pith. sign in

arxiv: 2507.16594 · v2 · submitted 2025-07-22 · 💻 cs.NI · cs.AI· cs.DC

Optimizing Split Learning Latency in TinyML-Based IoT Systems

Zied Jenhani , Mounir Bensalem , Jasenka Dizdarevi\'c , Admela Jukan This is my paper

Pith reviewed 2026-05-19 03:24 UTC · model grok-4.3

classification 💻 cs.NI cs.AIcs.DC
keywords split learningTinyMLIoTESP-NOWbeam searchlatency optimizationESP32-S3wireless protocols
0
0 comments X

The pith

ESP-NOW protocol with beam search split-point optimization achieves the lowest end-to-end latency for TinyML split learning on ESP32-S3 devices.

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

The paper runs the first latency benchmarks for split learning on ESP32-S3 boards, comparing four wireless protocols across MobileNet-V2 and ResNet50 models. ESP-NOW records the shortest round-trip time of 3.6 seconds, while the authors' beam search algorithm selects split points that keep total inference time near the optimum and finishes its own search in 0.1 seconds even with five devices. A reader would care because the work turns split learning from a theoretical idea into something that can actually run on constrained sensors without long delays or heavy computation on the edge.

Core claim

Through direct measurements the authors show that ESP-NOW yields the lowest round-trip time of 3.6 seconds among UDP, TCP, ESP-NOW and BLE, and that their Beam Search-based algorithm for choosing model split points produces near-optimal end-to-end latency while requiring only 0.1 seconds of processing time for five devices, outperforming Greedy Search, First-Fit and Random-Fit and approaching the exhaustive Brute Force result.

What carries the argument

The Beam Search-based algorithm for split point optimization, which explores promising split locations across the model layers to jointly minimize communication and computation overhead under a chosen wireless protocol.

If this is right

  • ESP-NOW becomes the preferred protocol for latency-critical split learning on similar low-power hardware.
  • The beam search method lets practitioners optimize splits for small numbers of devices without exhaustive search.
  • Split-point choice trades off communication volume against local compute time differently for each model architecture.
  • End-to-end latency can be driven close to the theoretical minimum while keeping optimization overhead low.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same search technique could be extended to larger device counts if beam width and pruning rules are adjusted.
  • Adding energy cost as a second objective inside the beam search might produce splits that also save battery life.
  • The measured protocol ranking could guide protocol selection in other edge-AI settings that use constrained radios.

Load-bearing premise

The assumption that latency numbers measured on ESP32-S3 boards under the authors' controlled test conditions without interference or changing network loads will match real IoT deployments.

What would settle it

Re-running the identical protocol and model tests while adding realistic wireless interference or variable background traffic and observing whether the reported 3.6 s RTT and 0.1 s optimization time still hold.

Figures

Figures reproduced from arXiv: 2507.16594 by Admela Jukan, Jasenka Dizdarevi\'c, Mounir Bensalem, Zied Jenhani.

Figure 1
Figure 1. Figure 1: SL-TinyML framework architecture and communication processes between IoT devices [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed Mobilenet-V2 architecture and split points. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Split learning (SL) addresses the limitation of running deep learning inference directly on low-power edge/IoT nodes, in which it executes part of the inference process on the sensor and offloading the remainder to a companion device. Despite its promise, the inference latency of SL on constrained hardware under realistic low-power wireless protocols remains unexplored. This paper presents the first experimental latency benchmark of TinyML-based SL on ESP32-S3 boards, comparing four wireless communication protocol solutions (UDP, TCP, ESP-NOW, BLE). We also analyze the impact of the choice of different split points across different models (MobileNet-V2 and ResNet50) in terms of communication and computation overhead as a way to minimize the end-to-end inference latency. We propose a Beam Search-based algorithm for split point optimization that minimizes end-to-end latency, and compare it with other methods, including Greedy Search, First-Fit, Random-Fit, and Brute Force. ESP-NOW achieves the best RTT (3.6 s) and serves as the base protocol for the algorithm, which delivers near-optimal latency with processing time of 0.1 s for 5 devices.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. This paper presents the first experimental latency benchmark of TinyML-based split learning on ESP32-S3 boards. It compares four wireless protocols (UDP, TCP, ESP-NOW, BLE) for communication in split inference, analyzes the impact of split points on models including MobileNet-V2 and ResNet50 in terms of communication and computation overhead, and proposes a Beam Search-based algorithm for split-point optimization. The algorithm is compared to Greedy Search, First-Fit, Random-Fit, and Brute Force, with the central claims that ESP-NOW achieves the lowest RTT of 3.6 s and that Beam Search delivers near-optimal end-to-end latency with a processing time of 0.1 s for 5 devices.

Significance. If the reported measurements are reproducible and the experimental conditions are representative, the work supplies useful empirical data on protocol selection and heuristic optimization for latency-critical split learning in constrained IoT hardware. The practical runtime of the Beam Search heuristic (0.1 s) and the identification of ESP-NOW as the lowest-latency protocol among the tested set constitute concrete, deployable insights for TinyML system designers.

major comments (2)
  1. [Experimental Evaluation / Abstract] The abstract and experimental sections report concrete performance numbers (ESP-NOW RTT of 3.6 s, Beam Search processing time of 0.1 s for 5 devices) yet supply no information on the number of measurement repetitions, standard deviation, confidence intervals, or statistical tests used to establish that one protocol or algorithm is superior. Without these details the comparative claims rest on single-point observations whose reliability cannot be assessed.
  2. [Experimental Setup] The latency results are obtained under controlled conditions on ESP32-S3 boards; the manuscript does not describe tests that introduce realistic wireless interference, channel contention, or background traffic. Because the ordering of protocols (ESP-NOW best) and the near-optimality of Beam Search are central to the contribution, the absence of such stress tests leaves the robustness of the reported rankings open to question.
minor comments (1)
  1. [Abstract] The abstract states that split-point choice affects communication and computation overhead but does not quantify the trade-off (e.g., bytes transferred versus FLOPs at each split point) for the two models; a short table or sentence would clarify the motivation for the optimization study.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and insightful comments. We address each of the major comments below and indicate the revisions made to the manuscript.

read point-by-point responses

  1. Referee: [Experimental Evaluation / Abstract] The abstract and experimental sections report concrete performance numbers (ESP-NOW RTT of 3.6 s, Beam Search processing time of 0.1 s for 5 devices) yet supply no information on the number of measurement repetitions, standard deviation, confidence intervals, or statistical tests used to establish that one protocol or algorithm is superior. Without these details the comparative claims rest on single-point observations whose reliability cannot be assessed.

    Authors: We thank the referee for this observation. The manuscript currently omits details regarding the number of repetitions and statistical measures. We will revise the experimental sections to include the number of measurement repetitions performed, standard deviations, confidence intervals, and any statistical tests used to support the comparisons. revision: yes

  2. Referee: [Experimental Setup] The latency results are obtained under controlled conditions on ESP32-S3 boards; the manuscript does not describe tests that introduce realistic wireless interference, channel contention, or background traffic. Because the ordering of protocols (ESP-NOW best) and the near-optimality of Beam Search are central to the contribution, the absence of such stress tests leaves the robustness of the reported rankings open to question.

    Authors: We agree that evaluating the system under realistic interference and contention would provide a more comprehensive assessment of robustness. Our current experiments focus on baseline performance in a controlled setting to clearly compare the protocols and optimization algorithms without confounding factors. We will revise the manuscript to explicitly state the controlled nature of the experiments and add a section discussing potential impacts of interference as a limitation, along with suggestions for future work in more dynamic environments. revision: partial

Circularity Check

0 steps flagged

No significant circularity in experimental benchmarking and algorithm comparison

full rationale

The paper conducts direct latency measurements on ESP32-S3 hardware across protocols (UDP, TCP, ESP-NOW, BLE) and models (MobileNet-V2, ResNet50), then evaluates a proposed Beam Search algorithm against Greedy Search, First-Fit, Random-Fit, and Brute Force baselines. All central claims derive from these empirical timings and comparisons rather than from equations, fitted parameters renamed as predictions, or self-citation chains. No load-bearing step reduces by construction to its own inputs; the work is self-contained experimental analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claims rest on experimental measurements of latency under specific hardware and protocol conditions rather than on mathematical axioms or free parameters fitted to data. No invented entities or ad-hoc assumptions are evident from the abstract.

pith-pipeline@v0.9.0 · 5750 in / 1176 out tokens · 39578 ms · 2026-05-19T03:24:04.700837+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

  1. [1]

    Ai-powered iot: A survey on integrating artificial intelligence with iot for enhanced security, efficiency, and smart appli- cations,

    V . M. U and et. al., “Ai-powered iot: A survey on integrating artificial intelligence with iot for enhanced security, efficiency, and smart appli- cations,” IEEE Access, vol. 13, pp. 50 296–50 339, 2025

    work page 2025

  2. [2]

    Lightweight deep learning for resource-constrained environments: A survey,

    H.-I. Liu, M. Galindo, H. Xie, L.-K. Wong, H.-H. Shuai, Y .-H. Li, and W.-H. Cheng, “Lightweight deep learning for resource-constrained environments: A survey,” ACM Comput. Surv. , vol. 56, no. 10, Jun

    work page

  3. [3]

    Available: https://doi.org/10.1145/3657282

    [Online]. Available: https://doi.org/10.1145/3657282

    work page doi:10.1145/3657282

  4. [4]

    Tinyml applications, research challenges, and future research directions,

    H. Oufettoul, R. Chaibi, and S. Motahhir, “Tinyml applications, research challenges, and future research directions,” in 2024 21st Learning and Technology Conference (L&T), 2024, pp. 86–91

    work page 2024

  5. [5]

    Adaptive split learning over energy-constrained wireless edge networks,

    Z. Li, W. Wu, S. Wu, and W. Wang, “Adaptive split learning over energy-constrained wireless edge networks,” in IEEE INFOCOM 2024- IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2024, pp. 1–6

    work page 2024

  6. [6]

    Split computing and early exiting for deep learning applications: Survey and research challenges,

    Y . Matsubara, M. Levorato, and F. Restuccia, “Split computing and early exiting for deep learning applications: Survey and research challenges,” ACM Computing Surveys , vol. 55, no. 5, pp. 1–30, 2022

    work page 2022

  7. [7]

    A review on the emerging technology of tinyml,

    V . Tsoukas, A. Gkogkidis, E. Boumpa, and A. Kakarountas, “A review on the emerging technology of tinyml,” ACM Comput. Surv. , vol. 56, no. 10, Jun. 2024. [Online]. Available: https://doi.org/10.1145/3661820

    work page doi:10.1145/3661820 2024

  8. [8]

    Performance evaluation of split computing with tinyml on iot devices,

    F. Bove, S. Colli, and L. Bedogni, “Performance evaluation of split computing with tinyml on iot devices,” in 2024 IEEE 21st Consumer Communications & Networking Conference (CCNC) , 2024, pp. 1–6

    work page 2024

  9. [9]

    Latency-aware split learning optimization via genetic algorithms,

    L. H. Trung, T. Y . Nguyen, D. D. Le, T. T. Dang, and T. A. Khoa, “Latency-aware split learning optimization via genetic algorithms,” in Proceedings of the 6th Workshop on Intelligent Cross-Data Analysis and Retrieval, 2025, pp. 14–19

    work page 2025

  10. [10]

    Split learning in 6g edge networks,

    Z. Lin and et. al., “Split learning in 6g edge networks,” IEEE Wireless Communications, vol. 31, no. 4, pp. 170–176, 2024

    work page 2024

  11. [11]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 4510–4520

    work page 2018

  12. [12]

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

    B. Jacob and et. al., “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    work page 2018

  13. [13]

    Improving congestion control of tcp for constrained iot networks,

    C. Lim, “Improving congestion control of tcp for constrained iot networks,” Sensors, vol. 20, no. 17, 2020. [Online]. Available: https://www.mdpi.com/1424-8220/20/17/4774

    work page 2020