pith. sign in

arxiv: 2605.31226 · v1 · pith:RV3YJH4Bnew · submitted 2026-05-29 · 💻 cs.LG · cs.AI

What changes after deployment? A survey on On-device Learning in TinyML

Massimo Pavan , Luca Pezzarossa , Fabrizio Pittorino , Manuel Roveri , Xenofon Fafoutis This is my paper

Pith reviewed 2026-06-28 23:09 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords on-device learningTinyMLdistribution changesurveymicrocontrollermachine learning deploymentadaptationedge devices
0
0 comments X

The pith

Organizing on-device learning papers by post-deployment distribution change regime shows how change type shapes feasible applications, hardware, and solution structures in TinyML.

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

The paper surveys roughly 70 works on on-device learning for microcontroller devices under the single organizing principle of distribution change regimes that appear after deployment. It examines how each regime type affects which applications become addressable on the device, which hardware platforms are chosen, and how the learning methods themselves are built. The survey also documents a consistent mismatch between the benchmarks used in research papers and the conditions encountered in actual deployments. A sympathetic reader would care because static models break when real-world data shifts, and on-device learning is offered as the remedy, yet without mapping change types to solution requirements it is hard to know which remedy fits which situation.

Core claim

The survey claims that distribution change regimes supply the primary lens for characterizing on-device learning solutions: each regime determines the applications that can be handled on-device, the hardware that is employed, and the architectural form the solutions take, while also exposing a persistent gap between the evaluation settings used in methodological work and the conditions of real deployments.

What carries the argument

Distribution change regime, used as the single organizing principle that groups works according to the manner in which data distributions shift after deployment.

If this is right

  • Applications that can run on-device are limited or enabled according to the specific distribution change regime present.
  • Hardware selection for TinyML devices is driven by the requirements of the prevailing change regime.
  • The internal structure of on-device learning algorithms varies systematically with the change regime they target.
  • Evaluation benchmarks in current papers do not match the distribution changes that occur in deployed systems.

Where Pith is reading between the lines

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

  • Designers of future on-device learning systems could first identify the expected change regime and then select matching hardware and algorithm families.
  • Closing the benchmark-to-deployment gap would require new evaluation protocols that inject realistic distribution shifts drawn from target application domains.
  • Hardware platforms might be co-designed with specific change regimes in mind rather than treated as general-purpose accelerators.

Load-bearing premise

That grouping the literature by distribution change regime captures the main structural differences among on-device learning solutions and that the chosen papers represent the field.

What would settle it

A new collection of on-device learning papers whose applications, hardware choices, and solution structures do not align with the distribution change regimes assigned to them would show the organizing principle does not hold.

Figures

Figures reproduced from arXiv: 2605.31226 by Fabrizio Pittorino, Luca Pezzarossa, Manuel Roveri, Massimo Pavan, Xenofon Fafoutis.

Figure 1
Figure 1. Figure 1: Number of papers published over the years in ODL for TinyML. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The general architecture of an ODL solution [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The sequence of configurations of a batch learning solution [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The configuration of an incremental learning solution [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
read the original abstract

Machine learning models on microcontroller-class devices (TinyML) face a fundamental challenge: post-deployment distribution change undermines static models. On-device learning (ODL) addresses this by running the learning process directly on the device. The existing literature has not characterized how distribution change occurs or how different change types require different solutions. Approximately 70 ODL works are surveyed under one principle: the distribution change regime. The survey analyzes how different types of distribution change influence the applications addressable on-device, the hardware employed, and the structure of the solutions. A persistent gap between methodological benchmarks and real-world deployment scenarios is also identified.

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

1 major / 0 minor

Summary. The manuscript is a survey of approximately 70 on-device learning (ODL) works in TinyML, organized under the principle of distribution change regimes. It analyzes how different regimes affect addressable applications, hardware choices, and solution structures, while identifying a persistent gap between typical methodological benchmarks and real-world deployment conditions.

Significance. If the surveyed corpus is representative and the distribution-change axis proves to be the dominant structural lens, the work could usefully organize a fragmented literature and highlight the need for more deployment-oriented evaluation protocols in TinyML.

major comments (1)
  1. [Methodology / Survey Scope] The central claim that the ~70 works are representative and that distribution change regime is the primary organizing principle rests on an unstated or undescribed paper-selection process. No search protocol, databases, keywords, or inclusion/exclusion criteria are supplied, preventing verification that the corpus adequately covers the field or that alternative TinyML constraints (memory footprint, power envelope) are not equally or more load-bearing.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our survey. The major comment identifies a clear gap in methodological transparency, which we will address through revision.

read point-by-point responses

  1. Referee: The central claim that the ~70 works are representative and that distribution change regime is the primary organizing principle rests on an unstated or undescribed paper-selection process. No search protocol, databases, keywords, or inclusion/exclusion criteria are supplied, preventing verification that the corpus adequately covers the field or that alternative TinyML constraints (memory footprint, power envelope) are not equally or more load-bearing.

    Authors: We agree that the absence of an explicit paper-selection protocol is an oversight that undermines verifiability of the corpus and the choice of organizing axis. In the revised manuscript we will insert a new 'Survey Methodology' subsection (placed after the introduction) that specifies: (1) databases queried (Google Scholar, arXiv, IEEE Xplore, ACM DL); (2) search strings combining 'TinyML'/'microcontroller' with 'on-device learning', 'continual learning', 'online learning', 'distribution shift', 'concept drift', and hardware terms; (3) temporal scope (2018–2024); and (4) inclusion criteria (peer-reviewed or preprint works that implement at least one learning step on an MCU-class device with documented memory <1 MB and power constraints) together with exclusion criteria (pure simulation studies, cloud-offloaded training, or works lacking hardware evaluation). We will also add a short justification paragraph explaining why distribution-change regime was elevated as the primary lens: it directly governs which learning primitives are feasible on-device, whereas memory and power act as secondary filters applied uniformly across regimes. These additions will allow readers to assess coverage and the relative importance of the chosen axis without altering the survey's core findings or structure. revision: yes

Circularity Check

0 steps flagged

No circularity: survey organizes external literature without derivations or self-referential reductions

full rationale

This is a survey paper whose core activity is classifying ~70 external works by distribution change regime. No equations, fitted parameters, predictions, or mathematical derivations exist that could reduce to inputs by construction. The organization principle is an explicit classification choice stated in the abstract, not a claim derived from self-citation or ansatz. Selection of works is presented as a survey corpus without any load-bearing uniqueness theorem or self-citation chain. The paper is self-contained against external benchmarks as a literature review and receives the default non-finding for absence of any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a literature survey paper; it introduces no free parameters, mathematical axioms, or invented entities. All claims derive from analysis of cited prior works.

pith-pipeline@v0.9.1-grok · 5640 in / 1065 out tokens · 28771 ms · 2026-06-28T23:09:03.576100+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

111 extracted references · 84 canonical work pages · 9 internal anchors

  1. [1]

    Norhen Abdennadher, Danilo Pau, and Arcangelo Bruna. 2021. Fixed complexity tiny reservoir heterogeneous network for on-line ECG learning of anomalies. In2021 IEEE 10th Global Conference on Consumer Electronics (GCCE). IEEE, Piscataway, NJ, USA, 233–237. doi:10.1109/GCCE53005.2021. 9622022 ISSN: 2378-8143

    work page doi:10.1109/gcce53005.2021 2021

  2. [2]

    Omar Abozaid, Bassant Selim, and Brigitte Jaumard. 2025. Adaptive Embedded Machine Learning on IoT Devices: A Review. In2025 International Telecommunications Conference (ITC-Egypt). IEEE, Piscataway, NJ, USA, 553–560. doi:10.1109/ITC-Egypt66095.2025.11186607

    work page doi:10.1109/itc-egypt66095.2025.11186607 2025

  3. [3]

    Anguita, Alessandro Ghio, L

    D. Anguita, Alessandro Ghio, L. Oneto, Xavier Parra, and Jorge Luis Reyes-Ortiz. 2013. A Public Domain Dataset for Human Activity Recognition using Smartphones. InThe European Symposium on Artificial Neural Networks. springer, Bruges (Belgium). https://api.semanticscholar.org/CorpusID: 6975432

    2013

  4. [4]

    2009.Cortex-M0 Technical Reference Manual

    Arm Limited. 2009.Cortex-M0 Technical Reference Manual. Technical Report ARM DDI 0432C. Arm Limited. https://developer.arm.com/ documentation/ddi0432/c/ Accessed: May 2026

    2009

  5. [5]

    2010.Cortex-M4 Technical Reference Manual

    Arm Limited. 2010.Cortex-M4 Technical Reference Manual. Technical Report ARM DDI 0439B. Arm Limited. https://developer.arm.com/ documentation/ddi0439/be/ Accessed: May 2026

    2010

  6. [6]

    2015.Arm Cortex-M7 Processor Technical Reference Manual

    Arm Limited. 2015.Arm Cortex-M7 Processor Technical Reference Manual. Technical Report ARM DDI 0489F. Arm Limited. https://developer.arm. com/documentation/ddi0489/f/ Accessed: May 2026

    2015

  7. [7]

    2020.Arm Ethos-U55: Product Brief

    Arm Limited. 2020.Arm Ethos-U55: Product Brief. Technical Report. Arm Limited. https://armkeil.blob.core.windows.net/developer/Files/pdf/ product-brief/arm-ethos-u55-product-brief.pdf Accessed: May 2026

    2020

  8. [8]

    2020.Arm Ethos-U65 microNPU: Product Brief

    Arm Limited. 2020.Arm Ethos-U65 microNPU: Product Brief. Technical Report. Arm Limited. https://developer.arm.com/-/media/Arm%20Developer% 20Community/PDF/Machine%20Learning/Arm_Ethos_U65_Product_Brief.pdf Accessed: May 2026

    2020

  9. [9]

    Colby Banbury, Vijay Janapa Reddi, Peter Torelli, Jeremy Holleman, Nat Jeffries, Csaba Kiraly, Pietro Montino, David Kanter, Sebastian Ahmed, Danilo Pau, et al. 2021. MLPerf Tiny Benchmark.Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks(2021)

    2021

  10. [10]

    Simone Benatti, Fabio Montagna, Victor Kartsch, Abbas Rahimi, Davide Rossi, and Luca Benini. 2019. Online Learning and Classification of EMG-Based Gestures on a Parallel Ultra-Low Power Platform Using Hyperdimensional Computing.IEEE Transactions on Biomedical Circuits and Systems13, 3 (June 2019), 516–528. doi:10.1109/TBCAS.2019.2914476

    work page doi:10.1109/tbcas.2019.2914476 2019

  11. [11]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101 – Mining Discriminative Components with Random Forests. InComputer Vision – ECCV 2014, David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars (Eds.). Springer International Publishing, Cham, 446–461

    2014

  12. [12]

    Burrello, A

    A. Burrello, A. Garofalo, N. Bruschi, G. Tagliavini, D. Rossi, and F. Conti. 2021. DORY: Automatic End-to-End Deployment of Real-World DNNs on Low-Cost IoT MCUs.IEEE Trans. Comput.(2021), 1–1. doi:10.1109/TC.2021.3066883

    work page doi:10.1109/tc.2021.3066883 2021

  13. [13]

    Alessio Burrello, Marcello Zanghieri, Cristian Sarti, Leonardo Ravaglia, Simone Benatti, and Luca Benini. 2021. Tackling Time-Variability in sEMG-based Gesture Recognition with On-Device Incremental Learning and Temporal Convolutional Networks. In2021 IEEE Sensors Applications Symposium (SAS). IEEE, Piscataway, NJ, USA, 1–6. doi:10.1109/SAS51076.2021.9530007

    work page doi:10.1109/sas51076.2021.9530007 2021

  14. [14]

    Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. 2021. TinyTL: Reduce Activations, Not Trainable Parameters for Efficient On-Device Learning. doi:10.48550/arXiv.2007.11622 arXiv:2007.11622 [cs]

    work page doi:10.48550/arxiv.2007.11622 2021

  15. [15]

    Elia Cereda, Alessandro Giusti, and Daniele Palossi. 2024. Training on the Fly: On-device Self-supervised Learning aboard Nano-drones within 20 mW. doi:10.48550/arXiv.2408.03168 arXiv:2408.03168 [cs]

    work page doi:10.48550/arxiv.2408.03168 2024

  16. [16]

    Elia Cereda, Manuele Rusci, Alessandro Giusti, and Daniele Palossi. 2024. On-device Self-supervised Learning of Visual Perception Tasks aboard Hardware-limited Nano-quadrotors. doi:10.48550/arXiv.2403.04071 arXiv:2403.04071 [cs]

    work page doi:10.48550/arxiv.2403.04071 2024

  17. [17]

    Kwon, and Cecilia Mascolo

    Jagmohan Chauhan, Young D. Kwon, and Cecilia Mascolo. 2022. Exploring On-Device Learning Using Few Shots for Audio Classification. In 2022 30th European Signal Processing Conference (EUSIPCO). IEEE, Piscataway, NJ, USA, 424–428. doi:10.23919/EUSIPCO55093.2022.9909551 ISSN: 2076-1465

    work page doi:10.23919/eusipco55093.2022.9909551 2022

  18. [18]

    Mahesh Chowdhary and Swapnil Sayan Saha. 2023. On-Sensor Online Learning and Classification Under 8 KB Memory. In2023 26th International Conference on Information Fusion (FUSION). IEEE, Piscataway, NJ, USA, 1–8. doi:10.23919/FUSION52260.2023.10224228

    work page doi:10.23919/fusion52260.2023.10224228 2023

  19. [19]

    Aakanksha Chowdhery, Pete Warden, Jonathon Shlens, Andrew Howard, and Rocky Rhodes. 2019. Visual Wake Words Dataset. arXiv:1906.05721 [cs.CV] https://arxiv.org/abs/1906.05721 Manuscript submitted to ACM What changes after deployment? A survey on On-device Learning in TinyML 29

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  20. [20]

    Cristian Cioflan, Lukas Cavigelli, and Luca Benini. 2024. Boosting keyword spotting through on-device learnable user speech characteristics. doi:10.48550/arXiv.2403.07802 arXiv:2403.07802 [cs]

    work page doi:10.48550/arxiv.2403.07802 2024

  21. [21]

    Cristian Cioflan, Lukas Cavigelli, Manuele Rusci, Miguel De Prado, and Luca Benini. 2022. Towards On-device Domain Adaptation for Noise-Robust Keyword Spotting. In2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, Piscataway, NJ, USA, 82–85. doi:10.1109/AICAS54282.2022.9869990

    work page doi:10.1109/aicas54282.2022.9869990 2022

  22. [22]

    Cristian Cioflan, Lukas Cavigelli, Manuele Rusci, Miguel de Prado, and Luca Benini. 2024. On-Device Domain Learning for Keyword Spotting on Low-Power Extreme Edge Embedded Systems. doi:10.48550/arXiv.2403.10549 arXiv:2403.10549 [cs]

    work page doi:10.48550/arxiv.2403.10549 2024

  23. [23]

    Michele Craighero, Davide Quarantiello, Beatrice Rossi, Diego Carrera, Pasqualina Fragneto, and Giacomo Boracchi. 2024. On-Device Personalization for Human Activity Recognition on STM32.IEEE Embedded Systems Letters16, 2 (June 2024), 106–109. doi:10.1109/LES.2023.3293458

    work page doi:10.1109/les.2023.3293458 2024

  24. [24]

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. 2021. A continual learning survey: Defying forgetting in classification tasks.IEEE transactions on pattern analysis and machine intelligence44, 7 (2021), 3366–3385

    2021

  25. [25]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. ImageNet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 248–255. doi:10.1109/CVPR.2009.5206848

    work page doi:10.1109/cvpr.2009.5206848 2009

  26. [26]

    Mark Deutel, Frank Hannig, Christopher Mutschler, and Jürgen Teich. 2025. On-Device Training of Fully Quantized Deep Neural Networks on Cortex-M Microcontrollers.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems44, 4 (April 2025), 1250–1261. doi:10.1109/TCAD.2024.3484354

    work page doi:10.1109/tcad.2024.3484354 2025

  27. [27]

    Simone Disabato and Manuel Roveri. 2020. Incremental On-Device Tiny Machine Learning. InProceedings of the 2nd International Workshop on Challenges in Artificial Intelligence and Machine Learning for Internet of Things (AIChallengeIoT ’20). Association for Computing Machinery, New York, NY, USA, 7–13. doi:10.1145/3417313.3429378

    work page doi:10.1145/3417313.3429378 2020

  28. [28]

    Simone Disabato and Manuel Roveri. 2024. Tiny Machine Learning for Concept Drift.IEEE Transactions on Neural Networks and Learning Systems 35, 6 (June 2024), 8470–8481. doi:10.1109/TNNLS.2022.3229897

    work page doi:10.1109/tnnls.2022.3229897 2024

  29. [29]

    Gregory Ditzler, Manuel Roveri, Cesare Alippi, and Robi Polikar. 2015. Learning in nonstationary environments: A survey.IEEE Computational Intelligence Magazine10, 4 (2015), 12–25

    2015

  30. [30]

    R. A. Fisher. 1936. Iris. UCI Machine Learning Repository. doi:10.24432/C56C76

    work page doi:10.24432/c56c76 1936

  31. [31]

    Joao Gama, Raquel Sebastiao, and Pedro Pereira Rodrigues. 2013. On evaluating stream learning algorithms.Machine learning90 (2013), 317–346

    2013

  32. [32]

    Toni Heittola, Annamaria Mesaros, and Tuomas Virtanen. 2020. Acoustic scene classification in DCASE 2020 Challenge: generalization across devices and low complexity solutions. arXiv:2005.14623 [eess.AS] https://arxiv.org/abs/2005.14623

    work page arXiv 2020

  33. [33]

    Benchmarking Neural Network Robustness to Common Corruptions and Perturbations

    Dan Hendrycks and Thomas G. Dietterich. 2019. Benchmarking Neural Network Robustness to Common Corruptions and Perturbations.CoRR abs/1903.12261 (2019). arXiv:1903.12261 http://arxiv.org/abs/1903.12261

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  34. [34]

    Maxwell L Hutchinson, Erin Antono, Brenna M Gibbons, Sean Paradiso, Julia Ling, and Bryce Meredig. 2017. Overcoming data scarcity with transfer learning.arXiv preprint arXiv:1711.05099(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  35. [35]

    Andras Janosi. 1989. Heart Disease. UCI Machine Learning Repository. doi:10.24432/C52P4X

    work page doi:10.24432/c52p4x 1989

  36. [36]

    Clancy, Metin Akay, and Wei Chen

    Xinyu Jiang, Xiangyu Liu, Jiahao Fan, Xinming Ye, Chenyun Dai, Edward A. Clancy, Metin Akay, and Wei Chen. 2021. Open Access Dataset, Toolbox and Benchmark Processing Results of High-Density Surface Electromyogram Recordings.IEEE Transactions on Neural Systems and Rehabilitation Engineering29 (2021), 1035–1046. doi:10.1109/TNSRE.2021.3082551

    work page doi:10.1109/tnsre.2021.3082551 2021

  37. [37]

    Osama Khan, Gwanjong Park, and Euiseong Seo. 2023. DaCapo: An On-Device Learning Scheme for Memory-Constrained Embedded Systems. ACM Trans. Embed. Comput. Syst.22, 5s (2023), 142:1–142:23. doi:10.1145/3609121

    work page doi:10.1145/3609121 2023

  38. [38]

    Byeonggeun Kim, Mingu Lee, Jinkyu Lee, Yeonseok Kim, and Kyuwoong Hwang. 2019. Query-by-example on-device keyword spotting. arXiv:arXiv:1910.05171

    work page arXiv 2019

  39. [39]

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3D Object Representations for Fine-Grained Categorization. In2013 IEEE International Conference on Computer Vision Workshops. IEEE, Piscataway, NJ, USA, 554–561. doi:10.1109/ICCVW.2013.77

    work page doi:10.1109/iccvw.2013.77 2013

  40. [40]

    2009.Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. 2009.Learning multiple layers of features from tiny images. Technical Report 0. University of Toronto, Toronto, Ontario. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf

    2009

  41. [41]

    Kwapisz, Gary M

    Jennifer R. Kwapisz, Gary M. Weiss, and Samuel A. Moore. 2011. Activity recognition using cell phone accelerometers.SIGKDD Explor. Newsl.12, 2 (March 2011), 74–82. doi:10.1145/1964897.1964918

    work page doi:10.1145/1964897.1964918 2011

  42. [42]

    Kwon, Jagmohan Chauhan, Hong Jia, Stylianos I

    Young D. Kwon, Jagmohan Chauhan, Hong Jia, Stylianos I. Venieris, and Cecilia Mascolo. 2023. LifeLearner: Hardware-Aware Meta Continual Learning System for Embedded Computing Platforms. doi:10.48550/arXiv.2311.11420 arXiv:2311.11420 [cs]

    work page doi:10.48550/arxiv.2311.11420 2023

  43. [43]

    Kwon, Jagmohan Chauhan, Abhishek Kumar, Pan Hui HKUST, and Cecilia Mascolo

    Young D. Kwon, Jagmohan Chauhan, Abhishek Kumar, Pan Hui HKUST, and Cecilia Mascolo. 2021. Exploring System Performance of Continual Learning for Mobile and Embedded Sensing Applications. In2021 IEEE/ACM Symposium on Edge Computing (SEC). IEEE, Piscataway, NJ, USA, 319–332. doi:10.1145/3453142.3491285

    work page doi:10.1145/3453142.3491285 2021

  44. [44]

    Kwon, Rui Li, Stylianos I

    Young D. Kwon, Rui Li, Stylianos I. Venieris, Jagmohan Chauhan, Nicholas D. Lane, and Cecilia Mascolo. 2024. TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge. doi:10.48550/arXiv.2307.09988 arXiv:2307.09988 [cs]

    work page doi:10.48550/arxiv.2307.09988 2024

  45. [45]

    Lake, Ruslan Salakhutdinov, and Joshua B

    Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. 2015. Human-level concept learning through probabilistic program induction. Science350, 6266 (2015), 1332–1338. arXiv:https://www.science.org/doi/pdf/10.1126/science.aab3050 doi:10.1126/science.aab3050

    work page doi:10.1126/science.aab3050 2015

  46. [46]

    Laine, S

    Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. 1998. Gradient-based learning applied to document recognition.Proc. IEEE86, 11 (1998), 2278–2324. doi:10.1109/5.726791 Manuscript submitted to ACM 30 Pavan et al

    work page doi:10.1109/5.726791 1998

  47. [47]

    Clayton Frederick Souza Leite and Yu Xiao. 2022. Resource-Efficient Continual Learning for Sensor-Based Human Activity Recognition.ACM Trans. Embed. Comput. Syst.21, 6 (2022), 85:1–85:25. doi:10.1145/3530910

    work page doi:10.1145/3530910 2022

  48. [48]

    David Leroy, Alice Coucke, Thibaut Lavril, Thibault Gisselbrecht, and Joseph Dureau. 2019. Federated Learning for Keyword Spotting. arXiv:1810.05512 [eess.AS] https://arxiv.org/abs/1810.05512

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  49. [49]

    Ji Lin, Ligeng Zhu, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, and Song Han. 2024. On-Device Training Under 256KB Memory. doi:10.48550/ arXiv.2206.15472 arXiv:2206.15472 [cs]

    work page arXiv 2024

  50. [50]

    Vincenzo Lomonaco and Davide Maltoni. 2017. CORe50: a New Dataset and Benchmark for Continuous Object Recognition. InProceedings of the 1st Annual Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 78), Sergey Levine, Vincent Vanhoucke, and Ken Goldberg (Eds.). PMLR, 17–26. https://proceedings.mlr.press/v78/lomonaco17a.html

    2017

  51. [51]

    Afonso Lourenço, João Rodrigo, João Gama, and Goreti Marreiros. 2025. DFDT: Dynamic Fast Decision Tree for IoT Data Stream Mining on Edge Devices. doi:10.48550/arXiv.2502.14011 arXiv:2502.14011 [cs]

    work page doi:10.48550/arxiv.2502.14011 2025

  52. [52]

    Afonso Lourenço, João Rodrigo, João Gama, and Goreti Marreiros. 2025. On-device edge learning for IoT data streams: a survey. doi:10.48550/arXiv. 2502.17788 arXiv:2502.17788 [cs]

    work page internal anchor Pith review doi:10.48550/arxiv 2025

  53. [53]

    Jie Lu, Anjin Liu, Fan Dong, Feng Gu, Joao Gama, and Guangquan Zhang. 2018. Learning under concept drift: A review.IEEE transactions on knowledge and data engineering31, 12 (2018), 2346–2363

    2018

  54. [54]

    EmbodiTTA: Resource-Efficient Test-Time Adaptation for Embodied Visual Systems

    Xiao Ma, Young D. Kwon, and Dong Ma. 2025. On-demand Test-time Adaptation for Edge Devices. doi:10.48550/arXiv.2505.00986 arXiv:2505.00986 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.00986 2025

  55. [55]

    S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. 2013.Fine-Grained Visual Classification of Aircraft. Technical Report. arXiv:1306.5151 [cs- cv]

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  56. [56]

    Mark Mazumder, Sharad Chitlangia, Colby Banbury, Yiping Kang, Juan Manuel Ciro, Keith Achorn, Daniel Galvez, Mark Sabini, Peter Mattson, David Kanter, Greg Diamos, Pete Warden, Josh Meyer, and Vijay Janapa Reddi. 2021. Multilingual Spoken Words Corpus. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round ...

    2021

  57. [57]

    2022.Personalized Machine Learning

    Julian McAuley. 2022.Personalized Machine Learning. Cambridge University Press

    2022

  58. [58]

    Lan Mei, Cristian Cioflan, Thorir Mar Ingolfsson, Victor Kartsch, Andrea Cossettini, Xiaying Wang, and Luca Benini. 2024. Train-On-Request: An On-Device Continual Learning Workflow for Adaptive Real-World Brain Machine Interfaces. In2024 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, Piscataway, NJ, USA, 1–5. doi:10.1109/BioCAS61083.2024....

    work page doi:10.1109/biocas61083.2024.10798357 2024

  59. [59]

    Lan Mei, Thorir Mar Ingolfsson, Cristian Cioflan, Victor Kartsch, Andrea Cossettini, Xiaying Wang, and Luca Benini. 2025. An Ultra-Low Power Wearable BMI System With Continual Learning Capabilities.IEEE Transactions on Biomedical Circuits and Systems19, 3 (June 2025), 511–522. doi:10.1109/TBCAS.2024.3457522

    work page doi:10.1109/tbcas.2024.3457522 2025

  60. [60]

    Moody and R.G

    G.B. Moody and R.G. Mark. 2001. The impact of the MIT-BIH arrhythmia database.IEEE engineering in medicine and biology magazine : the quarterly magazine of the Engineering in Medicine & Biology Society20 (06 2001), 45–50. doi:10.1109/51.932724

    work page doi:10.1109/51.932724 2001

  61. [61]

    Davide Nadalini, Manuele Rusci, Giuseppe Tagliavini, Leonardo Ravaglia, Luca Benini, and Francesco Conti. 2022. PULP-TrainLib: Enabling On-Device Training for RISC-V Multi-core MCUs Through Performance-Driven Autotuning. InEmbedded Computer Systems: Architectures, Modeling, and Simulation, Alex Orailoglu, Marc Reichenbach, and Matthias Jung (Eds.). Spring...

    2022

  62. [62]

    Maria-Elena Nilsback and Andrew Zisserman. 2008. Automated Flower Classification over a Large Number of Classes. In2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. 722–729. doi:10.1109/ICVGIP.2008.47

    work page doi:10.1109/icvgip.2008.47 2008

  63. [63]

    German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. 2019. Continual lifelong learning with neural networks: A review.Neural networks113 (2019), 54–71

    2019

  64. [64]

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. 2012. Cats and dogs. In2012 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 3498–3505. doi:10.1109/CVPR.2012.6248092

    work page doi:10.1109/cvpr.2012.6248092 2012

  65. [65]

    Patil, Paras Jain, Prabal Dutta, Ion Stoica, and Joseph E

    Shishir G. Patil, Paras Jain, Prabal Dutta, Ion Stoica, and Joseph E. Gonzalez. 2022. POET: Training Neural Networks on Tiny Devices with Integrated Rematerialization and Paging. doi:10.48550/arXiv.2207.07697 arXiv:2207.07697 [cs]

    work page doi:10.48550/arxiv.2207.07697 2022

  66. [66]

    Danilo Pau, Prem Kumar Ambrose, Andrea Pisani, and Fabrizio M. Aymone. 2023. TinyRCE: Forward Learning Under Tiny Constraints. In2023 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE). IEEE, Piscataway, NJ, USA, 295–300. doi:10.1109/MetroXRAINE58569.2023.10405784

    work page doi:10.1109/metroxraine58569.2023.10405784 2023

  67. [67]

    Danilo Pau, Abderrahim Khiari, and Davide Denaro. 2021. Online learning on tiny micro-controllers for anomaly detection in water distribution systems. In2021 IEEE 11th International Conference on Consumer Electronics (ICCE-Berlin). IEEE, Piscataway, NJ, USA, 1–6. doi:10.1109/ICCE- Berlin53567.2021.9720009 ISSN: 2166-6822

    work page doi:10.1109/icce- 2021

  68. [68]

    Danilo Pietro Pau, Andrea Pisani, Fabrizio Maria Aymone, and Gianluigi Ferrari. 2023. TinyRCE: Multipurpose Forward Learning for Resource Restricted Devices.IEEE Sensors Letters7, 10 (Oct. 2023), 1–4. doi:10.1109/LSENS.2023.3307119

    work page doi:10.1109/lsens.2023.3307119 2023

  69. [69]

    Danilo Pietro Pau, Simone Tognocchi, and Marco Marcon. 2025. Learning Online MEMS Calibration with Time-Varying and Memory-Efficient Gaussian Neural Topologies.Sensors25, 12 (Jan. 2025), 3679. doi:10.3390/s25123679 Publisher: Multidisciplinary Digital Publishing Institute

    work page doi:10.3390/s25123679 2025

  70. [70]

    Massimo Pavan, Claudio Galimberti, and Manuel Roveri. 2025. TActiLE: Tiny Active LEarning for wearable devices. In2025 International Joint Conference on Neural Networks (IJCNN). IEEE, IEEE, Piscataway, NJ, USA, 1–8

    2025

  71. [71]

    Massimo Pavan, Gioele Mombelli, Francesco Sinacori, and Manuel Roveri. 2025. TinySV: Speaker Verification in TinyML with On-device Learning. InProceedings of the 4th International Conference on AI-ML Systems (AIMLSystems ’24). Association for Computing Machinery, New York, NY, USA, Manuscript submitted to ACM What changes after deployment? A survey on On-...

    work page doi:10.1145/3703412.3703415 2025

  72. [72]

    Massimo Pavan, Eugeniu Ostrovan, Armando Caltabiano, and Manuel Roveri. 2024. TyBox: An Automatic Design and Code Generation Toolbox for TinyML Incremental On-Device Learning.ACM Trans. Embed. Comput. Syst.23, 3 (2024), 42:1–42:27. doi:10.1145/3604566

    work page doi:10.1145/3604566 2024

  73. [73]

    Pereira, Leonardo S

    Eduardo S. Pereira, Leonardo S. Marcondes, and Josemar M. Silva. 2023. On-Device Tiny Machine Learning for Anomaly Detection Based on the Extreme Values Theory.IEEE Micro43, 6 (Nov. 2023), 58–65. doi:10.1109/MM.2023.3316918

    work page doi:10.1109/mm.2023.3316918 2023

  74. [74]

    Visal Rajapakse, Ishan Karunanayake, and Nadeem Ahmed. 2023. Intelligence at the Extreme Edge: A Survey on Reformable TinyML.ACM Comput. Surv.55, 13s (2023), 282:1–282:30. doi:10.1145/3583683

    work page doi:10.1145/3583683 2023

  75. [75]

    Leonardo Ravaglia, Manuele Rusci, Davide Nadalini, Alessandro Capotondi, Francesco Conti, and Luca Benini. 2021. A TinyML Platform for On-Device Continual Learning with Quantized Latent Replays.IEEE Journal on Emerging and Selected Topics in Circuits and Systems11, 4 (Dec. 2021), 789–802. doi:10.1109/JETCAS.2021.3121554 arXiv:2110.10486 [cs]

    work page doi:10.1109/jetcas.2021.3121554 2021

  76. [76]

    Attila Reiss and Didier Stricker. 2012. Introducing a New Benchmarked Dataset for Activity Monitoring.2012 16th International Symposium on Wearable Computers(2012), 108–109. https://api.semanticscholar.org/CorpusID:10337279

    2012

  77. [77]

    Haoyu Ren, Darko Anicic, Xue Li, and Thomas Runkler. 2024. On-device Online Learning and Semantic Management of TinyML Systems.ACM Trans. Embed. Comput. Syst.23, 4 (2024), 55:1–55:32. doi:10.1145/3665278

    work page doi:10.1145/3665278 2024

  78. [78]

    Haoyu Ren, Darko Anicic, and Thomas Runkler. 2021. TinyOL: TinyML with Online-Learning on Microcontrollers. doi:10.48550/arXiv.2103.08295 arXiv:2103.08295 [cs]

    work page doi:10.48550/arxiv.2103.08295 2021

  79. [79]

    Flavio Renzi, Haoyu Ren, Alessio Bernardo, Giacomo Ziffer, Darko Anicic, and Emanuele Della Valle. 2025. Online Learning Techniques for Occupancy Detection on Resource Constrained Devices. In2025 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW). IEEE, Piscataway, NJ, USA, 1019–1026. doi:10.1109/IPDPSW66978.2025.00159 ISS...

    work page doi:10.1109/ipdpsw66978.2025.00159 2025

  80. [80]

    Manuele Rusci, Francesco Paci, Marco Fariselli, Eric Flamand, and Tinne Tuytelaars. 2025. Self-Learning for Personalized Keyword Spotting on Ultra-Low-Power Audio Sensors.IEEE Internet of Things Journal12, 8 (April 2025), 10210–10221. doi:10.1109/JIOT.2024.3515143 arXiv:2408.12481 [cs]

    work page doi:10.1109/jiot.2024.3515143 2025

Showing first 80 references.