electronics

Review
Advancements in TinyML: Applications, Limitations, and Impact
on IoT Devices
Abdussalam Elhanashi 1, * , Pierpaolo Dini 1 , Sergio Saponara 1 and Qinghe Zheng 2

1 Department of Information Engineering, University of Pisa, 56126 Pisa, Italy; [email protected] (P.D.);
[email protected] (S.S.)
2 School of Intelligent Engineering, Shandong Management University, Jinan 250357, China; [email protected]
* Correspondence: [email protected]

Abstract: Artificial Intelligence (AI) and Machine Learning (ML) have experienced rapid growth
in both industry and academia. However, the current ML and AI models demand significant
computing and processing power to achieve desired accuracy and results, often restricting their use
to high-capability devices. With advancements in embedded system technology and the substantial
development in the Internet of Things (IoT) industry, there is a growing desire to integrate ML
techniques into resource-constrained embedded systems for ubiquitous intelligence. This aspiration
has led to the emergence of TinyML, a specialized approach that enables the deployment of ML
models on resource-constrained, power-efficient, and low-cost devices. Despite its potential, the
implementation of ML on such devices presents challenges, including optimization, processing
capacity, reliability, and maintenance. This article delves into the TinyML model, exploring its
background, the tools that support it, and its applications in advanced technologies. By understanding
these aspects, we can better appreciate how TinyML is transforming the landscape of AI and ML in
embedded and IoT systems.

Keywords: TinyML; embedded systems; internet of things (IoT); Machine Learning optimization;
resource-constrained devices

Citation: Elhanashi, A.; Dini, P.;

Saponara, S.; Zheng, Q. Advancements
in TinyML: Applications, Limitations,
1. Introduction
and Impact on IoT Devices. Electronics The Internet of Things (IoT) aims to leverage edge computing, a paradigm that enables
2024, 13, 3562. https://doi.org/ seamless and real-time processing of data from millions of interconnected sensors and
10.3390/electronics13173562 devices. Edge computing refers to a range of devices and networks near the user, facilitating
Academic Editors: Muhammad
local data processing to enhance efficiency and reduce latency [1]. One significant advantage
Shahid Anwar, Maria Torres Vega and of IoT devices is their requirement for low computing and processing power, as they are
Muhammad Salman Pathan deployed at the network edge, leading to a low memory footprint [2]. IoT devices collect
sensory data and transmit them either to a nearby location or cloud platforms for further
Received: 13 August 2024 processing. Edge computing stores and processes these data, providing the necessary
Revised: 31 August 2024
infrastructure to support distributed computing [3].
Accepted: 4 September 2024
The implementation of edge computing in IoT devices offers several benefits: enhanced
Published: 8 September 2024
security, privacy, and reliability for end-users, reduced latency, and increased availability
and throughput response for applications and services. By processing data closer to the
source, edge computing mitigates the risks associated with transmitting sensitive infor-
Copyright: © 2024 by the authors.
mation across long distances, thereby reducing potential vulnerabilities. Additionally, it
Licensee MDPI, Basel, Switzerland. supports real-time analytics, enabling faster decision making and improving the overall
This article is an open access article responsiveness of IoT systems. This is particularly crucial for time-sensitive applications,
distributed under the terms and such as autonomous vehicles, healthcare monitoring, and industrial automation, where
conditions of the Creative Commons delays in data processing could have significant consequences. Furthermore, edge comput-
Attribution (CC BY) license (https:// ing alleviates the burden on centralized cloud systems, optimizing bandwidth usage and
creativecommons.org/licenses/by/ ensuring that critical functions remain operational even in the face of network disruptions.
4.0/). Additionally, edge devices can collaborate with sensors and cloud platforms to process

Electronics 2024, 13, 3562. https://doi.org/10.3390/electronics13173562 https://www.mdpi.com/journal/electronics

Electronics 2024, 13, 3562 2 of 19

data at the network edge rather than solely relying on cloud processing. This collaboration
results in effective data management, persistence, delivery, and content caching. Edge
computing significantly improves network services, particularly in applications involving
human-to-machine interaction and modern healthcare [4,5].
Recent research has highlighted the potential of implementing Machine Learning (ML)
techniques in various IoT use cases. However, traditional ML models often demand signifi-
cant computing power, processing capabilities, and high memory capacity, limiting their
implementation in IoT devices and applications [6–8]. Current edge computing technology
also faces challenges such as limited transmission capacity and power efficiency, leading
to heterogeneous systems. This necessitates a harmonious and holistic infrastructure for
updating, training, and deploying ML models [9]. Figure 1 presents a detailed taxonomy
of the main applications of TinyML, showcasing the diverse areas where TinyML is mak-
ing substantial impacts. This includes healthcare, environmental monitoring, industrial
Electronics 2024, 13, x FOR PEER REVIEW 3 of 20
automation, and consumer electronics. The taxonomy emphasizes specific use cases and
their benefits, offering a clear perspective on how TinyML is transforming various industries
through its capability to execute complex Machine Learning tasks on ultra-low-power devices.

Figure 1. Comprehensive taxonomy categorizing the primary applications of TinyML. It encompasses

Figure 1. Comprehensive taxonomy categorizing the primary applications of TinyML. It encom-
various domains where TinyML is making significant impacts, highlighting specific use cases and
passes various domains where TinyML is making significant impacts, highlighting specific use cases
their respective advantages. From healthcare and environmental monitoring to industrial automation
and their respective advantages. From healthcare and environmental monitoring to industrial auto-
and consumer
mation electronics,
and consumer this taxonomy
electronics, provides provides
this taxonomy a clear overview of how TinyML
a clear overview of howisTinyML
revolutionizing
is revo-
different sectors
lutionizing with its
different abilitywith
sectors to perform complex
its ability Machine
to perform Learning
complex tasks onLearning
Machine ultra-low-power devices.
tasks on ultra-
low-power devices.
Furthermore, the architecture designed for embedded devices poses another challenge
due to varyingofhardware
Contributions This Workand software requirements, making it difficult to build a standard
ML architecture for IoT networks [10]. Currently, data generated by different devices are
This research review on “Advancements in TinyML” provides a comprehensive
often sent to cloud platforms for processing due to the computationally intensive nature of
overviewimplementations.
network of the current state MLand future
models potential
typically of on
rely TinyML. It makesapplication-specific
deep learning several significant
contributions to the field, which are as follows:
integrated circuits (ASICs) and graphic processing units (GPUs) for data processing, which
• This
require review highlights
substantial power and an memory.
intuitive understanding of TinyML,
Therefore, deploying providing
full-fledged MLdetailed
models in-
on
sights into
IoT devices itsfeasible
is not fundamental
due toprinciples andcomputing
their limited applications, making
power and itstorage
a valuable resource
capacity [11].
for readers
The demand seeking to grasp bothlow-power
for miniaturizing the theoretical and practical
embedded devicesaspects of this emerg-
and optimizing ML
ingfor
models technology.
better power and memory efficiency has paved the way for TinyML. TinyML
•
aims Ittoexplores
implement the wide range ofand
ML models applications
practicesofonTinyML
edge IoTacross different
devices and fields, including
frameworks. It
healthcare,
enables agriculture,
signal processing industrial
on IoT automation,
devices and and environmental
provides embedded intelligence, monitoring,
eliminating
the need to transfer data
demonstrating to cloud platforms
the transformative for processing.
potential of TinyMLThe successful
in these areas.implementation
• TinyML
of on IoTprovides
This review devices can enhance privacy
an in-depth and of
exploration efficiency
the keywhile reducing
enablers operating
driving the ad-
vancement of TinyML technology, while also showcasing a diverse range of use cases
and applications across various industries, highlighting its transformative impact on
resource-constrained environments.
• This paper explores the current and future challenges in TinyML research, offers
practical solutions, and provides a forward-looking perspective on advancing re-
search in this rapidly evolving field, emphasizing the importance of innovation and
Electronics 2024, 13, 3562 3 of 19

costs. Moreover, TinyML is particularly appealing because it can provide on-premise

analytics even in cases of inadequate connectivity [12].

Contributions of This Work

This research review on “Advancements in TinyML” provides a comprehensive
overview of the current state and future potential of TinyML. It makes several signifi-
cant contributions to the field, which are as follows:
• This review highlights an intuitive understanding of TinyML, providing detailed insights
into its fundamental principles and applications, making it a valuable resource for readers
seeking to grasp both the theoretical and practical aspects of this emerging technology.
• It explores the wide range of applications of TinyML across different fields, includ-
ing healthcare, agriculture, industrial automation, and environmental monitoring,
demonstrating the transformative potential of TinyML in these areas.
• This review provides an in-depth exploration of the key enablers driving the ad-
vancement of TinyML technology, while also showcasing a diverse range of use cases
and applications across various industries, highlighting its transformative impact on
resource-constrained environments.
• This paper explores the current and future challenges in TinyML research, offers
practical solutions, and provides a forward-looking perspective on advancing re-
search in this rapidly evolving field, emphasizing the importance of innovation and
collaboration to overcome these hurdles.
The rest of this paper is organized as follows. Section 2 presents an overview of
TinyML Section 3 explores Methodological Insights and Analysis. Conclusions and future
directions are drawn in Section 4.

2. Overview of TinyML
Tiny Machine Learning represents a groundbreaking approach to embedding Ma-
chine Learning algorithms into resource-constrained devices, enabling intelligent decision
making at the edge of the network. This paradigm shift leverages advancements in hard-
ware, software, and algorithms to process data locally on microcontrollers [13], drastically
reducing latency, power consumption [14], and dependency on cloud computing. The
advent of TinyML has opened up a plethora of applications across various domains, such
as healthcare, agriculture, manufacturing, and environmental monitoring, by bringing AI
capabilities to the smallest of devices. One of the primary drivers of TinyML is the con-
tinuous improvement in microcontroller units (MCUs), which are becoming increasingly
powerful yet energy efficient. These MCUs, often equipped with ARM Cortex-M proces-
sors, are now capable of running complex neural network models within a constrained
power budget. This makes it feasible to deploy Machine Learning models in scenarios
where energy efficiency is paramount, such as battery-operated sensors in remote locations
or wearable devices. The software landscape for TinyML is equally vital, with frameworks
like TensorFlow Lite [15] for Microcontrollers, uTensor, and Edge Impulse providing the
tools necessary to optimize and deploy models on these low-power devices. TensorFlow
Lite for Microcontrollers, for instance, enables the conversion of trained models into a
format that can run on MCUs with as little as 16 KB of RAM, demonstrating the remarkable
strides made in reducing the computational footprint of Machine Learning algorithms.
The optimization process often involves techniques such as model quantization, which
reduces the precision of the model’s weights and activations from a 32-bit floating point
to 8-bit integers without significant loss in accuracy. This not only decreases the model
size but also enhances the inference speed and reduces power consumption. Pruning,
another technique, removes redundant parameters from the model, further shrinking its
size and computational requirements. These optimizations are crucial for enabling real-
time inference on devices with limited processing power and memory. The applications
of TinyML are diverse and transformative. In healthcare, for example, wearable devices
equipped with TinyML models can continuously monitor vital signs and detect anomalies,
Electronics 2024, 13, 3562 4 of 19

providing early warnings for conditions like arrhythmias or sleep apnea. In agriculture,
TinyML-powered sensors can monitor soil moisture, temperature, and crop health, enabling
precision farming practices that enhance yield and resource efficiency. In industrial settings,
predictive maintenance systems equipped with TinyML can monitor machinery for signs
of wear and tear, preventing costly breakdowns and downtime. Environmental monitoring
is another area where TinyML is making a significant impact. TinyML sensors can be
deployed in large numbers across forests, oceans, and urban areas to monitor air quality,
detect forest fires, or track wildlife movements. These sensors operate autonomously for
extended periods, thanks to their low power consumption, and provide valuable data for
environmental conservation and management. The ability to process data at the edge also
enhances privacy and security, as sensitive information does not need to be transmitted to
the cloud. This is particularly important in applications like smart homes and healthcare,
where data privacy is a major concern. By processing data locally, TinyML reduces the risk
of data breaches and ensures that personal information remains secure. The development
of TinyML is supported by a vibrant ecosystem of hardware manufacturers, software
developers, and research institutions. Companies like ARM, NVIDIA, and Qualcomm
are continuously pushing the boundaries of what is possible with edge AI, developing
new hardware and software solutions that enhance the capabilities of TinyML. Academic
institutions and research labs are also contributing to the field by developing novel algo-
rithms and optimization techniques that further improve the efficiency and performance of
TinyML models. Despite the significant progress made, TinyML faces several challenges
that need to be addressed to fully realize its potential. One of the main challenges is the
limited computational and memory resources available on microcontrollers, which con-
strain the complexity and accuracy of the models that can be deployed. While techniques
like quantization and pruning help mitigate these limitations, there is an ongoing need for
more efficient algorithms and model architectures that can deliver high performance within
these constraints. Another challenge is the lack of standardized tools and frameworks for
developing and deploying TinyML models. While frameworks like TensorFlow Lite for Mi-
crocontrollers provide a good starting point, there is a need for more comprehensive tools
that can streamline the entire development process, from model training and optimization
to deployment and monitoring. Additionally, the integration of TinyML into existing sys-
tems and workflows requires significant expertise in both Machine Learning and embedded
systems, highlighting the need for better educational resources and training programs. The
future of TinyML holds exciting possibilities, with the potential to revolutionize industries
and improve quality of life in numerous ways. As hardware continues to advance and new
algorithms and optimization techniques are developed, the capabilities of TinyML will
continue to expand, enabling even more sophisticated applications. The convergence of
TinyML with other emerging technologies, such as the Internet of Things (IoT) and 5G, will
further enhance its impact, creating a connected world where intelligent decision making
is embedded in every device Figure 2 presents Tiny Machine Learning (TinyML) is focused
on revolutionizing application development by embedding Machine Learning models
directly into small, resource-constrained devices. This integration allows for advanced data
processing and decision-making capabilities on devices with limited computational power
and storage, enabling a range of new, innovative applications across various industries. By
leveraging TinyML, developers can create more intelligent, responsive, and autonomous
systems, enhancing functionalities and user experiences in everyday technology.
TinyML represents a significant leap forward in the field of Machine Learning, enabling
intelligent processing at the edge of the network on resource-constrained devices. Through
advancements in hardware, software, and algorithms, TinyML is unlocking new applications
across various domains, from healthcare and agriculture to industrial automation and envi-
ronmental monitoring. Despite the challenges, the continued development and adoption of
TinyML hold the promise of a smarter, more efficient, and more connected world.
 bling intelligent processing at the edge of the network on resource-constrained devices.
Through advancements in hardware, software, and algorithms, TinyML is unlocking new
applications across various domains, from healthcare and agriculture to industrial auto-
mation and environmental monitoring. Despite the challenges, the continued develop-
Electronics 2024, 13, 3562 ment and adoption of TinyML hold the promise of a smarter, more efficient, and more 5 of 19
connected world.

Figure
Figure 2. 2. Tiny
Tiny Machine
Machine Learning
Learning (TinyML)
(TinyML) aims
aims to to create
create newnew applications
applications byby integrating
integrating Machine
Machine
Learning models into embedded systems.
Learning models into embedded systems.

TinyMachine
Tiny MachineLearning
Learning(TinyML)
(TinyML)represents
represents aa significant
significant advancement
advancement in in the
theintegra-
inte-
gration of Machine Learning with Internet of Things (IoT) devices. This technologyenables
tion of Machine Learning with Internet of Things (IoT) devices. This technology ena-
thethe
bles deployment
deployment of of
Machine
Machine Learning
Learning models
models ononultra-low-power
ultra-low-powerdevices,
devices,allowing
allowingfor
forreal-time
real-timedata
dataprocessing
processing and
anddecision
decisionmaking
making without
without reliance onon
reliance cloud computing.
cloud computing. The
Thefollowing
followingsections provide
sections provideanan
overview
overviewof of
thethe
state-of-the-art
state-of-the-artin in
TinyML,
TinyML,its its
applications,
applica-
challenges, and future directions. TinyML, which involves deploying
tions, challenges, and future directions. TinyML, which involves deploying Machine Machine Learning
algorithms on resource-constrained devices, has become a transformative technology in
the Internet of Things (IoT) domain. This overview highlights recent advancements and
current trends in TinyML for IoT applications. Modern microcontrollers, such as those from
the ARM Cortex-M series, and specialized processors like Google’s Edge TPU and Intel’s
Movidius Myriad X, are increasingly used to run TinyML models efficiently, handling
computational needs while consuming minimal power [16,17]. Innovations in hardware
accelerators, including FPGAs and custom ASICs, enhance the performance and efficiency
of ML operations on edge devices by optimizing tasks like matrix multiplications and
convolution operations [18,19]. Techniques for model compression, such as quantization,
pruning, and knowledge distillation, are crucial for reducing the size of ML models without
significantly compromising performance, with quantization converting model weights to
lower-bit representations to reduce memory usage [11,20]. Efficient algorithms designed
for resource-constrained environments include models like MobileNet and EfficientNet,
which minimize computational and memory requirements [21,22]. Software frameworks
and tools like TensorFlow Lite for Microcontrollers (TFLite Micro) which has been made by
Google that enables deployment on embedded devices with limited resources, providing
a lightweight runtime and tools for model conversion and optimization [23], while plat-
forms like Edge Impulse simplify the development and deployment of TinyML models
on edge devices [24]. In application areas, TinyML facilitates real-time data processing
on smart sensors and actuators for tasks such as anomaly detection, predictive mainte-
nance, and environmental monitoring [10,25], as well as continuous health monitoring and
activity recognition in wearables, improving privacy by reducing the need for constant
cloud communication [26,27]. Challenges and future directions include optimizing power
consumption to extend battery life in IoT devices [28], addressing security and privacy
concerns through techniques like local data encryption and secure model updates [29],
and exploring solutions for managing and updating large numbers of devices as IoT de-
ployments scale [30]. Figure 3 presents a detailed framework that depicts the cohesive
 Electronics 2024, 13, 3562 6 of 19

integration of Internet of Things (IoT) applications with cloud computing, edge computing,
and TinyML technologies. It highlights the synergistic interplay among these components,
illustrating how data are processed and managed across various layers, from edge devices
to centralized cloud infrastructures. The framework underscores TinyML’s critical role in
delivering on-device intelligence and facilitating real-time analytics, which significantly
boosts the efficiency and responsiveness of IoT systems. By demonstrating the interplay
between local processing at the edge and broader data management in the cloud, this
Electronics 2024, 13, x FOR PEER REVIEW 7 of 20
framework showcases how these technologies collectively enhance the performance and
capabilities of IoT solutions.

Figure
Figure A comprehensive
3. comprehensive
3. A framework
framework illustrating
illustrating the the integration
integration of IoT
of IoT applications
applications with
with cloud
cloud
computing, edge computing, and TinyML. This framework highlights the synergistic
computing, edge computing, and TinyML. This framework highlights the synergistic interaction interaction between
these technologies,
between showcasing
these technologies, how data
showcasing howaredata
processed and managed
are processed at different
and managed levels from
at different edge
levels
from edgeto
devices devices to centralized
centralized cloudIt systems.
cloud systems. emphasizes It emphasizes the role
the role of TinyML in of TinyML
enabling in enabling
on-device on-
intelligence
device intelligence
and real-time and real-time
analytics, analytics,
enhancing enhancing
the overall efficiencytheand
overall efficiency and
responsiveness responsiveness
of IoT solutions. of
IoT solutions.
Deep Learning is a core component of TinyML, enabling complex tasks like image
Deep Learning
recognition is a core
and natural component
language of TinyML,
processing enabling complex tasks
on resource-constrained likeHowever,
devices. image
recognition and naturaldemands
their computational language require
processing on resource-constrained
specialized hardware (HW) devices. However,on
to run efficiently
their computational
these demands
small platforms. Tensorrequire specialized
Processing hardware
Units (TPUs) (HW)
are one suchtospecialized
run efficiently on
hardware
these small platforms.
designed to accelerateTensor
DNNProcessing Units
operations, (TPUs) are
particularly forone such specialized
inferencing tasks, byhardware
optimizing
designed to accelerate
for the low power and DNN operations,
latency particularly
requirements for inferencing
of TinyML. tasks,processing
FP5A, a flexible by optimizing
unit, is
foranother
the lowexample
power andthatlatency
supports various bit-widths
requirements to enhance
of TinyML. FP5A,efficiency
a flexibleinprocessing
running DNNs unit,on
is embedded systems.
another example Convolutional
that Neuralbit-widths
supports various Networksto (CNNs),
enhance a type of DNN
efficiency inparticularly
running
effective
DNNs for image processing
on embedded tasks, have been
systems. Convolutional adapted
Neural into microarchitectures
Networks (CNNs), a type oflike DNNCNN
Micro, which
particularly is optimized
effective for image to processing
fit within the tight
tasks, power
have beenand memory
adapted intoconstraints typical
microarchitec-
of TinyML
tures like CNN applications.
Micro, which These components
is optimized to fitwork in the
within tandem
tight to ensure
power and that sophisticated
memory con-
Machine
straints Learning
typical models
of TinyML can be deployed
applications. Theseeffectively
components in embedded environments,
work in tandem to ensuredriving
that
innovation in edge AI.
sophisticated Machine Learning models can be deployed effectively in embedded envi-
ronments, driving innovation in edge AI.
The integration of cloud computing, edge computing, and TinyML forms a powerful
framework for IoT applications. cloud computing provides scalable storage and pro-
cessing power, enabling complex analytics and data management. Edge computing brings
processing closer to the data source, reducing latency and enhancing real-time decision
making. TinyML, with its efficient Machine Learning models, facilitates low-power, on-
Electronics 2024, 13, 3562 7 of 19

The integration of cloud computing, edge computing, and TinyML forms a powerful
framework for IoT applications. cloud computing provides scalable storage and processing
power, enabling complex analytics and data management. Edge computing brings process-
ing closer to the data source, reducing latency and enhancing real-time decision making.
TinyML, with its efficient Machine Learning models, facilitates low-power, on-device an-
alytics, making IoT devices smarter and more responsive. This synergistic framework
enhances data processing efficiency, reduces bandwidth requirements, and supports ad-
vanced applications like predictive maintenance, smart homes, and healthcare monitoring,
creating a robust and intelligent IoT ecosystem.

3. Methodological Insights and Analysis

We conducted a thorough search of online databases, including Google Scholar, IEEE
Xplore, and PubMed, using specific keywords related to TinyML and its applications in
IoT. This initial search aimed to gather a broad range of relevant studies and reports,
ensuring a comprehensive understanding of the current landscape. Our search strategy
was meticulously designed to capture the most pertinent literature, encompassing both
foundational research and cutting-edge advancements in the field.
To ensure the quality and relevance of the selected papers, we established stringent
criteria for selection. These criteria included the publication date, with a focus on the most
recent studies from 2020 to 2023, to capture the latest developments and trends. We also
considered the relevance of TinyML and IoT, ensuring that the papers directly addressed the
intersection of these technologies. Additionally, we prioritized the availability of full-text
papers in English to facilitate thorough analysis and accessibility for a broader audience.
From the selected papers, we extracted key information related to optimization techniques,
applications, challenges, and future directions of TinyML. This extraction process was
systematic and detailed, allowing us to organize the information comprehensively. We
categorized the data to facilitate a structured analysis, ensuring that each aspect of TinyML
was thoroughly examined. Our analysis involved identifying common themes, significant
advancements, and gaps in the current research. We synthesized the findings into coherent
sections, highlighting the main contributions of TinyML to various domains. This synthesis
provided a clear and detailed overview of the state of the art in TinyML, emphasizing its
potential and identifying areas for future research. The synthesized information was re-
viewed and validated by multiple authors to ensure accuracy and comprehensiveness. This
collaborative review process involved cross-checking the data and resolving any discrepan-
cies through discussion and consensus. By leveraging the expertise of multiple authors, we
ensured that the final document was both accurate and comprehensive, providing valuable
insights into the advancements in TinyML.

3.1. Tools and Framework for TinyML

The survey highlights the diversity and innovation within the TinyML landscape,
focusing on tools and frameworks pivotal for IoT and edge devices. Table 1 presents a
detailed examination of several tools and frameworks commonly utilized in TinyML appli-
cations. It includes information on the developers of each tool, the platforms they support,
the typical applications they are used for, and key references associated with them. The
table features well-known tools such as TensorFlow Lite, uTensor, Edge Impulse, NanoEdge
AI Studio, and PyTorch Mobile. Each tool’s unique features, diverse functionalities, and
impact on the advancement of TinyML technology are highlighted, providing insight into
their roles and contributions within the field. TensorFlow Lite [12]: Developed by the
Google Brain Team, TensorFlow Lite supports a range of platforms including Android,
iOS, embedded Linux, and microcontrollers. It is widely utilized for applications such as
image/audio classification and object detection. The framework’s versatility and robust
model support make it ideal for real-time processing in smart cameras, voice-activated
assistants, and various IoT devices. TensorFlow Lite’s comprehensive cross-platform sup-
port enables seamless deployment of Machine Learning models on diverse hardware, from
Electronics 2024, 13, 3562 8 of 19

smartphones to embedded systems, enhancing its utility in creating intelligent applications

across multiple environments. uTensor [31]: A product of ARM, uTensor is optimized for
ARM’s microcontroller ecosystem and supports platforms like Mbed and ST K64 ARM
boards. It excels in image classification and gesture recognition, making it crucial for smart
devices and wearables. uTensor’s design emphasizes low latency and power efficiency,
essential for battery-powered devices. This focus on optimization for ARM hardware
ensures reliable and continuous operation in portable and wearable technology, position-
ing uTensor as a key player in enhancing the functionality of small, resource-constrained
devices. PyTorch Mobile [32]: Originating from Meta AI, PyTorch Mobile extends the
capabilities of the popular PyTorch framework to mobile and embedded devices, support-
ing Android, iOS, and Linux CPU platforms. It is tailored for applications in computer
vision and natural language processing (NLP), including image recognition, object de-
tection, and speech recognition. PyTorch Mobile’s strength lies in its ability to efficiently
run complex models on edge devices, enabling advanced AI capabilities in mobile and
embedded systems. This flexibility and computational power make it a top choice for devel-
opers aiming to deploy sophisticated Machine Learning models across various platforms,
thereby enhancing mobile and embedded applications with cutting-edge AI features. Edge
Impulse [33]: This tool is particularly adept at edge-based asset tracking and predictive
maintenance. Edge Impulse supports a variety of hardware platforms and is designed for
rapid prototyping and deployment of TinyML models in edge devices. Its emphasis on
ease of use and integration into existing workflows makes it suitable for applications in
industrial and commercial environments where real-time data processing and analysis are
critical. The platform’s focus on practical deployment and actionable insights helps organi-
zations leverage edge-based analytics for improved operational efficiency and predictive
capabilities. NanoEdge AI Studio [34]: Focused on industrial applications, NanoEdge AI
Studio specializes in anomaly detection and predictive maintenance. It provides tools for
building and deploying Machine Learning models that can detect anomalies in operational
data, contributing to increased efficiency and reduced downtime in industrial settings. Its
emphasis on industrial-grade solutions highlights the importance of robust and reliable AI
tools for critical operational environments, where precision and reliability are paramount.
MediaPipe [35], another tool developed by Google, provides a cross-platform frame-
work for building real-time perception pipelines. It supports Android, iOS, Linux, and
web platforms, and is widely used in applications such as hand and face tracking and pose
estimation. MediaPipe’s ability to handle complex computer vision tasks efficiently on
edge devices makes it a crucial tool in the TinyML ecosystem.
Web of Science AI [36], developed by Clarivate, is a powerful research analytics tool
designed to assist researchers in curating, analysing, and interpreting vast amounts of
scientific literature. It leverages Artificial Intelligence to streamline the process of discover-
ing trends, identifying key research areas, and understanding the impact of publications
across various disciplines. Web of Science AI’s integration with platforms like Windows,
macOS, and Web allows researchers to access comprehensive datasets and utilize advanced
metrics for citation analysis, collaboration networks, and research performance benchmark-
ing. This tool is instrumental in providing insights that drive research strategies, funding
opportunities, and policy decisions, making it a cornerstone for academic and industrial
research environments.
On the other hand, OpenVINO [37], developed by Intel, is an open-source toolkit
designed for optimizing and deploying deep learning models across a wide range of
hardware platforms, including CPUs, GPUs, and specialized accelerators. It is particularly
well-suited for computer vision applications, such as object detection, image segmentation,
and facial recognition, due to its robust support for convolutional neural networks (CNNs)
and other deep learning architectures. OpenVINO’s versatility extends to its compatibility
with Linux, Windows, and macOS, providing developers with tools for model optimization,
inference acceleration, and seamless deployment across edge and cloud environments.
Its ability to convert and optimize models from popular frameworks like TensorFlow
Electronics 2024, 13, 3562 9 of 19

and PyTorch further enhances its utility in developing high-performance AI applications.

Together, Web of Science AI and OpenVINO represent cutting-edge tools in their respective
domains, driving innovation in research analytics and deep learning inference. Each
tool has unique strengths tailored to specific platforms and applications. TensorFlow
Lite and PyTorch Mobile offer broad support and versatility, making them suitable for a
wide range of use cases across different industries. uTensor and Edge Impulse provide
specialized solutions for particular hardware and application needs, with uTensor focusing
on ARM-based microcontrollers and Edge Impulse excelling in edge-based asset tracking
and predictive maintenance. NanoEdge AI Studio caters to industrial applications with
its focus on anomaly detection and predictive maintenance, offering critical insights and
operational efficiencies in industrial settings.

Table 1. Overview of various tools and frameworks used in TinyML applications, detailing their
developers, supported platforms, typical applications, and key references. This includes popular
tools such as TensorFlow Lite, uTensor, Edge Impulse, NanoEdge AI Studio, PyTorch Mobile, and
MediaPipe, highlighting their diverse capabilities and contributions to the field.

Tool/Framework Developer Supporting Platforms Applications References

Image and Audio
TensorFlow Lite Google Brain Team Embedded Linux, Android [12]
recognition, Object Tracking
Audio Classification,
uTensor ARM Mbed, ST K64 ARM Boards [31]
Gesture Recognition
Android, iOS, Embedded Asset Tracking,
Edge Impulse Zach Shelby, Jan Jongboom [33]
Linux, Microcontrollers Predictive Maintenance
NanoEdge AI Studio Cartesiam Android, Linux Predictive Maintenance [34]
PyTorch Mobile Meta AI Android, iOS, Linux CPU Computer Vision [32]
Hand/Face Tracking,
MediaPipe Google Android, iOS, Linux, Web [35]
Pose Estimation
Research Analytics,
Web of Science Clarivate Windows, macOS, Web [36]
Data Curation
OpenVINO Intel Linux, Windows, macOS Deep Learning, Inference [37]

The continuous development and refinement of these tools, driven by both major tech
companies and innovative startups, are essential for advancing the field of TinyML. This
diverse ecosystem enables smarter, more efficient IoT and edge devices by leveraging the
latest advancements in hardware, software, and Machine Learning algorithms. As TinyML
capabilities continue to expand, these tools and frameworks will play a crucial role in
bringing intelligent decision making to the edge, transforming industries, and improving
quality of life through enhanced technology solutions.

Programming Languages in TinyML Development

The development of TinyML applications involves a range of programming languages,
each tailored to meet the unique demands of building Machine Learning models that can
operate on resource-constrained devices like microcontrollers and edge systems. These
languages are integral to different stages of TinyML development, from model training and
optimization to deployment and execution on tiny hardware. Below is an analysis of the
primary programming languages used in the TinyML ecosystem, highlighting their specific
roles, strengths, and limitations.
• Python
Python is widely used for the initial development and training of Machine Learning
models. It is the predominant language for frameworks like TensorFlow and PyTorch. Ease
of Use: Python’s simple syntax and extensive libraries make it accessible for both beginners
and experts. Python has a vast array of libraries and frameworks (e.g., NumPy, TensorFlow,
PyTorch) that streamline the development of Machine Learning models. A large and active
community provides extensive resources, tutorials, and support. Python is an interpreted
Electronics 2024, 13, 3562 10 of 19

language and is slower than compiled languages like C++. This can be a limitation in
environments where performance is critical. Python’s runtime environment consumes
significant memory and processing power, which is problematic for resource-constrained
edge devices.
• C/C++
C and C++ are commonly used for the deployment and optimisation of TinyML
models on microcontrollers. Libraries like TensorFlow Lite Micro and uTensor are written
in C++ to ensure efficiency.: C/C++ provides high performance and low-level control
over hardware, making it ideal for optimizing code to run on constrained devices. C/C++
allows precise management of memory usage, which is critical for devices with limited
resources. These languages are supported by virtually all microcontrollers and embedded
systems. C/C++ has a steeper learning curve compared to Python, requiring more effort
to write and maintain code. Unlike Python, C/C++ lacks high-level libraries for Machine
Learning, making the development process more time-consuming.
• JavaScript
JavaScript is used for developing web-based TinyML applications. Frameworks like
TensorFlow.js allow models to be deployed directly in the browser or on Node.js servers
JavaScript’s primary advantage is its seamless integration with web technologies, making it
possible to run ML models in the browser. JavaScript is platform-independent, running on
any device with a web browser. The JavaScript ecosystem is large, with a wealth of libraries
and tools available for web-based ML. Like Python, JavaScript is an interpreted language,
leading to performance limitations in resource-constrained environments. JavaScript lacks
the extensive Machine Learning libraries and frameworks found in Python, making it less
suitable for complex ML tasks.

3.2. Applications of TinyML

Table 2 provides a comprehensive survey of state-of-the-art TinyML applications,
illustrating its transformative impact across multiple domains by embedding Machine
Learning capabilities into resource-constrained devices. In the realm of speech-based
applications, TinyML facilitates real-time speech detection, recognition, and enhancement,
as explored by Fedorov et al. [38] and Kwon and Park [39]. These advancements enable
devices to understand and process spoken language locally, improving user interaction
and maintaining privacy by minimizing cloud dependency. Vision-based applications,
covered by Paul et al. [40] and Mohan et al. [41], leverage TinyML for tasks such as image
classification, object detection, and gesture recognition.
These capabilities allow devices to interpret visual data directly at the edge, enhancing
performance in areas like surveillance, security, and interactive systems, where immediate
visual analysis is crucial. Data pattern classification, as demonstrated by Ren et al. [13],
focuses on classifying and compressing data patterns on edge devices, optimizing data
management and reducing bandwidth usage. This is essential for IoT devices that collect
and analyse data continuously, enabling more efficient operations and conserving resources.
In the healthcare domain, Dutta and Bharali [42] highlight how TinyML supports diagnostic
applications such as respiratory monitoring and pose estimation. These applications
enable continuous, real-time health monitoring and early detection of medical conditions,
which is particularly valuable in remote or underserved areas with limited access to
healthcare professionals. Edge computing benefits from TinyML, as discussed by Guleria
et al. [43], by enhancing the performance and efficiency of edge devices through local
data processing. This reduces the reliance on cloud infrastructure, speeds up decision
making, and improves system scalability and reliability. Prado et al. [44] illustrate the
application of TinyML in autonomous vehicles, where it is used for navigation and object
detection, crucial for the safe and efficient operation of autonomous systems. By enabling
real-time sensory data processing, TinyML helps autonomous vehicles detect obstacles
and make decisions without constant cloud communication, thus improving safety and
Electronics 2024, 13, 3562 11 of 19

functionality. Overall, the table underscores TinyML’s broad and transformative potential
across diverse applications, driving advancements in real-time data processing, efficiency,
and intelligent decision making at the edge of the network. The table presents an overview
of the various domains in which TinyML technology is applied. It encompasses a range
of areas such as speech recognition, computer vision, pattern classification, healthcare
diagnostics, edge computing, and autonomous vehicles. For each category, the table
includes a concise description of the application and provides references for additional
reading. This comprehensive summary aims to highlight the versatility and impact of
TinyML in different sectors. Environmental monitoring is another application area where
TinyML is used to monitor conditions such as air quality, temperature, and humidity [45].
This application is crucial for maintaining environmental health and safety and providing
timely data for decision making. In wearable technology, TinyML is integrated into devices
for activity tracking, health monitoring, and fitness applications [46]. This integration
allows for more personalized and accurate monitoring of users’ health and fitness metrics.
Smart agriculture leverages TinyML for precision farming, crop monitoring, and livestock
management [47]. This application helps in optimizing agricultural practices, leading to
better yield and resource management. Security and surveillance systems utilize TinyML
for real-time threat detection and monitoring capabilities [48]. This application enhances
the effectiveness of surveillance systems, providing timely alerts and improving security.
TinyML applications, while transformative, face significant cybersecurity challenges due
to their resource constraints, which limit the implementation of robust security measures.
The small footprint of these devices increases their vulnerability to physical tampering
and cyberattacks such as model extraction, adversarial attacks, and data poisoning. To
address these issues, lightweight cryptography can secure data transmission and storage
without overburdening device resources, while secure boot processes and firmware updates
prevent unauthorized code execution. Real-time anomaly detection can help identify
and mitigate threats proactively. Additionally, protecting models through obfuscation
and secure enclaves safeguards intellectual property. Current trends enhancing TinyML
cybersecurity include federated learning, which maintains data privacy by keeping data
local while enabling collaborative model training, adversarial training to fortify models
against manipulated inputs, blockchain integration for decentralized, tamper-resistant
data exchange, and edge-to-cloud security frameworks for comprehensive protection. As
TinyML expands across domains such as healthcare, surveillance, and autonomous vehicles,
these advanced cybersecurity strategies are essential for protecting sensitive data, ensuring
system integrity, and maintaining user trust.
TinyML utilizes lightweight algorithms optimized for resource-constrained edge
devices. The most common algorithms include variants of neural networks such as Con-
volutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and
Decision Trees. These algorithms are tailored to run efficiently on microcontrollers with
limited memory, processing power, and energy.
Key features of these algorithms include model compression, quantization, and prun-
ing. Model compression reduces the model size without significantly sacrificing accuracy.
Quantization converts model parameters to lower precision, such as from 32-bit floating
point to 8-bit integers, reducing memory usage and computational demand. Pruning elimi-
nates less critical parts of the network, streamlining the model further. TinyML architectures
often follow a modular design, with layers dedicated to input preprocessing, feature extrac-
tion, and decision making. For instance, in a CNN, the convolutional layers handle feature
extraction, while fully connected layers make decisions based on the extracted features.
Practical examples include voice recognition systems like Google’s Keyword Spotting,
image classification for real-time object detection, and anomaly detection in IoT devices.
Future improvements in TinyML will focus on developing more efficient algorithms, en-
hancing model interpretability, and improving adaptability to various hardware platforms,
enabling broader adoption in fields like healthcare, agriculture, and smart cities.
Electronics 2024, 13, 3562 12 of 19

Table 2. TinyML applications across various domains. This table summarizes the different application
areas where TinyML is utilized, including speech, vision, data pattern classification, health diagnosis,
edge computing, and autonomous vehicles. Each entry provides a brief description and references
for further reading.

Application Area Description References

Includes speech detection, recognition, and
Speech-Based Applications [38,39]
enhancement for devices.
Encompasses image classification, object
Vision-Based Applications [40,41]
detection, and gesture recognition.
Focuses on classifying and compressing data
Data Pattern Classification [13]
patterns on edge devices.
Involves diagnostic applications like
Health Diagnosis [42]
respiratory monitoring and pose estimation.
Utilizes TinyML for enhancing edge device
Edge Computing [43]
performance and efficiency.
Applies TinyML for navigation and object
Autonomous Vehicles [44]
detection in autonomous systems.
Used for monitoring environmental
Environmental Monitoring conditions such as air quality, [45]
temperature, humidity.
Integrates TinyML in wearable devices for
Wearable Technology activity tracking, health monitoring, and [46]
fitness applications
Implements TinyML for precision farming,
Smart Agriculture [47]
crop monitoring, and livestock management
Enables intelligent surveillance systems with
Security and Surveillance real-time threat detection and [48]
monitoring capabilities.

Emulators for TinyML Development

Emulators play a crucial role in the development and testing of TinyML applications,
allowing developers to simulate the behaviour of microcontrollers and other edge devices in
a controlled environment before deployment on actual hardware. These tools are essential
for debugging and ensuring the reliability of code, particularly in resource-constrained
settings. Open-source emulators like QEMU and Renode are popular choices for their
cost-effectiveness and flexibility. QEMU supports a wide range of architectures, including
ARM, making it a versatile option for simulating microcontroller environments, though
it requires manual setup and configuration. Renode, specifically designed for IoT and
edge devices, offers advanced debugging capabilities and can simulate entire networks
of devices, making it particularly useful for complex TinyML projects. However, open-
source emulators often come with the challenges of complex setup and potentially slower
performance, which can affect the accuracy of performance testing. On the other hand,
commercial emulators such as Keil MDK and Proteus provide user-friendly interfaces
and professional support, making them easier to use and integrate into development
workflows. Keil MDK is particularly well suited for ARM-based microcontrollers, offering
advanced features and seamless integration with hardware debugging tools. Proteus
supports a wide range of microcontroller platforms and provides extensive simulation
features. Despite their advantages, commercial emulators can be expensive, limiting their
accessibility to smaller development teams or hobbyists, and may also lead to vendor lock-
in, restricting flexibility in hardware choices. Table 3 presents a comprehensive comparison
of TinyML-compatible hardware, focusing on their processors, memory, computational
power, power consumption, and typical use cases. The STWIN platform [49] features an
Electronics 2024, 13, 3562 13 of 19

ARM Cortex-M4 processor with 640 KB of SRAM and 2 MB of Flash memory, operating at
80 MHz. Its low power consumption makes it well suited for industrial IoT applications,
where reliability and efficient sensor integration are critical. This hardware is designed for
applications such as condition monitoring and vibration analysis, leveraging its industrial-
grade sensors and BLE connectivity to provide robust data transmission and extensive
interfacing capabilities. The Arduino Nano 33 BLE [50], also equipped with an ARM Cortex-
M4 processor, offers 256 KB of SRAM and 1 MB of Flash memory, running at 64 MHz.
It is known for its ultra-low power consumption, making it ideal for battery-operated
devices. Its integrated Bluetooth 5.0 and 9-axis IMU facilitate advanced applications like
wearable health monitoring and motion tracking, where its compact design and precise
sensing are key advantages. The Edge TPU [51], a custom ASIC, excels in hardware
acceleration for Machine Learning tasks, delivering 4 TOPS (Tera Operations Per Second)
with very low power consumption (approximately 2 W). This specialized hardware is
optimized for inferencing tasks such as image classification and object detection, providing
high computational efficiency and minimal latency, which are essential for real-time AI
applications in edge computing environments. The Raspberry Pi Pico [52] features an
ARM Cortex-M0+ dual-core processor, with 264 KB of SRAM and 2 MB of Flash memory,
operating at 133 MHz. It balances low power consumption with adequate performance
for basic Machine Learning tasks and educational projects. Its cost-effectiveness and
simplicity make it a popular choice for prototyping and learning applications, where
straightforward functionality and affordability are prioritized. This analysis underscores
how each hardware option caters to different TinyML requirements, offering a range of
memory capacities, power efficiencies, and computational capabilities to address various
application needs effectively.

Table 3. Detailed specifications and capabilities of TinyML hardware options.

Hardware Processor/MCU Special Features Typical Use Cases References

ARM Industrial-grade, sensors,
STWIN Industrial IoTs [49]
Cortex-M4 BLE connectivity
ARM Bluetooth
Arduino Nano 33 BLE Wearable, remote sensing [50]
Cortex-M4 5.0, 9-axis IMU
Hardware acceleration Image classification, and
Edge TPU Custom ASIC [51]
For ML object detection
ARM Simple Machine
Raspberry Pi Dual-core processors [52]
Cortex-M0+ Learning Tasks

3.3. Limitations of TinyML

The deployment of TinyML on IoT and edge devices presents several intricate chal-
lenges that affect the performance and feasibility of Machine Learning models in resource-
constrained environments. One of the foremost concerns is power consumption. Edge
devices are typically limited in their energy resources, making it challenging to run complex
ML models efficiently. To mitigate this, techniques such as model quantization and pruning
are vital. Quantization reduces the precision of the model’s weights and activations, while
pruning eliminates less significant components, both contributing to reduced energy con-
sumption without substantial losses in functionality [36]. For example, Google’s Edge TPU
employs quantization to deliver high performance with minimal power usage, illustrating
how these techniques can effectively balance power and performance.
Memory constraints further complicate these issues, as the limited memory on edge de-
vices restricts the size and complexity of deployable models. This often necessitates the use
of compact neural network architectures and model compression strategies. Research into
lightweight models, such as MobileNet and EfficientNet, highlights how these architectures
are specifically designed to fit within memory limitations while maintaining competitive
performance [53]. For instance, MobileNet’s use of depthwise separable convolutions
Electronics 2024, 13, 3562 14 of 19

significantly reduces the model size and computational load, making it suitable for edge
devices with limited memory. Additionally, the processing capabilities of edge devices are
generally lower than those of traditional servers, leading to slower inference times and
reduced responsiveness, particularly in real-time applications. To address this, hardware
accelerators like specialized processors and FPGAs are employed. These accelerators are
designed to handle specific computations more efficiently, thus enhancing performance
without imposing excessive computational demands [54]. An example is Intel’s Movid-
ius Myriad X, which is optimized for neural network inference and provides substantial
performance improvements over conventional processors. Another significant issue is the
lack of standardization in TinyML frameworks and benchmarks, which results in inconsis-
tencies and compatibility issues across different devices and applications. Standardizing
tools, protocols, and evaluation metrics is crucial for facilitating deployment and ensuring
interoperability [55]. For instance, the lack of uniform benchmarks makes it challenging
to compare the performance of TinyML models across different platforms, hindering the
development of universally applicable solutions. Latency is another critical challenge,
particularly for real-time applications that require extremely low latency. Achieving such
performance with TinyML on edge devices is difficult due to their limited processing power
and memory. Techniques to address latency include optimizing algorithms and leveraging
edge–cloud computing architectures, where edge devices handle simple tasks and offload
complex computations to cloud servers [56]. This hybrid approach can help mitigate the
processing constraints of edge devices. Data privacy is also a major concern when handling
sensitive information locally on edge devices. Ensuring robust privacy-preserving tech-
niques is essential to safeguard user data from unauthorized access and breaches [57]. For
example, techniques like federated learning enable models to be trained across multiple
devices without transferring raw data, thus enhancing privacy. Deploying models on
millions of distributed edge devices presents logistical challenges in updating and main-
taining these models. Streamlining this process requires effective strategies to manage
and distribute updates across a vast network [58]. Solutions such as over-the-air updates
and version control mechanisms are critical for ensuring that all devices remain current
and functional. Addressing these challenges through innovative optimization techniques,
efficient hardware design, and the development of universal standards is essential for
advancing TinyML’s integration into diverse IoT and edge computing scenarios. Table 4
details the challenges and considerations for deploying TinyML on edge devices

Table 4. Challenges and considerations in deploying TinyML on edge devices. This table outlines key
obstacles such as power consumption, memory constraints, processing capacity limitations, and the
lack of standardization, along with relevant references for further reading.

Challenge Description References

Edge devices have limited power, making
Power Consumption [36]
efficient ML model execution challenging.
TinyML models need to operate within
Memory Constraints [53]
limited memory available on edge devices.
Edge devices have lower processing
Processing Capacity [54]
capabilities compared to traditional servers.
Lack of standardized frameworks and
Standardization benchmarks for TinyML deployment. [55]

Real-time applications require low latency,

Latency Requirements which can be difficult to achieve with [56]
TinyML on edge devices.
Handling sensitive data locally on edge
Data Privacy devices necessitates strong [57]
privacy-preserving techniques.
Deploying and updating models on millions
Model Deployment of distributed edge devices can be [58]
logistically challenging.
Electronics 2024, 13, 3562 15 of 19

4. Conclusions and Future Directions

4.1. Future Directions
Future research directions for advancing TinyML in the context of IoT and edge devices
highlight several critical areas of focus. One significant area is the development of energy-
efficient TinyML models, which aims to optimize Machine Learning algorithms for low-
power devices without compromising accuracy (Immonen and Hämäläinen, 2022) [59]. This
involves implementing model optimization techniques such as quantization, pruning, and
hardware integration to enhance power efficiency while maintaining model performance.
Another crucial direction is enhanced memory management, which seeks to improve
techniques for managing memory to accommodate more complex ML models on resource-
constrained devices (Han and Siebert, 2022) [60]. Table 5 provides an overview of significant
research areas in TinyML. It includes initiatives aimed at creating energy-efficient models,
improving memory management, optimizing the integration of edge and cloud computing,
and advancing standardization efforts. Each area is detailed with pertinent references
to recent research. This includes employing memory optimization techniques, efficient
data handling strategies, and advanced compression algorithms to manage and reduce
the memory footprint of models. Collaborative edge–cloud models are also a focal point,
exploring how edge devices and cloud servers can work together to optimize computing
resources and performance Bao et al. [61]. This approach leverages hybrid processing,
effective data aggregation, and load balancing to combine the strengths of both edge and
cloud computing, reducing latency and bandwidth usage while enhancing overall system
efficiency. Additionally, standardization efforts are essential for developing consistent
frameworks and benchmarks for TinyML Shafique et al. [55]. Establishing standardized
tools, benchmarks, and metrics will facilitate the creation, deployment, and evaluation
of TinyML models across diverse platforms, ensuring consistency, interoperability, and
improved integration. Research in this area, as discussed in [47], aims to make AI more
environmentally friendly by reducing energy consumption and resource usage. TinyML
for real-time applications is a critical focus, particularly for mission-critical tasks such as
autonomous vehicles. Ensuring that TinyML models can operate effectively in real-time
scenarios is essential for applications that require immediate decision making and response.
The challenges and solutions in this area are explored in [62]. Hardware-software co-design
represents an integrated approach where both hardware and software are designed together
to optimize performance in TinyML applications. This co-design strategy, discussed by [39],
aims to create more efficient and powerful systems by aligning the capabilities of hardware
with the requirements of software.
Subsequent investigations in TinyML should prioritise the following domains in order
to tackle the recognised obstacles and propel the discipline forward:
• Compact and power-saving TinyML models: It is essential to develop Machine Learn-
ing models that are optimised for energy conservation while maintaining high accuracy
in order to ensure the sustainable implementation of TinyML on-edge devices. Re-
search should investigate innovative methods for model quantisation pruning and
compression. Enhanced Memory Management: It is crucial to enhance memory man-
agement strategies in order to enable the deployment of more intricate Machine Learn-
ing models on edge devices with limited memory. This involves creating algorithms
that adaptively optimise memory utilisation according to the available resources. Col-
laborative edge–cloud models involve the exploration and optimisation of computing
models that leverage the combined capabilities of edge devices and cloud servers to
efficiently handle Machine Learning workloads. This approach is particularly useful
in overcoming the processing capacity constraints of edge devices. Efforts to establish
a set of standardised practices or guidelines: Facilitating the development of standard-
ised frameworks and benchmarks for TinyML would promote its wider adoption and
compatibility across many platforms and applications.
Electronics 2024, 13, 3562 16 of 19

Table 5. Table summarizing key research directions in TinyML, including efforts to develop energy-
efficient models, enhance memory management, optimize collaborative edge–cloud computing, and
advance standardization. Each direction is described with relevant references to recent studies.

Research Direction Description References

Developing ML models that are optimized for energy
Energy-Efficient TinyML Models [59]
efficiency without sacrificing accuracy.
Improving memory management techniques to allow the
Enhanced Memory Management [60]
deployment of more complex ML models.
Exploring and optimizing collaborative computing models
Collaborative Edge–Cloud Models [61]
where edge devices and cloud servers work together.
Contributing to the development of standardized
Standardization Efforts [55]
frameworks and benchmarks for TinyML.
Exploring the use of TinyML in promoting sustainable AI
Sustainable AI with TinyML [47]
practices, minimizing carbon footprints of deployments
Focusing on the deployment of TinyML models in real-time,
TinyML for Real-Time Applications [62]
mission-critical applications like autonomous vehicles.
Creating integrated approaches where HW and SW are
Hardware-Software Co-Design [39]
co-designed to optimize performance in TinyML.

4.2. Conclusions
The advancements in TinyML represent a significant leap forward in the integration of
Machine Learning with resource-constrained devices, enabling intelligent decision making
at the edge of the network. This review has highlighted the transformative potential of
TinyML across various domains, including healthcare, agriculture, industrial automation,
and environmental monitoring. By leveraging optimization techniques such as model
quantization and pruning, TinyML models can operate efficiently on low-power devices,
making them suitable for a wide range of applications. Despite the promising develop-
ments, several challenges remain, including power consumption, memory constraints,
and the need for standardized frameworks. Addressing these challenges requires ongo-
ing research and innovation in both hardware and software. The integration of TinyML
with IoT devices offers significant benefits, such as enhanced real-time data processing
and reduced latency, which are crucial for the effective deployment of intelligent systems.
Future research directions should focus on developing energy-efficient models, improving
memory management techniques, and exploring collaborative edge–cloud computing
models. Additionally, efforts to standardize tools and frameworks will be essential for the
widespread adoption of TinyML. As the field continues to evolve, the potential for TinyML
to revolutionize industries and improve quality of life becomes increasingly apparent.
TinyML stands at the forefront of a new era in Machine Learning, bringing powerful AI
capabilities to the smallest of devices. The continued advancement and adoption of TinyML
hold the promise of a smarter, more efficient, and more connected world.

Author Contributions: Conceptualization, A.E.; methodology, A.E. and P.D.; software, A.E. and Q.Z.;
validation, A.E.; formal analysis, A.E. and S.S.; investigation, A.E. and Q.Z.; resources, A.E., P.D., S.S.
and Q.Z.; data curation, A.E.; writing—original draft preparation, A.E. and P.D.; writing—review and
editing, A.E. and Q.Z.; visualization, A.E. and Q.Z.; supervision, A.E. and S.S.; project administration,
A.E., P.D., S.S. and Q.Z.; funding acquisition, A.E. and S.S. All authors have read and agreed to the
published version of the manuscript.
Funding: This research was partially funded by the “Horizon Europe program under grant agreement
101092850 (AERO project)”; by the “European High-Performance Computing Joint Undertaking (JU)
program under grant agreement 101033975 (EUPEX)”; and by “PNRR project CN1 Big Data, HPC
and Quantum Computing in Spoke 6 multiscale modelling and engineering applications”.
Electronics 2024, 13, 3562 17 of 19

Data Availability Statement: No new data were created or analyzed in this study. Data sharing is
not applicable to this article.
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Shi, W.; Dustdar, S. The Promise of Edge Computing. Computer 2016, 49, 78–81. [CrossRef]
2. Satyanarayanan, M. The Emergence of Edge Computing. Computer 2017, 50, 30–39. [CrossRef]
3. Shi, W.; Cao, J.; Zhang, Q.; Li, Y.; Xu, L. Edge Computing: Vision and Challenges. IEEE Internet Things J. 2016, 3, 637–646.
[CrossRef]
4. Gubbi, J.; Buyya, R.; Marusic, S.; Palaniswami, M. Internet of Things (IoT): A vision, architectural elements, and future directions.
Future Gener. Comput. Syst. 2013, 29, 1645–1660. [CrossRef]
5. Elhanashi, A.; Saponara, S.; Zheng, Q. Classification and Localization of Multi-Type Abnormalities on Chest X-rays Images. IEEE
Access 2023, 11, 83264–83277. [CrossRef]
6. Zheng, Q.; Wang, R.; Tian, X.; Yu, Z.; Wang, H.; Elhanashi, A.; Saponara, S. A real-time transformer discharge pattern recognition
method based on CNN-LSTM driven by few-shot learning. Electr. Power Syst. Res. 2023, 219, 109241. [CrossRef]
7. Elhanashi, A.; Saponara, S.; Dini, P.; Zheng, Q.; Morita, D.; Raytchev, B. An integrated and real-time social distancing, mask
detection, and facial temperature video measurement system for pandemic monitoring. J. Real-Time Image Process. 2023, 20, 95.
[CrossRef]
8. Sze, V.; Chen, Y.-H.; Yang, T.-J.; Emer, J.S. Efficient Processing of Deep Neural Networks: A Tutorial and Survey. Proc. IEEE 2017,
105, 2295–2329. [CrossRef]
9. Zhao, Z.; Zhang, L.; Xu, Y.; Liang, W. Machine learning-based networking: Concepts, applications and challenges. IEEE Netw.
2018, 32, 78–85.
10. Lane, N.D.; Bhattacharya, S.; Mathur, A.; Georgiev, P.; Forlivesi, C.; Kawsar, F. Squeezing Deep Learning into Mobile and
Embedded Devices. IEEE Pervasive Comput. 2015, 14, 82–88. [CrossRef]
11. Han, S.; Mao, H.; Dally, W.J. Deep compression: Compressing deep neural networks with pruning, trained quantization and
Huffman coding. In Proceedings of the International Conference on Learning Representations (ICLR), San Juan, Puerto Rico,
2–4 May 2016.
12. Warden, P.; Situnayake, D. TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers; O’Reilly
Media: Sebastopol, CA, USA, 2019.
13. Ren, H.; Anicic, D.; Runkler, T.A. Tinyol: Tinyml with online-learning on microcontrollers. In Proceedings of the 2021 International
Joint Conference on Neural Networks (IJCNN), Shenzhen, China, 18–22 July 2021.
14. Giordano, M.; Baumann, N.; Crabolu, M.; Fischer, R.; Bellusci, G.; Magno, M. Design and Performance Evaluation of an Ultralow-
Power Smart IoT Device With Embedded TinyML for Asset Activity Monitoring. IEEE Trans. Instrum. Meas. 2022, 71, 1–11.
[CrossRef]
15. Sudharsan, B.; Salerno, S.; Yadav, P.; Breslin, J.G. Approach for Remote, On-Demand loading and Execution of TensorFlow
Lite ML Models on Arduino IoT Boards. In Proceedings of the 2022 21st ACM/IEEE International Conference on Information
Processing in Sensor Networks (IPSN), Milano, Italy, 4–6 May 2022.
16. Google AI Blog. Introducing Edge TPU: Our New Edge Computing Platform for Machine Learning; Google AI Blog: Mountain View,
CA, USA, 2018.
17. Intel Corporation. Intel Movidius Myriad X Vision Processing Unit (VPU); Intel: Santa Clara, CA, USA, 2019.
18. Wu, X.; Wang, M.; Lin, J.; Wang, Z. Amoeba: An Efficient and Flexible FPGA-Based Accelerator for Arbitrary-Kernel CNNs. IEEE
Trans. Very Large Scale Integr. (VLSI) Syst. 2024, 32, 1086–1099. [CrossRef]
19. Zaman, K.S.; Reaz, M.B.I.; Ali, S.H.M.; Bakar, A.A.A.; Chowdhury, M.E.H. Custom hardware architectures for deep learning on
portable devices: A review. IEEE Trans. Neural Netw. Learn. Syst. 2021, 33, 6068–6088. [CrossRef]
20. Jacob, B.; Kligys, S.; Chen, B.; Zhu, M.; Tang, M.; Howard, A.; Adam, H.; Kalenichenko, D. Quantization and Training of Neural
Networks for Efficient Integer-Arithmetic-Only Inference. arXiv 2017, arXiv:1712.05877.
21. Howard, A.G.; Zhu, M.; Chen, B.; Kalenichenko, D.; Wang, W.; Weyand, T.; Andreetto, M.; Adam, H. MobileNets: Efficient
Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017, arXiv:1704.04861.
22. Tan, M.; Le, Q.V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv 2019, arXiv:1905.11946.
23. TensorFlow. TensorFlow Lite for Microcontrollers; TensorFlow: Mountain View, CA, USA, 2020.
24. Edge Impulse. Edge Impulse Documentation; Edge Impulse: San Jose, CA, USA, 2021.
25. Charan Sai, K. An auto-encoder based TinyML approach for real-time anomaly detection. SAE Int. J. Adv. Curr. Pract. Mobil. 2022,
5, 1496–1501. [CrossRef]
26. Diab, M.S.; Rodriguez-Villegas, E. Embedded machine learning using microcontrollers in wearable and ambulatory systems for
health and care applications: A review. IEEE Access 2022, 10, 98450–98474. [CrossRef]
27. Zhang, S.; Li, Y.; Zhang, S.; Shahabi, F.; Xia, S.; Deng, Y.; Alshurafa, N. Deep learning in human activity recognition with wearable
sensors: A review on advances. Sensors 2022, 22, 1476. [CrossRef] [PubMed]
Electronics 2024, 13, 3562 18 of 19

28. Kumar, M.; Zhang, X.; Liu, L.; Wang, Y.; Shi, W. Energy-efficient machine learning on the edges. In Proceedings of the 2020 IEEE
International Parallel and Distributed Processing Symposium Workshops (IPDPSW), New Orleans, LA, USA, 18–22 May 2020.
29. Zhou, X.; Xu, K.; Wang, N.; Jiao, J.; Dong, N.; Han, M.; Xu, H. A secure and privacy-preserving machine learning model sharing
scheme for edge-enabled IoT. IEEE Access 2021, 9, 17256–17265. [CrossRef]
30. Rahman, M.A.; Asyhari, A.T.; Leong, L.S.; Satrya, G.B.; Tao, M.H.; Zolkipli, M.F. Scalable machine learning-based intrusion
detection system for IoT-enabled smart cities. Sustain. Cities Soc. 2020, 61, 102324. [CrossRef]
31. Profentzas, C.; Almgren, M.; Landsiedel, O. Performance of deep neural networks on low-power IoT devices. In Proceedings of
the Workshop on Benchmarking Cyber-Physical Systems and Internet of Things, Virtual, 3 May 2021.
32. Ignatov, A.; Malivenko, G.; Plowman, D.; Shukla, S.; Timofte, R. Fast and accurate single-image depth estimation on mobile
devices, mobile ai 2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, Nashville, TN, USA, 20–25 June 2021.
33. Hymel, S.; Banbury, C.; Situnayake, D.; Elium, A.; Ward, C.; Kelcey, M.; Baaijens, M.; Majchrzycki, M.; Plunkett, J.; Tischler, D.;
et al. Edge impulse: An mlops platform for tiny machine learning. In Proceedings of the Machine Learning and Systems 5, Dalian,
China, 21–23 July 2023.
34. Cartesiam. NanoEdge AI Studio: The Leading Tool for Anomaly Detection. 2021. Available online: https://blog.st.com/nanoedge-
ai-studio/ (accessed on 3 September 2024).
35. Pauzi, A.S.B.; Mohd Nazri, F.B.; Sani, S.; Bataineh, A.M.; Hisyam, M.N.; Jaafar, M.H.; Ab Wahab, M.N.; Mohamed, A.S.A.
Movement estimation using mediapipe blazepose. In Proceedings of the Advances in Visual Informatics: 7th International Visual
Informatics Conference, IVIC 2021, Kajang, Malaysia, 23–25 November 2021; Proceedings 7; Springer International Publishing:
Berlin/Heidelberg, Germany, 2021.
36. Alajlan, N.N.; Ibrahim, D.M. TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI
Applications. Micromachines 2022, 13, 851. [CrossRef] [PubMed]
37. de Prado Escudero, M. On Automation for Optimised and Robust Deployment of Neural Networks on Edge Devices. Ph.D.
Thesis, ETH Zurich, Zurich, Switzerland, 2021.
38. Fedorov, I.; Stamenovic, M.; Jensen, C.; Yang, L.-C.; Mandell, A.; Gan, Y.; Mattina, M.; Whatmough, P.N. TinyLSTMs: Efficient
neural speech enhancement for hearing aids. arXiv 2020, arXiv:2005.11138.
39. Kwon, J.; Park, D. Hardware/Software Co-Design for TinyML Voice-Recognition Application on Resource Frugal Edge Devices.
Appl. Sci. 2021, 11, 11073. [CrossRef]
40. Paul, A.J.; Mohan, P.; Sehgal, S. Rethinking Generalization in American Sign Language Prediction for Edge Devices with Extremely
Low Memory Footprint. In Proceedings of the 2020 IEEE Recent Advances in Intelligent Computational Systems, RAICS 2020,
Thiruvananthapuram, India, 3–5 December 2020; pp. 147–152.
41. Mohan, P.; Paul, A.J.; Chirania, A. A tiny cnn architecture for medical face mask detection for resource-constrained endpoints. In
Advances in Intelligent Systems and Computing; Springer: Berlin/Heidelberg, Germany, 2021; Volume 657, pp. 657–670.
42. Dutta, D.L.; Bharali, S. TinyML Meets IoT: A Comprehensive Survey. Internet Things 2021, 16, 100461. [CrossRef]
43. Guleria, C.; Das, K.; Sahu, A. A survey on mobile edge computing: Efficient energy management system. In Proceedings of the
2021 Innovations in Energy Management and Renewable Resources (52042), Kolkata, India, 5–7 February 2021; pp. 1–4.
44. de Prado, M.; Rusci, M.; Capotondi, A.; Donze, R.; Benini, L.; Pazos, N. Robustifying the deployment of tinyML models for
autonomous mini-vehicles. Sensors 2021, 21, 1339. [CrossRef]
45. González Rivero, R.A.; Morera Hernández, L.E.; Schalm, O.; Hernández Rodríguez, E.; Alejo Sánchez, D.; Morales Pérez, M.C.;
Nuñez Caraballo, V.; Jacobs, W.; Martinez Laguardia, A. A low-cost calibration method for temperature, relative humidity, and
carbon dioxide sensors used in air quality monitoring systems. Atmosphere 2023, 14, 191. [CrossRef]
46. Ometov, A.; Shubina, V.; Klus, L.; Skibińska, J.; Saafi, S.; Pascacio, P.; Flueratoru, L.; Gaibor, D.Q.; Chukhno, N.; Chukhno, O.; et al.
A survey on wearable technology: History, state-of-the-art and current challenges. Comput. Netw. 2021, 193, 108074. [CrossRef]
47. Abadade, Y.; Temouden, A.; Bamoumen, H.; Benamar, N.; Chtouki, Y.; Hafid, A.S. A Comprehensive Survey on TinyML. IEEE
Access 2023, 11, 96892–96922. [CrossRef]
48. Ahmed, A.A.; Echi, M. Hawk-Eye: An AI-Powered Threat Detector for Intelligent Surveillance Cameras. IEEE Access 2021, 9,
63283–63293. [CrossRef]
49. Loizzi, G. Development of a Wearable Device for Breathing Rate Monitoring Using Stretchable Sensors. Ph.D. Thesis, Politecnico
di Torino, Turin, Italy, 2021.
50. Kurniawan, A. IoT Projects with Arduino Nano 33 BLE Sense; Apress: Berkeley, CA, USA, 2021; Volume 129.
51. Reidy, B.C.; Mohammadi, M.; Elbtity, M.E.; Zand, R. Efficient deployment of transformer models on edge tpu accelerators: A
real system evaluation. In Architecture and System Support for Transformer Models (ASSYST@ ISCA 2023); ISCA 2023 Workshop:
Orlando, FL, USA, 2023.
52. Jolles, J.W. Broad-scale applications of the Raspberry Pi: A review and guide for biologists. Methods Ecol. Evol. 2021, 12, 1562–1579.
[CrossRef]
53. Goudarzi, M.; Palaniswami, M.S.; Buyya, R. A Distributed Deep Reinforcement Learning Technique for Application Placement in
Edge and Fog Computing Environments. IEEE Trans. Mob. Comput. 2021, 22, 2491–2505. [CrossRef]
54. Muhammad, G.; Hossain, M.S. Emotion Recognition for Cognitive Edge Computing Using Deep Learning. IEEE Internet Things J.
2021, 8, 16894–16901. [CrossRef]
Electronics 2024, 13, 3562 19 of 19

55. Shafique, M.; Theocharides, T.; Reddy, V.J.; Murmann, B. TinyML: Current Progress, Research Challenges, and Future Roadmap.
In Proceedings of the 58th ACM/IEEE Design Automation Conference (DAC), San Francisco, CA, USA, 5–9 December 2021;
pp. 1303–1306.
56. Xu, K.; Zhang, H.; Li, Y.; Zhang, Y.; Lai, R.; Liu, Y. An Ultra-Low Power TinyML System for Real-Time Visual Processing at Edge.
IEEE Trans. Circuits Syst. II Express Briefs 2023, 70, 2640–2644. [CrossRef]
57. Bi, M.; Wang, Y.; Cai, Z.; Tong, X. A privacy-preserving mechanism based on local differential privacy in edge computing. China
Commun. 2020, 17, 50–65. [CrossRef]
58. Belcastro, L.; Marozzo, F.; Orsino, A.; Talia, D.; Trunfio, P. Edge-Cloud Continuum Solutions for Urban Mobility Prediction and
Planning. IEEE Access 2023, 11, 38864–38874. [CrossRef]
59. Immonen, R.; Hämäläinen, T. Tiny Machine Learning for Resource-Constrained Microcontrollers. J. Sens. 2022, 2022, 7437023.
[CrossRef]
60. Han, H.; Siebert, J. TinyML: A Systematic Review and Synthesis of Existing Research. In Proceedings of the International Conference on
Artificial Intelligence in Information and Communication (ICAIIC), Jeju Island, Republic of Korea, 21–24 February 2022; pp. 269–274.
61. Bao, G.; Guo, P. Federated learning in cloud-edge collaborative architecture: Key technologies, applications and challenges.
J. Cloud Comput. 2022, 11, 94. [CrossRef]
62. Koufos, K.; EI Haloui, K.; Dianati, M.; Higgins, M.; Elmirghani, J.; Imran, M.A.; Tafazolli, R. Trends in intelligent communication
systems: Review of standards, major research projects, and identification of research gaps. J. Sens. Actuator Netw. 2021, 10, 60.
[CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.