Pith. sign in

REVIEW 4 major objections 5 minor 48 references

INSIGHT: A Survey of In-Network Systems for Intelligent, High-Efficiency AI and Topology Optimization

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 The paper argues that AI computation — gradient aggregation during training and line-rate inference — can move off servers into programmable switches, routers, and NICs, and that the migration now has a stable, mappable design space.

desk verdict A useful but flawed survey: the taxonomy and tables are solid, while unsourced comparative figures and a title/body mismatch need fixing before it can serve as a trustworthy reference. read the letter →

arxiv 2505.24269 v2 pith:7S44XFY3 submitted 2025-05-30 cs.NI cs.AI

Aleksandr Algazinov , Joydeep Chandra , Matt Laing This is my paper · ORCID
classification cs.NIcs.AI
keywords in-networkcomputationdistributedlearningsoftware-definednetworkingprogrammabledataplanesmodelcompressionaggregationfederatededgeAI
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

This survey sets out to establish that AI computation can be moved off end hosts and into the network fabric itself — programmable switches, routers, and NICs — and that this migration is both technically feasible and worth doing. Its central claim is that the field has a coherent design space, organized along four axes: mapping ML models onto memory-constrained switch pipelines, compressing models to fit those pipelines, aggregating gradients inside the network for distributed training, and simplifying development through frameworks such as Planter and Quark. If the synthesis is right, the practical takeaway is that the communication bottleneck in distributed AI training is addressable at the switch, not just at the host, and that a meaningful class of inference tasks can run at line rate on devices previously judged too weak for machine learning. The paper further claims that applications such as network monitoring, intrusion detection, traffic management, and edge AI already demonstrate the payoff, and it maps the open problems — runtime programmability, standardized benchmarks, support for larger models — that gate further progress.

What carries the argument

The load-bearing object is the programmable data plane (PDP): a P4-programmable switch pipeline with fixed stages, SRAM and TCAM memory, and bitwise arithmetic that becomes the execution substrate for AI. The argument is carried by mapping techniques that translate trained models into this substrate — depth-based and encode-based decision-tree layouts, bitwise XNOR-and-popcount deployments for binary neural networks, and lookup tables for statistical models — together with quantization that substitutes for the missing floating-point unit. The second mechanism is in-network aggregation: programmable switches that sum gradients from many workers mid-flight, using chunking, fixed-point arithmetic, and switch-side scoreboarding for reliability, which turns the communication bottleneck into a switch-resource problem. Frameworks are the packaging machinery: Planter compiles many model families into switch executables with one-click deployment, and Quark runs pruned, quantized convolutional neural networks entirely on the data plane at line rate.

What would settle it

A direct audit against the primary literature would settle the survey's reliability: check whether the cited SwitchML paper (NSDI 2021) actually reports the 'up to 3 times' and 'up to 5.5 times' training-time reductions credited in Section 4.2, whether GOAT and Serene support the 'up to 40%' figure, and whether any published comparison backs Fig. 1's ranking of SwitchML as best. If those numbers cannot be located in the cited papers, or the comparisons rest on unrelated testbeds, the survey's central comparative claims fail; if they check out, the synthesis stands.

Watch

Extended reading notes

Core claim

The paper's central claim, stated on its own terms, is that in-network computation is the convergence point of two movements — 'AI for Networking' and 'Networking for AI' — and that this convergence is now mature enough to be surveyed as a coherent field. It argues that programmable data planes, particularly P4-programmable switch ASICs, make it possible to run real machine-learning workloads where packets flow, provided three problems are solved: models must be mapped onto fixed, shallow pipelines with no native floating-point support; models must be compressed (via pruning, quantization, low-rank factorization, knowledge distillation, and architecture search) to fit switch memory; and distributed training must be reorganized so that switches aggregate gradients in transit rather than shipping them to a central parameter server. Empirically, the survey asserts that in-network aggregation frameworks deliver measured gains — SwitchML reduces training time by up to 3 times and up to 5.5 times in optimized scenarios, while GOAT and Serene show up to 40% reductions — and that frameworks like Planter and Quark turn these techniques into usable tooling. The survey's overall assertion is that intelligent, efficient, responsive networks are achievable by embedding AI in the fabric itself.

Load-bearing premise

The survey is only as reliable as its reporting of other people's measurements: if the speedups it attributes to SwitchML, GOAT, and Serene are misquoted, or come from incomparable experimental setups, its comparative conclusions lose their ground.

Editorial extensions

If this is right

  • If the survey's synthesis is accurate, the synchronization bottleneck in distributed AI training shifts from network bandwidth to switch resources, making in-network gradient aggregation a primary lever for faster training.
  • A practical class of models — decision trees, binary neural networks, statistical classifiers, and compressed CNNs — can be deployed at line rate on production switch ASICs, making line-rate inference a realistic design target for monitoring and security systems.
  • Framework-level automation of model-to-pipeline compilation means in-network ML no longer requires simultaneous expertise in machine learning and P4 programming; one-click deployment becomes a viable workflow.
  • The reported gains — up to 3x to 5.5x training-time reduction from in-network aggregation and sub-100 microsecond inference for anomaly detection — suggest the remaining obstacles are engineering problems such as runtime programmability and benchmarks, not fundamental feasibility.
  • Federated learning combined with in-network aggregation becomes a more plausible route to privacy-preserving training at scale, because intermediate aggregation at switches reduces both server load and communication latency.

Reading between the lines

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

  • Editorial note: despite the title's mention of topology optimization, the body of the paper never treats network topology optimization; the survey's actual scope is model mapping, compression, in-network aggregation, and development frameworks.
  • The comparison of aggregation frameworks is presented without the underlying experimental setup, so the natural next step — one the paper itself endorses — is a standardized benchmark suite that re-measures SwitchML, GOAT, PANAMA, and Serene on a single testbed.
  • The convergence thesis implies a testable trajectory: as programmable-switch resources grow, the set of deployable model families should expand along the compression frontier the survey maps.
  • The paper's repeated emphasis on missing floating-point support suggests that a widely supported low-precision arithmetic standard for data-plane languages could unlock a large part of this design space; the survey gestures at the problem but leaves that engineering implication implicit.
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 manuscript is a survey of in-network computation for AI workloads. It covers the evolution from SDN to programmable data planes, methods for mapping ML models onto network devices, model compression, distributed training with in-network aggregation, frameworks such as Planter and Quark, applications in monitoring, security, traffic management, and edge AI, and future research directions. The abstract and title also announce topology optimization, but the body does not contain a section on that topic. The central claim is that the paper provides a comprehensive and reliable synthesis of the field.

Significance. If the synthesis were properly sourced, the survey would be a useful entry point for researchers and practitioners: it assembles a broad set of recent references, gives a structured account of mapping and compression techniques, and identifies concrete open problems such as runtime programmability and standardized benchmarks. The paper makes no formal derivations, so machine-checkable proofs are not applicable; the relevant verification burden is fidelity to primary sources. That burden is not currently met because several comparative claims rest on figures without provenance, which weakens the survey's reliability as a map of the literature.

major comments (4)
  1. [Section 4.2, Fig. 1] The sentence 'SwitchML performs the best in the speed up scenarios compared to the other frameworks like GOAT, PANAMA, and Serene' is presented as a comparative ranking, but Fig. 1 has no source, axis labels, units, or experimental methodology. The four systems come from different papers with different synchronization models, workloads, and switch hardware, and no cited reference appears to benchmark all four under identical conditions. Because the survey's purpose is to provide a reliable synthesis, this unsupported ranking is load-bearing. Please remove the ranking, or replace it with a provenance table (system, workload, hardware, metric, source), or explicitly rephrase it as an informal observation and cite each quantitative result. The adjacent numeric claims (up to 3x, 5.5x, and 40% reductions) also need inline citations and a statement of the conditions under which they were reported.
  2. [Section 5.2, Fig. 2] The claim that 'in-network computation can surpass the performance of the traditional server-based systems in both latency and throughput parameters' is not supported by any cited source. Fig. 2 has no caption provenance, axes, units, or methodology. It is unclear whether the figure is reproduced from a specific paper, is schematic, or is the authors' own measurement. Please provide the source, or, if the figure is illustrative, say so explicitly and remove the performance comparison. This is the same provenance problem as in Fig. 1 and directly affects the paper's central promise of an accurate synthesis.
  3. [Title and scope] The title promises 'Topology Optimization', but no section, subsection, table, or figure in the body defines or discusses topology optimization. If topology optimization is intended as a contribution, the paper needs a dedicated section (for example, network topology design for AI traffic or placement of aggregation points). If it is not, the title should be changed to match the actual scope. As written, the title overstates the coverage and misleads readers about the survey's contents.
  4. [Section 6.1] The text states that N-BaIoT [9] 'employs deep autoencoders to flag anomalous IoT flows' and 'has demonstrated sub-100us inference times and high detection performance in real deployments.' The cited paper is a study introducing a dataset and a detection method for IoT botnets; it is not a programmable-switch framework, and the sub-100us inference claim does not appear to be supported by that source. Please verify the claim and correct the attribution, or remove the specific performance figure.
minor comments (5)
  1. [Section 3.2] The text says 'deeper architectures like conventional neural networks often go beyond hardware capabilities'; in context this should read 'convolutional neural networks'.
  2. [Figure 1 caption] The caption 'Comparison of Various In-Network Aggregations Frameworks' should read 'Aggregation Frameworks'.
  3. [References] References [21] and [22] both list the NSDI 2021 paper by Sapio et al.; they should be merged into a single entry with complete venue information.
  4. [Section 4.3] The sentence introducing federated learning cites only [26], which is a study of in-network computation for wireless edge networks; a more general FL reference would be helpful for readers unfamiliar with the paradigm.
  5. [Abstract and body text] There are numerous line-break and spacing artifacts (for example, 'In-networkcomputation' in the abstract) that should be cleaned up during copyediting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a survey that restates external literature and introduces no fitted parameters, predictions, or self-cited load-bearing results.

full rationale

This is a survey/taxonomy paper with no mathematical derivation, no fitted parameters, and no predictive model. I checked each circularity pattern in priority order. (1) No quantity is defined in terms of another quantity and then presented as a prediction; section-level claims such as 'SwitchML reduced training time by up to 3 times' are citations of external empirical results, not derivations from inputs. (2) No parameter is fitted to a subset of data and then 'predicted'; the paper performs no fitting at all. (3) The authors do not cite their own prior work as load-bearing support; the cited surveys and systems papers are by other research groups and are independent secondary or primary sources. (4) No uniqueness theorem is imported from the authors' prior work. (5) No ansatz is smuggled in via citation; the frameworks and compression techniques described are attributed to external references. (6) No known result is renamed as organization or unification; the paper is an ordinary literature synthesis. The main weaknesses are provenance and scope, not circularity: Fig. 1 asserts 'SwitchML performs the best' and Fig. 2 asserts latency/throughput advantages, but neither figure carries a source, axis labels, units, or methodology, so the comparative conclusions are not independently verifiable from the paper; likewise the title promises 'Topology Optimization' but the body never addresses it. These are correctness/completeness concerns about unsupported reporting, not cases where an output reduces to its input by construction. Because the survey is self-contained as a secondary source and no load-bearing step is definitionally circular or dependent on a self-citation chain, the circularity score is 0.

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

No derivation or fitted parameters are present. The survey rests on the accuracy of secondary reporting and on representative coverage; these are domain assumptions about the cited literature rather than mathematical axioms.

assumptions (3)
  • domain assumption Cited empirical results (SwitchML speedups, GOAT/Serene reductions, Quark latency, detection accuracies) are accurately transcribed from the primary sources.
    The survey makes no primary measurements; all quantitative claims come from secondary reporting of other papers, e.g., Section 4.2 and Section 6.1.
  • domain assumption The selection of frameworks and applications (SDN, PDPs, Planter, Quark, SwitchML, GOAT, PANAMA, Serene, N-BaIoT) is representative of the field.
    The 'comprehensive' framing depends on inclusion choices; no systematic search or inclusion/exclusion criteria are described.
  • domain assumption Categorizations in Tables 1 and 2 correctly map each technique and framework to its properties.
    The tables' rows are qualitative summaries; errors in mapping would propagate to the survey's guidance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of INSIGHT: A Survey of In-Network Systems for Intelligent, High-Efficiency AI and Topology Optimization." pith.science (2026). https://pith.science/paper/7S44XFY3

@misc{pith2026250524269,
  author       = {Pith},
  title        = {Pith review of: INSIGHT: A Survey of In-Network Systems for Intelligent, High-Efficiency AI and Topology Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7S44XFY3}},
  note         = {Machine review of arXiv:2505.24269}
}
read the original abstract

In-network computation represents a transformative approach to addressing the escalating demands of Artificial Intelligence (AI) workloads on network infrastructure. By leveraging the processing capabilities of network devices such as switches, routers, and Network Interface Cards (NICs), this paradigm enables AI computations to be performed directly within the network fabric, significantly reducing latency, enhancing throughput, and optimizing resource utilization. This paper provides a comprehensive analysis of optimizing in-network computation for AI, exploring the evolution of programmable network architectures, such as Software-Defined Networking (SDN) and Programmable Data Planes (PDPs), and their convergence with AI. It examines methodologies for mapping AI models onto resource-constrained network devices, addressing challenges like limited memory and computational capabilities through efficient algorithm design and model compression techniques. The paper also highlights advancements in distributed learning, particularly in-network aggregation, and the potential of federated learning to enhance privacy and scalability. Frameworks like Planter and Quark are discussed for simplifying development, alongside key applications such as intelligent network monitoring, intrusion detection, traffic management, and Edge AI. Future research directions, including runtime programmability, standardized benchmarks, and new applications paradigms, are proposed to advance this rapidly evolving field. This survey underscores the potential of in-network AI to create intelligent, efficient, and responsive networks capable of meeting the demands of next-generation AI applications.

Figures

Figures reproduced from arXiv: 2505.24269 by the authors.

Figure 1
Figure 1. Comparison of Various In-Network Aggregations Frameworks 4.2 Collaborative In-Network Aggregation Strategies and Frameworks Various frameworks were introduced to support collaborative in-network aggre￾gation. A key challenge, however, lies in handling asynchronous gradient ar￾rivals caused by network latency among workers. For example, GOAT [27] em￾ploys a scheduling mechanism that partitions models into sub-compone… view at source ↗
Figure 2
Figure 2. Latency Reduction and Throughput Gains with In-Network Computation 5.1 Overview of Existing Frameworks: Planter and Quark [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 28 canonical work pages

  1. [9]

    N-BaIoT—Network-Based Detection of IoT Botnet Attacks Using Deep Autoencoders

    Yair Meidan et al. “N-BaIoT—Network-Based Detection of IoT Botnet Attacks Using Deep Autoencoders”. In:IEEE Pervasive Computing 17.3 (2018), pp. 12–22.doi: 10.1109/MPRV.2018.03367731

    arXiv 2018

  2. [1]

    Isolation Forest

    Fei Tony Liu, Kai Ting, and Zhi-Hua Zhou. “Isolation Forest”. In: Jan. 2009, pp. 413–422.doi: 10.1109/ICDM.2008.17

    doi:10.1109/icdm.2008.17 2009

  3. [2]

    Tanenbaum and David J

    Andrew S. Tanenbaum and David J. Wetherall.Computer Networks. 5th. USA: Prentice Hall Press, 2010.isbn: 0132126958.doi: 10.5555/1942194

    work page doi:10.5555/1942194 2010

  4. [3]

    Low-rank matrix factorization for Deep Neural Network training with high-dimensional output targets

    Tara N. Sainath et al. “Low-rank matrix factorization for Deep Neural Network training with high-dimensional output targets”. In:2013 IEEE International Conference on Acoustics, Speech and Signal Processing.2013, pp. 6655–6659. doi: 10.1109/ICASSP.2013.6638949

    arXiv 2013

  5. [4]

    The Algorithmic Foundations of Dif- ferential Privacy

    Cynthia Dwork and Aaron Roth. “The Algorithmic Foundations of Dif- ferential Privacy”. In:Foundations and Trends® in Theoretical Computer Science 9.3–4(2014),pp.211–407. doi: 10.1561/0400000042. url: https: //www.cis.upenn.edu/~aaroth/Papers/privacybook.pdf

    doi:10.1561/0400000042 2014

  6. [5]

    XGBoost: A Scalable Tree Boosting System

    Tianqi Chen and Carlos Guestrin. “XGBoost: A Scalable Tree Boosting System”. In:Proceedings of the 22nd ACM SIGKDD International Confer- ence on Knowledge Discovery and Data Mining. KDD ’16. San Francisco, California, USA: Association for Computing Machinery, 2016, pp. 785–

    work page 2016

  7. [6]

    Vishakh Hegde and Sheema Usmani.Parallel and Distributed Deep Learn- ing.Tech.rep.CME323:DistributedAlgorithmsandOptimization,Spring

    work page

  8. [7]

    BotGuard: Lightweight real-time botnet detection in soft- ware defined networks

    Jing Chen et al. “BotGuard: Lightweight real-time botnet detection in soft- ware defined networks”. In:Wuhan University Journal of Natural Sciences 22 (Apr. 2017), pp. 103–113.doi: 10.1007/s11859-017-1223-8

    work page doi:10.1007/s11859-017-1223-8 2017

Show all 48 references
  1. [8]

    Neural Architecture Search with Reinforce- mentLearning

    Barret Zoph and Quoc Le. “Neural Architecture Search with Reinforce- mentLearning”.In: International Conference on Learning Representations

  2. [10]

    Generating Long Sequences with Sparse Transformers

    Rewon Child et al. Generating Long Sequences with Sparse Transformers

  3. [11]

    Secure Multiparty Computation (MPC).CryptologyePrint Archive, Paper 2020/300

    YehudaLindell. Secure Multiparty Computation (MPC).CryptologyePrint Archive, Paper 2020/300. 2020. doi: 10 . 1145 / 3387108. url: https : //eprint.iacr.org/2020/300

    2020

  4. [12]

    An Overview of Neural Network Compression

    James O’ Neill. An Overview of Neural Network Compression. 2020. arXiv: 2006.03669 [cs.LG]. url: https://arxiv.org/abs/2006.03669

    2020 arXiv

  5. [13]

    Autonomous Vehicles

    Christoph Bartneck et al. “Autonomous Vehicles”. In:An Introduction to Ethics in Robotics and AI. Cham: Springer International Publishing, 2021, pp. 83–92. isbn: 978-3-030-51110-4. doi: 10.1007/978- 3- 030- 51110- 4_10. url: https://doi.org/10.1007/978-3-030-51110-4_10

    2021 doi

  6. [14]

    NetFC: Enabling Accurate Floating-point Arithmetic on Programmable Switches

    Penglai Cui et al. “NetFC: Enabling Accurate Floating-point Arithmetic on Programmable Switches”. In:2021 IEEE 29th International Conference on Network Protocols (ICNP). 2021, pp. 1–11.doi: 10.1109/ICNP52444. 2021.9651946

    2021

  7. [15]

    In-network Aggrega- tion for Shared Machine Learning Clusters

    Nadeen Gebara, Manya Ghobadi, and Paolo Costa. “In-network Aggrega- tion for Shared Machine Learning Clusters”. In:Proceedings of Machine Learning and Systems. Ed. by A. Smola, A. Dimakis, and I. Stoica. Vol. 3. 2021, pp. 829–844. url: https : / / proceedings . mlsys . org / p...

    2021

  8. [16]

    Knowledge Distillation: A Survey

    Jianping Gou et al. “Knowledge Distillation: A Survey”. In:International Journal of Computer Vision129.6 (Mar. 2021), pp. 1789–1819.issn: 1573-

    2021

  9. [17]

    P4Pi: P4 on Raspberry Pi for networking education

    Sándor Laki et al. “P4Pi: P4 on Raspberry Pi for networking education”. In: 51.3 (July 2021), pp. 17–21.issn: 0146-4833. doi: 10.1145/3477482. 3477486. url: https://doi.org/10.1145/3477482.3477486

    2021

  10. [18]

    The Programmable Data Plane: Abstractions, Ar- chitectures, Algorithms, and Applications

    Oliver Michel et al. “The Programmable Data Plane: Abstractions, Ar- chitectures, Algorithms, and Applications”. In:ACM Comput. Surv.54.4 (May 2021). issn: 0360-0300. doi: 10.1145/3447868. url: https://doi. org/10.1145/3447868

    2021 doi

  11. [19]

    A White Paper on Neural Network Quantization

    Markus Nagel et al. “A White Paper on Neural Network Quantization”. In: ArXiv abs/2106.08295 (2021). url: https://api.semanticscholar. org/CorpusID:235435934

    2021 arXiv

  12. [20]

    Network for AI and AI for Network: Challenges and Opportunities for Learning-Oriented Networks

    Jianping Pan et al. “Network for AI and AI for Network: Challenges and Opportunities for Learning-Oriented Networks”. In:IEEE Network 35.6 (2021), pp. 270–277.doi: 10.1109/MNET.101.2100118

    2021 doi

  13. [21]

    ScalingDistributedMachineLearningwithIn-Network Aggregation

    AmedeoSapioetal.“ScalingDistributedMachineLearningwithIn-Network Aggregation”. In: (Apr. 2021), pp. 785–808.url: https://www.usenix. org/conference/nsdi21/presentation/sapio. 16 Aleksandr Algazinov 1, Joydeep Chandra2, Matt Laing3

    2021

  14. [22]

    ScalingDistributedMachineLearningwithIn-Network Aggregation

    AmedeoSapioetal.“ScalingDistributedMachineLearningwithIn-Network Aggregation”. In: 18th USENIX Symposium on Networked Systems De- sign and Implementation (NSDI 21). USENIX Association, Apr. 2021, pp. 785–808. isbn: 978-1-939133-21-2. url: https://www.usenix.org/ conference/nsd...

    2021

  15. [23]

    Energy-Efficient Data Transfer Optimization via Decision-Tree Based Uncertainty Reduction

    Hasibul Jamil et al. “Energy-Efficient Data Transfer Optimization via Decision-Tree Based Uncertainty Reduction”. In:2022 International Con- ference on Computer Communications and Networks (ICCCN).2022,pp.1–

    2022

  16. [24]

    Software-Defined Networking (SDN): A Review

    Quadri Waseem et al. “Software-Defined Networking (SDN): A Review”. In: (2022), pp. 30–35.doi: 10.1109/ICOIACT55506.2022.9972067

    2022

  17. [25]

    IIsy: Practical In-Network Classification

    Changgang Zheng et al. IIsy: Practical In-Network Classification. 2022. arXiv: 2205.08243 [cs.NI]. url: https://arxiv.org/abs/2205.08243

    2022 arXiv

  18. [26]

    In-Network Computation for Large-Scale Fed- erated Learning Over Wireless Edge Networks

    Thinh Quang Dinh et al. “In-Network Computation for Large-Scale Fed- erated Learning Over Wireless Edge Networks”. In:IEEE Transactions on Mobile Computing22.10 (2023), pp. 5918–5932.doi: 10.1109/TMC.2022. 3190260

    2023 doi

  19. [27]

    GOAT:GradientSchedulingwithCollaborativeIn-Network Aggregation for Distributed Training

    JinFangetal.“GOAT:GradientSchedulingwithCollaborativeIn-Network Aggregation for Distributed Training”. In:2023 IEEE/ACM 31st Interna- tional Symposium on Quality of Service (IWQoS). 2023, pp. 1–10. doi: 10.1109/IWQoS57198.2023.10188783

    2023

  20. [28]

    A Review of the In-Network Computing and Its Role in the Edge-Cloud Continuum

    Manel Gherari et al. A Review of the In-Network Computing and Its Role in the Edge-Cloud Continuum. 2023. arXiv: 2312.00303 [cs.NI] . url: https://arxiv.org/abs/2312.00303

    2023 arXiv

  21. [29]

    doi: 10.1109/ICCCN54977.2022.9868866

    2022

  22. [30]

    AI-Driven Packet Forwarding With Programmable Data Plane: A Survey

    Wei Quan et al. “AI-Driven Packet Forwarding With Programmable Data Plane: A Survey”. In: IEEE Communications Surveys & Tutorials 25.1 (2023), pp. 762–790.doi: 10.1109/COMST.2022.3217613

    2023

  23. [31]

    AI-Driven Anomaly Detection in Network Monitoring Techniques and Tools

    Aakash Aluwala. “AI-Driven Anomaly Detection in Network Monitoring Techniques and Tools”. In:Journal of Artificial Intelligence & Cloud Com- puting (June 2024), pp. 1–6.doi: 10.47363/JAICC/2024(3)310

    2024 doi

  24. [32]

    Is AI a Trick or T(h)reat for SecurinA Programmable Data Planes?

    Enkeleda Bardhi, Mauro Conti, and Riccardo Lazzeretti. “Is AI a Trick or T(h)reat for SecurinA Programmable Data Planes?” In:IEEE Network 38.6 (2024), pp. 146–152.doi: 10.1109/MNET.2024.3451330

    2024

  25. [33]

    A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recom- mendations

    Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. “A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recom- mendations”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence 46.12 (2024), pp. 10558–10578. doi: 10.1109/TPAMI.2024. 3447085

    2024 doi

  26. [34]

    INTELLIGENTNETWORKOPTIMIZATION:REV- OLUTIONIZINGNETWORKMANAGEMENTTHROUGHAIANDML

    SaikatChoudhury.“INTELLIGENTNETWORKOPTIMIZATION:REV- OLUTIONIZINGNETWORKMANAGEMENTTHROUGHAIANDML”. In:INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING AND INSIGHT: A Survey of In-Network Systems 17 TECHNOLOGY 15 (Dec. 2024), pp. 2077–2085.doi: 10.34218/IJCET_ 15_06_178

    2024 doi

  27. [35]

    Programmable Data Plane Intelligence: Advances, Op- portunities, and Challenges

    Wai-Xi Liu et al. “Programmable Data Plane Intelligence: Advances, Op- portunities, and Challenges”. In:IEEE Network 37.5 (2023), pp. 122–128. doi: 10.1109/MNET.124.2200113

    2023 doi

  28. [36]

    Operationalizing AI/ML in Future Networks: A Bird’s EyeViewfromtheSystemPerspective

    Qiong Liu et al. “Operationalizing AI/ML in Future Networks: A Bird’s EyeViewfromtheSystemPerspective”.In: Comm. Mag.63.4(Sept.2024), pp. 176–182. issn: 0163-6804. doi: 10 . 1109 / MCOM . 001 . 2400033. url: https://doi.org/10.1109/MCOM.001.2400033

    2024 doi

  29. [37]

    Brain-on-switch: towards advanced intelligent network data plane via NN-driven traffic analysis at line-speed

    Jinzhu Yan et al. “Brain-on-switch: towards advanced intelligent network data plane via NN-driven traffic analysis at line-speed”. In:Proceedings of the 21st USENIX Symposium on Networked Systems Design and Imple- mentation. NSDI’24. Santa Clara, CA, USA: USENIX Association, 2...

    2024

  30. [38]

    In-NetworkMachineLearningUsingProgrammable Network Devices: A Survey

    ChanggangZhengetal.“In-NetworkMachineLearningUsingProgrammable Network Devices: A Survey”. In:IEEE Communications Surveys & Tuto- rials 26.2 (2024), pp. 1171–1200.doi: 10.1109/COMST.2023.3344351

    2024

  31. [39]

    Planter: Rapid Prototyping of In-Network Ma- chine Learning Inference

    Changgang Zheng et al. “Planter: Rapid Prototyping of In-Network Ma- chine Learning Inference”. In: SIGCOMM Comput. Commun. Rev. 54.1 (Aug. 2024), pp. 2–21.issn: 0146-4833. doi: 10.1145/3687230.3687232. url: https://doi.org/10.1145/3687230.3687232

    2024

  32. [40]

    No Worker Left (Too Far) Behind: Dynamic HybridSynchronizationforIn-NetworkMLAggregation

    Diego Cardoso Nunes et al. “No Worker Left (Too Far) Behind: Dynamic HybridSynchronizationforIn-NetworkMLAggregation”.In: International Journal of Network Management35.1 (2025). e2290 nem.2290, e2290.doi: https://doi.org/10.1002/nem.2290 . url: https://onlinelibrary. wiley.com...

    2025 doi

  33. [41]

    Accelerating Distributed Training With Collaborative In- Network Aggregation

    Jin Fang et al. “Accelerating Distributed Training With Collaborative In- Network Aggregation”. In:IEEE/ACM Transactions on Networking32.4 (2024), pp. 3437–3452.doi: 10.1109/TNET.2024.3387948

    2024

  34. [42]

    Quark: Implementing Convolutional Neural Networks Entirely on Programmable Data Plane

    Mai Zhang et al. Quark: Implementing Convolutional Neural Networks Entirely on Programmable Data Plane. 2025. arXiv:2501.15100 [cs.NI]. url: https://arxiv.org/abs/2501.15100

    2025 arXiv

  35. [47]

    Network traffic classification to improve quality of service (QoS)

    Ali Qasim Mohammed and Rana Ghani. “Network traffic classification to improve quality of service (QoS)”. In: Jan. 2025, p. 020007.doi: 10.1063/ 5.0264880

    2025

  36. [794]

    doi: 10.1145/2939672.2939785

    isbn: 9781450342322. doi: 10.1145/2939672.2939785. url: https: //doi.org/10.1145/2939672.2939785

  37. [1405]

    url: http://dx.doi.org/ 10.1007/s11263-021-01453-z

    doi: 10.1007/s11263- 021- 01453- z. url: http://dx.doi.org/ 10.1007/s11263-021-01453-z

    doi

  38. [2016]

    url: https : / / web

    Stanford University, 2016. url: https : / / web . stanford . edu / ~rezab/classes/cme323/S16/projects_reports/hedge_usmani.pdf

    2016

  39. [2017]

    INSIGHT: A Survey of In-Network Systems 15

    url: https://openreview.net/forum?id=r1Ue8Hcxg. INSIGHT: A Survey of In-Network Systems 15

  40. [2019]

    url: https://arxiv.org/abs/1904

    arXiv: 1904.10509 [cs.LG]. url: https://arxiv.org/abs/1904. 10509

    1904 arXiv

Pith tools

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