. Moreau, Z. Jiang, et al., “TVM: An automated end-to-end optimizing compiler for deep learning,” in OSDI, 2018

    2018

  27. [35]

    Fast inference of deep neural networks in FPGAs for particle physics,

    J. Duarte et al., “Fast inference of deep neural networks in FPGAs for particle physics,” Journal of Instrumentation , vol. 13, no. 7, P07027, 2018

    2018

  28. [36]

    Optimizing FPGA-based accelerator design for deep convolutional neural networks,

    C. Zhang, P. Li, G. Sun, Y . Guan, B. Xiao, and J. Cong, “Optimizing FPGA-based accelerator design for deep convolutional neural networks,” in FPGA, 2015

    2015

  29. [37]

    SODA: Stochastic superoptimization and design acceleration for DNNs on embedded devices,

    J. Wang, B. Wang, et al., “SODA: Stochastic superoptimization and design acceleration for DNNs on embedded devices,” in DAC, 2021

    2021

Pith tools

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