The DOI matches its Crossref record; 40 of the 40 references it cites were found in a registry or an index; 28 of the 56 sources its claims cite were checked for support.
TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications
Open accessAbstract
Recently, the Internet of Things (IoT) has gained a lot of attention, since IoT devices are placed in various fields. Many of these devices are based on machine learning (ML) models, which render them intelligent and able to make decisions. IoT devices typically have limited resources, which restricts the execution of complex ML models such as deep learning (DL) on them. In addition, connecting IoT devices to the cloud to transfer raw data and perform processing causes delayed system responses, exposes private data and increases communication costs. Therefore, to tackle these issues, there is a new technology called Tiny Machine Learning (TinyML), that has paved the way to meet the challenges of IoT devices. This technology allows processing of the data locally on the device without the need to send it to the cloud. In addition, TinyML permits the inference of ML models, concerning DL models on the device as a Microcontroller that has limited resources. The aim of this paper is to provide an overview of the revolution of TinyML and a review of tinyML studies, wherein the main contribution is to provide an analysis of the type of ML models used in tinyML studies; it also presents the details of datasets and the types and characteristics of the devices with an aim to clarify the state of the art and envision development requirements.
Overview
CoChat Summary
TL;DR
This review surveys TinyML, the deployment of compressed deep-learning inference on ultra-low-power IoT edge devices such as microcontrollers, and analyzes models, datasets, devices, frameworks, and limitations across 38 selected papers. It finds that TinyML offers energy, cost, latency, and privacy benefits, but remains constrained by device heterogeneity, limited processing power and memory, and a shortage of suitable DL models.
Key findings
- Across the reviewed TinyML studies, image datasets were the most common data type at 46%, followed by physiological/behavioral metrics at 29% and audio at 11%; the most frequently used image datasets were ImageNET and Visual Wake Words, while Google Speech Commands was the most used audio dataset.
- The most common models achieving best results were ANN at 21%, CNN at 12%, and MobileNet-V1 at 6%; non-neural ML models included Decision Tree at 9%, SVM at 6%, and KNN at 6%.
- Device usage was dominated by STM32 microcontrollers: STM32-F746ZG appeared in 22% of studies and STM32-H743ZI2 in 20%; reviewed device costs ranged approximately from USD 16.00 to USD 28.00, and MCUs were described as typically small (~1 cm³), low power (~1 mW), and cheap (~$1).
- Reported use-case results included ASL detection at 98.84% accuracy and 20 FPS after deployment, handwriting recognition at 100% prediction accuracy with 135.68 KB RAM, 668.97 KB flash, and 330 ms computation, medical face mask detection at 99.83% accuracy and 30 FPS with a 138 KB model, and MCUNet reaching 70.0% accuracy on STM32H743 with 0.49 MB SRAM and 0.9 MB flash.
Methods
- This is a systematic literature review following PRISMA guidelines, organized around research questions on datasets, models, devices, application domains, frameworks/libraries, and limitations in TinyML.
- Searches used Google Scholar, Clarivate’s Web of Science, Science Direct, IEEE, MDPI, and Springer; the first TinyML publication was dated 13 December 2019, and a total of 38 relevant published papers/articles were retrieved and screened using inclusion/exclusion criteria.
- Data extraction covered each study’s main contribution, implementation details including datasets, models, devices, and evaluation criteria; findings were synthesized descriptively into dataset, model, and device categories.
Limitations
- The paper states that TinyML faces device heterogeneity, with no standardized framework across manufacturers; lower processing power and speed than cloud-based systems; limited memory of only a few KB; and a lack of DL models that achieve high accuracy with small model size and low latency.
- Evident from the text: as a review, the findings are descriptive counts and summaries from 38 selected studies rather than new experimental validation, and the exclusion criteria omitted studies on compression techniques such as quantization and pruning, memory reduction, TinyML services, technical reports, and website-published experiments.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures






Verification
Claims the paper makes
23 Valid · 9 Not valid · 19 Not checked · 35 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, edge devices have not yet become cost-efficient enough to take training DL models due to limited resources [7,8,9].”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 1. Introduction
What the sources say
- Not checked — no full text of the sourcecites [7] Atitallah, 2020, Ben Leveraging Deep Learning and IoT big data analytics to support the smart cities development: Review and future directions, Comput. Sci. Rev. source
- Not checked — no full text of the sourcecites [8] Bhattacharya, 2020, A review on deep learning for future smart cities, Internet Technol. Lett. source
- Partly supportedcites [9] Wang, 2020, Deep learning for edge computing applications: A State-of-the-Art survey, IEEE Access sourceExperimental — accuracy not yet measured
The Model Training section says training deep models consumes massive CPU and GPU resources and that edge servers are usually challenged, or not cost-efficient, when undertaking training alone. The claim broadens this from edge servers to edge devices generally and omits the qualification that the servers undertake training solely; the paper distinguishes devices from edge servers in Model Inference. Page numbers are unavailable in our copy. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Authors in [28] proposed a model to detect the American Sign language (ASL) Alphabet and transcribed to text and speech in real time on tiny wearable IoT devices.”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 4. Related Work to TinyML > 4.1. TinyML Use Cases Studies
What the sources say
- Partly supportedcites [28] Paul, A.J., Mohan, P., and Sehgal, S. (2020, January 3–5). Rethinking Generalization in American Sign Language Prediction for Edge Devices with Extremely Low Memory Footprint. Proceedings of the 2020 IEEE Recent Advances in Intelligent Computational Systems, RAICS, Thiruvananthapuram, India. sourceExperimental — accuracy not yet measured
The Introduction (p. 1) describes a proposed ASL fingerspelling model small enough for wearable-IoT microcontrollers and envisions real-time transcription to text or speech. The Conclusion (p. 6) confirms real-time recognition on an OpenMV H7 board, but says a watch-like wearable device remains future work; our copy does not establish that text-and-speech transcription was implemented. (LLM, conf=0.88)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“It is computationally expensive, consuming massive CPU and GPU resources, power, memory, and time.”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 1. Introduction
What the sources say
- Not checked — no full text of the sourcecites [7] Atitallah, 2020, Ben Leveraging Deep Learning and IoT big data analytics to support the smart cities development: Review and future directions, Comput. Sci. Rev. source
- Not checked — no full text of the sourcecites [8] Bhattacharya, 2020, A review on deep learning for future smart cities, Internet Technol. Lett. source
- Partly supportedcites [9] Wang, 2020, Deep learning for edge computing applications: A State-of-the-Art survey, IEEE Access sourceExperimental — accuracy not yet measured
The Model Training section says training is often computation-intensive and consumes massive CPU and GPU resources. It does not specifically attribute substantial power, memory, or time consumption to training; the passage about time and energy concerns model inference instead. Page numbers are unavailable in our copy. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Furthermore, deployment of pre-trained DL models into tiny edge devices is now possible, after performing some techniques to compress DL models and optimize the inference.”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 1. Introduction
What the sources say
- Partly supportedcites [15] Wang, 2020, Pruning from Scratch, Proceedings of the AAAI Conference on Artificial Intelligence sourceExperimental — accuracy not yet measured
The Introduction (p. 1) says deep neural networks are deployed on mobile devices and that pruning compresses models and accelerates inference. It does not establish that these techniques make deployment of pre-trained models on tiny edge devices possible; the paper instead emphasizes pruning without pre-training. (LLM, conf=0.87)
- Supportedcites [16] Wardana, I.N.K., Gardner, J.W., and Fahmy, S.A. (2021). Optimising Deep Learning at the Edge for Accurate Hourly Air Quality Prediction. Sensors, 21. source
The Introduction describes compressing trained models for edge deployment and identifies post-training quantisation as a way to reduce model size and potentially improve latency. Section 3.5 reports that the authors optimised a trained model and deployed it to Raspberry Pi edge devices; Section 5.3 reports reductions in model size. (LLM, conf=0.90)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, inference of DL in edge devices is not yet prepared enough to be fully executed [3,4,6].”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 1. Introduction
What the sources say
- Partly supportedcites [3] Signoretti, G., Silva, M., Andrade, P., Silva, I., Sisinni, E., and Ferrari, P. (2021). An Evolving Tinyml Compression Algorithm for IoT Environments Based on Data Eccentricity. Sensors, 21. sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) says that computational restrictions lead TinyML researchers to reduce and optimize neural-network models for tiny devices. That supports a limitation on edge-device inference, but not the broader assertion that deep-learning inference is not yet ready to be fully executed on edge devices; the same section describes ML models being deployed there. (LLM, conf=0.84)
- Not checked — no full text of the sourcecites [4] Chen, 2020, Deep Learning on Mobile and Embedded Devices: State-of-the-Art, Challenges, and Future Directions, ACM Comput. Surv. source
- Supportedcites [6] Gorospe, J., Mulero, R., Arbelaitz, O., Muguerza, J., and Antón, M.Á. (2021). A Generalization Performance Study Using Deep Learning Networks in Embedded Systems. Sensors, 21. source
The Introduction states that DNNs are not prepared to be fully executed in edge-computing devices because their computing and memory demands exceed those devices’ resources. This directly supports the cited claim about on-device deep-learning inference. (LLM, conf=0.93)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
18 more claims · 1 with their cited sources checked
Create a free account to see all 86 claims and to verify the 8 references still open.
- Probably not a substantial claim. 45
- It may state two separate claims as one. 39
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 3. Research Methodology > 3.2. Search Strategy
“The first publication for TinyML was on 13 December 2019.”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 3. Research Methodology > 3.6. Data Selection
“The first published paper on the emerging topic was on 19 December 2019”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 4. Related Work to TinyML > 4.1. TinyML Use Cases Studies
“the best performance was that of ANN in all datasets except Sonar dataset.”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 4. Related Work to TinyML > 4.1. TinyML Use Cases Studies
“In contrast the SVM achieved the best results on the Heart and EnviroCar dataset.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 3. Research Methodology > 3.5. Data Extraction and Synthesis
“The information extracted from studies is described in Section 6.”
§ TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications > 5. Discussion and Findings
“## 5. Discussion and Findings Based on the related works of tinyML, we have derived the following findings concerning models, datasets, and devices.”
Citations
Cites
ImageNet: A large-scale hierarchical image database
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333
The PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews
Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., Shamseer, L., Tetzlaff, J. M., Akl, E. A., Brennan, S. E., et al. (2021) · International Journal of Surgery · doi:10.1016/j.ijsu.2021.105906 · Cited by 11,748
Automated Flower Classification over a Large Number of Classes
Nilsback, M.-E., & Zisserman, A. (2008) · 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing · doi:10.1109/icvgip.2008.47 · Cited by 3,442
ImageNet: A large-scale hierarchical image database
Deng, J. H., Dong, W. F., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvprw.2009.5206848 · Cited by 2,714
Parkhi, O. M., Vedaldi, A., Zisserman, A., & Jawahar, C. V. (2012) · 2012 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2012.6248092 · Cited by 1,533
Analysis of hidden units in a layered network trained to classify sonar targets
Gorman, R. P., & Sejnowski, T. J. (1988) · Neural Networks · doi:10.1016/0893-6080(88)90023-8 · Cited by 1,114 · cited 1×
“Data Virus dataset [54] Sonar dataset [55,56] Peugeot 14 [57] Peugeot 15 [57] EnviroCar [58] Air Quality Index (AQI) [59] Dset-2.”
MethodMachine Learning for Internet of Things Data Analysis: A Survey
Mahdavinejad, M. S., Rezvan, M., Barekatain, M., Adibi, P., Barnaghi, P. M., & Sheth, A. (2017) · Surrey Research Insight Open Access (The University of Surrey) · Cited by 1,008
Machine learning for internet of things data analysis: a survey
Mahdavinejad, M. S., Rezvan, M., Barekatain, M., Adibi, P., Barnaghi, P., & Sheth, A. P. (2018) · Digital Communications and Networks · doi:10.1016/j.dcan.2017.10.002 · Cited by 921 · cited 1×
“Developing a generalized framework is necessary for increasing adoption and awareness of TinyML [62,63].”
BackgroundIoT in smart cities: A survey of technologies, practices and challenges
Syed, A. S., Sierra-Sosa, D., Kumar, A., & Elmaghraby, A. (2021) · Smart Cities · doi:10.3390/smartcities4020024 · Cited by 527 · cited 1×
“The proliferation of IoT edge devices lead to collection of sensitive data from society, wherein transmission of data to the cloud may expose it to hacking and eavesdropping [13].”
BackgroundEdge Machine Learning for AI-Enabled IoT Devices: A Review
Merenda, M., Porcaro, C., & Iero, D. (2020) · Sensors · doi:10.3390/s20092533 · Cited by 523 · cited 4×
“[25] Handwritten Digits Handwritten digit images from (0 to 9) 70,000 images 60,000 10,000”
BackgroundMethodInfluential
Research works printed but not linked: 6
- Banbury, C.R., Reddi, V.J., Lam, M., Fu, W., Fazel, A., Holleman, J., Huang, X., Hurtado, R., Kanter, D., and Lokhmotov, A. (2020). Benchmarking TinyML Systems: Challenges and Direction. arXiv.
- Venzke, M., Klisch, D., Kubik, P., Ali, A., Missier, J.D., and Turau, V. (2020). Artificial Neural Networks for Sensor Data Classification on Small Embedded Systems. arXiv.
- Wong, A., Famouri, M., Pavlova, M., and Surana, S. (2020). TinySpeech: Attention Condensers for Deep Speech Recognition Neural Networks on Edge Devices. arXiv.
- Cai, H., Gan, C., Zhu, L., and Han, S. (2020). TinyTL: Reduce Memory, Not Parameters for Efficient On-Device Learning. arXiv.
- Vu, T.H., Olsson, C., Laptev, I., Oliva, A., and Sivic, J. (2014). Food-101–Mining Discriminative Components with Random Forests, Springer International Publishing. LNCS 8693.
- Chowdhery, A., Warden, P., Shlens, J., Howard, A., and Rhodes, R. (2019). Visual Wake Words Dataset. arXiv.
Cited by
TinyML for Ultra-Low Power AI and Large Scale IoT Deployments: A Systematic Review
Schizas, N., Karras, A., Karras, C., & Sioutas, S. (2022) · Future Internet · doi:10.3390/fi14120363 · Cited by 215
Oliveira, F., Costa, D. G., Assis, F., & Silva, I. (2024) · Internet of Things · doi:10.1016/j.iot.2024.101153 · Cited by 163
A Machine Learning-Oriented Survey on Tiny Machine Learning
Capogrosso, L., Cunico, F., Cheng, D. S., Fummi, F., & Cristani, M. (2024) · IEEE Access · doi:10.1109/access.2024.3365349 · Cited by 247
“In [31] and [37], the authors present the background of TinyML, list the tool for supporting TinyML, and the key enablers ( e.g ., model compression and quantization) for the improvement of TinyML systems.”
MethodAdvancements in TinyML: Applications, Limitations, and Impact on IoT Devices
Elhanashi, A., Dini, P., Saponara, S., & Zheng, Q. (2024) · Electronics · doi:10.3390/electronics13173562 · Cited by 152
TinyML: Tools, applications, challenges, and future research directions
Kallimani, R., Pai, K., Raghuwanshi, P., Iyer, S., & López, O. L. A. (2023) · Multimedia Tools and Applications · doi:10.1007/s11042-023-16740-9 · Cited by 149
“Toovercomethesedrawbacks,TinySpeech libraryhasbeenintroducedtobuildalowcomputationalarchitecturewithalowstorage facility using deep convolutional networks [19].”
Influential
Metadata
- Authors
- Alajlan, N. N., & Ibrahim, D. M.
- Affiliations
- Qassim University (SA)
- Corresponding author
- Ibrahim, Dina M.
- Venue
- Micromachines
- Publisher
- MDPI AG
- Published
- 2022-05-29
- Volume
- 13
- Issue
- 6
- Pages
- 851
- ISSN
- 2072-666X (ISSN-L)
- PMC
- PMC9227753
- OpenAlex
- W4282025104
- Cited by
- 224
- References
- 37
- Citations by year
- 2026: 67 · 2025: 71 · 2024: 56 · 2023: 26 · 2022: 4
- FWCI
- 17.41
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Micromachines (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Information Systems › Internet of Things and AI
- Keywords
- TinyML, edge computing, Internet of Things, ultra-low power electronics, microcontroller, deep learning inference, on-device learning, resource-constrained devices, IoT datasets
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications. https://app.cochat.ai/papers/10.3390/mi13060851
- The paper
- Alajlan, N. N., & Ibrahim, D. M. (2022). TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications. Micromachines. https://doi.org/10.3390/mi13060851
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.