Machine Translated by Google

Received 6 June 2023, accepted 7 July 2023, date of publication 10 July 2023, date of current version 12 September 2023.
Digital Object Identifier 10.1109/ ACCESS.2023.3294111

A Comprehensive Survey on TinyML

YOUSSEF ABADADE1, ANAS TEMOUDEN 2 , HATIM BAMOUMEN2, (Student Member, IEEE),
4
NABIL BENAMAR 2,3, YOUSRA CHTOUKI2, AND ABDELHAKIM SENHAJI HAFID
1System Engineering Laboratory, National School of Applied Sciences, Ibn Tofail University, Kenitra 14000, Morocco
2School of Sciences and Engineering, Al Akhawayn University, Ifrane 53000, Morocco
3Department of Computer Engineering, School of Technology, Moulay Ismail University, Meknes 50050, Morocco
4University of Montreal, Montreal, QC H3C 3J7, Canada

Corresponding author: Nabil Benamar ([email protected])

ABSTRACT Recent spectacular progress in computational technologies has led to an unprecedented boom in
the field of Artificial Intelligence (AI). AI is now used in a plethora of research areas and has demonstrated its
capability to bring new approaches and solutions to various research problems. However, the extensive
computation required to train AI algorithms comes with a cost. Driven by the need to reduce the energy
consumption, the carbon footprint and the cost of computers running machine learning algorithms, TinyML is
nowadays considered as a promising AI alternative focusing on technologies and applications for extremely
low-profile devices. This paper presents the results of a literature survey of all TinyML applications and related
research efforts. Our survey builds a taxonomy of TinyML techniques that have been used so far to bring new
solutions to various domains, such as healthcare, smart farming, environment, and anomaly detection. Finally,
this survey highlights the remaining challenges and points out possible future research directions. We anticipate
that this survey will motivate further discussions on the various fields of applications of TinyML and the synergy
of resource-constrained devices and edge intelligence.

INDEX TERMS TinyML, embedded machine learning, deep learning, edge intelligence.

I. INTRODUCTION models become larger and larger so does their carbon foot-
Recent advances in computational technologies have enabled an unprecedented print. To overcome the side effect of AI on the ecosystem and
boom in the field of Artificial Intelligence (AI). AI has become ubiquitous in our to make it more affordable for researchers and practitioners,
daily life. Ranging from game playing [1] to healthcare monitoring [2], [3] and tiny machine learning (TinyML) has recently emerged as a
passing by Natural Language Processing (NLP) [4], com-puter vision (CV) [5], promising field of AI. Driven by the need to reduce the energy
social media [6], and Autonomous Driving [7], [8], more AI applications consumption and the CO2 emissions of computers running
continue to expand as more people embrace this technology. However, the machine learning algorithms, TinyML is an AI alternative that
extensive computation needed to reach such exciting results in AI-based exclusively uses extremely low-profile devices to process AI
research and projects comes with a cost. Financially, the price of dedicated algorithms [10].
hardware to run AI algorithms is increasing. TinyML is a big shift in AI. It pushes the intelligence to the
edge and makes use of tiny devices such as microcontollers
to execute AI algorithms. TinyML powers low-latency,
The cloud alternative does not always overcome the cost issue consumes low power, and uses low bandwidth. In modern
since the cloud compute time is proportional to the number of computing technologies, standard CPUs consume gener-ally
processes being executed [9]. This also impacts the up to 85 watts and standard GPUs consume between 200
environment, as a result of non-renewable energy sup-plied watts to 500 watts [11], [12]. A typical microcon-troller
to modern tensor processing hardware. Authors in [9] evaluate consumes thousand times less power, in the order of milliwatts
the estimated cost of training some benchmark Deep Learning or microwatts. Such low energy consumption of TinyML
(DL) algorithms, in terms of CO2 emissions and cloud compute devices enables running them unplugged for weeks
time. The results show that as the trained or even years.
In AI, training large ML models requires a significant amount
The associate editor coordinates the review of this manuscript and of data that should be processed in the cloud net-work. AI
approving it for publication was Gianluigi Ciocca . researchers are nowadays more concerned about
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
96892 VOLUME 11, 2023
Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

the ethical aspect of machine learning than the accuracy of the model [13]. 1) We provide a new classification of TinyML application-tions and
This has led to extensive works on privacy-preserving machine learning techniques by surveying all published papers on TinyML until 2023.
techniques for big data analysis [14]. There is a growing interest from the
ML com-munity in leveraging a plethora of cryptographic techniques to 2) We study in depth the main advantages of using TinyML compared
secure data both in the training phase and the testing phase. to other existing approaches 3) We highlight how TinyML
unlocks new advances for sustainable development.
However, by deploying TinyML where the source of data is created and
since this data does not leave the device, privacy concerns are largely 4) We highlight the remaining challenges in TinyML towards its worldwide
addressed. deployment in different research fields.
There are different promising domains where TinyML can bring a
significant impact. In industry, anomaly detection is highly critical and can
The rest of this paper is organized as follows. Section I dis-cusses
help reduce the downtime for repairs and thus increase the efficiency of
existing surveys, highlights their contributions, and the main gap we fill
the production process.
with the current survey. It also presents the scope of this survey and the
By deploying ML algorithms in the edge, one can detect and analyze
research methodology. Section II presents an overview of TinyML. Section
continuously the sound emitted by the machine during the production
III presents envi-ronmental TinyML applications. Section IV focuses on
process, which can inform of a possible break down in that device.
recent TinyML applications in healthcare. Section V covers recent advances
Analyzing different metrics such as sounds or vibrations in real time will
of TinyML in smart farming. Section VI presents TinyML use for anomaly
help save time in correcting or replacing any defective device without
detection. Section VI presents the remaining challenges in TinyML, open
unnec-essary delays. In environment [15], one of the recent research
research questions, and future research directions. Finally, Section VII
fields, where sensors have been extensively deployed, is the Internet of
concludes the paper. For ease of use, the acronyms are summarized in
animals [16], [17]. Understanding animal behavior continues to be a
Table 5.
grueling task for a majority of researchers.

It is indeed not an easy routine to track animals for hours or even days in
their living place to document their behavior.
Internet of Things (IoT) and especially TinyML can make this arduous work A. EXISTING SURVEYS
largely superfluous. It can help getting more detailed insights on animals In this section, we explore existing survey specific publica-tions to TinyML.
life and predict possible threats. As of the writing of this paper, there are a few surveys exploring TinyML.
In the elephant TinyML project [18], collars are attached to an elephant This is mainly due to the fact that TinyML is relatively a new research topic.
and capture its real-time movements using GPS. According to [19], the year 2019 was the year where TinyML was dis-
The embedded sensors take surrounding images that are con-tinuously cussed for the first time in research publications. In Table 1, we order the
processed and analyzed by TinyML and can predict events around each existing surveys by publication year. The table shows a comparison of
animal such as the presence of human predators. Other ML models can existing surveys in terms of the topics covered: Benefits of TinyML, model
also be applied to understand and detect the mood of the elephant, while compression techniques for TinyML, TinyML frameworks and inference
an accelerometer is used to further determine the movement of the engine, TinyML Hardware, taxonomy of the main TinyML application-tions,
elephant. TinyML in the environment, smart farming, anomaly detection, healthcare
TinyML brings new insights and unlocks new possibilities for sustainable and challenges. The table lists the depth in which each topic was
development. By reducing the latency, TinyML enables real time addressed (covered, partially cov-ered). It also includes the number of
applications to be deployed in the source of data, such as in the case of references and their year of publication, cited in each survey paper. We
image and speech recognition. can categorize the main topics surveyed as follows: surveys that discuss
TinyML models can also run even when there is no internet connection, how TinyML can optimize ML models to bring intelligence and autonomy
which can not be realized in the cloud context. to devices in specific fields such as, healthcare [20] and embedded vision
In TinyML, the processed data do not need to leave the device, which [21]. Those that focus on general aspects of TinyML implementations.
significantly improves user privacy and thus complies with data protection These include, 1) the benefits, 2) use cases in TinyML, 3) frameworks and
regulations. input engine, 4) hardware, 5) model compression techniques, 6) tools, and
The scope of this paper is to present TinyML as a major candidate for challenges and future roadmap [22], [23], [24], [25], [26]. A survey
the support of machine learning in small and constrained devices such as presented the challenges and directions of benchmarking in TinyML [27].
microcontrollers, by decom-posing various related technologies and The work in [21] explores existing TinyML engineering, workflow and its
revealing research challenges and directions. This paper also highlights challenges specific to IoT embedded vision. In [28], the authors covered
the power that TinyML brings to the research field of AI in general and reformable TinyML, and listed existing workflows, deploy-ments schemes
Deep learning in particular. The current study shows to the AI researchers and sectors affected by reformable TinyML.
worldwide, who may have limited access to the high technology of
dedicated servers and data-centers, that TinyML is an affordable alternative.
specifically, the contributions of this paper can be summarized as follows:

VOLUME 11, 2023 96893

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 1. Organization of the survey.

The study in [29] presents a clear and complete closed-loop same remaining challenges including limitation of existing
workflow for the development and deployment of ML models benchmarks, hardware limitations when it comes to large
for MCUs. It outlines each step in the workflow and provides amount of data, software and hardware co-design which is
both qualitative and numerical insights. specifically addressed in [25]. However, existing surveys
All of these surveys have reported the same promised have not focused on the sectors where TinyML has a major
ben-efits of TinyML including energy efficiency, low cost, impact, such as healthcare, environment, smart farming, and
data integrity, privacy, security, and latency. They also reported the
anomaly detection. Most of the surveys demonstrated

96894 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

the implementation of TinyML in different use cases, such as gesture

recognition, speech recognition, keyword spot-ting, image recognition, and
visual wake words. Our survey attempts to address these limitations which
we discuss in detail in the following section.

B. THE CURRENT SURVEY

Our survey defines TinyML and presents its potential benefits in different
sectors. In addition, we report the current advances and challenges in
applying tinyML. We explore in depth existing contributions to critical
sectors where the use of TinyML brings a significant impact. To the best of
our knowledge, this survey is the first to include a taxonomy of the main
FIGURE 2. Types of used references.
TinyML applications. The main sectors that we survey are, the environment,
smart farming, healthcare and anomaly detection. The choice of these
sectors was driven by the amount of existing publications in these fields.
The potential of TinyML revolutionizing these specific sectors as well as this connection to any other device or thing [34]. Today, IoT is a huge

others can be sensed by the detailed overview of the existing work that we network that connects various devices including computers, smartphones,

share in this survey. We believe this survey is an important added value to drones to wearable and autonomous vehicles. IoT devices are getting more

the research field as well as a valuable reference to researchers. complex as machine learning models are integrated, allowing them to be
more intelligent. Because of the vast quantity of data that can be acquired
from IoT components, the coupling of machine learning with IoT leads to
overall development and better intelligence of IoT devices. The IoT
architecture can be broken down into four fundamental layers [35]: 1) the

C. REVIEW PROCEDURE perception layer: it is composed of sensors that gather data and physical
measurements, and actuators that execute tasks or actions based on
The research procedure used in this survey is a System-atic Literature
sensor data; 2) the network or transport layer: it comprises the infrastructure
Review (SLR) [30], which consists of the following steps: defining the
for internet gateways and data acquisition systems, to transmit and gather
research questions, retrieving the literature, evaluating the literature,
data from dif-ferent devices to an on-premise location; 3) the middleware or
extracting the data, and finally presenting and discussing the results. Our
processing layer; it includes high-performance machines for data analysis
aim is to present a fair evaluation of existing works on TinyML.
and data storage; and 4) the application or service/interface layer; it grants
users access to services and presents them through interfaces or APIs.
To this aim, we use different digital libraries: IEEE Xplore Digital Library,
However, when implementing tasks such as machine learning models, this
Science Direct, and Springer Link. We also used Google Scholar for
architecture may encounter difficulties such as high energy consumption. It
bibliographic databases. Differ-ent words and acronyms have been used
might also need to maintain a steady connection at the network layer in
in our research, such as : TinyML, Tiny-ML, embedded, tiny machine
order to preserve communication with devices and the cloud.
learning, embedded machine learning, edge intelligence, and TinyDL.

We noted the existence of some research papers dealing with TinyML as

the main topic without mentioning this term in the title neither in the
keywords [31], [32], [33].
The scope of our work in the current survey was guided by
The IoT architecture may be separated into three tiers based on
answering the three following research questions:
computing capacity, according to [36]. This hierarchical separation
1) What are the new AI related publications that use non traditional
increases efficiency and dependency among layers while also allowing for
computing resources and use mainly tiny devices?
cost and resource optimization [37]. Cloud computing in the top layer is
intended for high-level data processing and storage. However, the distance
2) What are the smart solutions proposed by TinyML developers and
between cloud servers and the data source causes challenges such as
researchers that mitigate the limitations of existing AI based
sluggish decision-making and latency. In 2011, Fog computing was
approaches?
introduced to overcome the increasing number of IoT devices, as a
3) What are the trending topics in TinyML, remaining challenges, and
decentralized computer layer between the cloud and the sensing
future research directions?
components [38].

II. BACKGROUND This layer performs data collection, medium processing, and decision-
A. IoT ARCHITECHURE making, decreasing the quantity of information coming to the cloud layer.
The term IoT has been introduced for the first time back in 1999 by Kevin Fog computing permits to cope with the huge amount of data to be
Ashton at MIT's Auto-ID Center to refer to a network that not only connects processed in the cloud by performing task offloading. The lowest layer is
computers but extends edge

VOLUME 11, 2023 96895

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 1. Comparison of related survey papers. Annotations: ''ÿ'' indicates that the topic is covered, ''ÿ'' indicates that topic is partially covered, ''ÿ''
indicates that the topic is not covered.

computing, which puts data storage and computation closer to the source have expanded, along with the requirements for low-latency and real-time
of the data (components of the sensing layer) [39]. decision making [23]. Mobile Machine Learn-ing is one example of the use
The fog and edge layers increase bandwidth and reduce latency, allowing of intelligence on the edge layer, especially on mobile devices such as
autonomous decision-making. These two layers improve the security and smartphones and tablets.
the privacy of the user data since it does not need to be sent outside the Mobile ML techniques are meant to perform effectively on mobile devices'
local network for further processing and storage. The edge layer is more limited computing resources (few gigabytes of RAM memory) and battery
dependent than the fog layer since the core tasks are handled locally, restrictions. A Neural Process-ing Unit (NPU) was incorporated in current
lowering device and network costs while providing quicker reaction time mobiles to enable the execution of ML algorithms. It is a customized
and offline availability. The computing paradigm is currently shifting from procedure designed to speed neural network workload calculation.
cloud computing to end-edge-cloud computing, which also supports AI
evolving from a centralized AI deployment to a distributed artistic intelligence NPU is simpler than traditional processors such as CPU or GPU [42] in
(DAI). This new paradigm is empowered by end-edge-cloud computing performing matrix multiplications, which are the essential operations of
(EECC), where the heterogeneous computing capabilities of on-devices, neural networks.
edge, and cloud servers are managed to meet the requirements raised by
resource-intensive and distributed AI computation [40]. TinyML focuses on deploying compressed and optimized machine
learning models on tiny, low-power devices such as battery-powered
microcontrollers, and embedded systems.
TinyML was inspired by Mobile ML's features [43] (low latency, resource
limits, moderate cost) and its development grew as a result of the technical
B. TinyML breakthrough in the field of IoT and MCUs. TinyML is successfully applied
Artificial intelligence is one of the sectors touched by the new IoT paradigm, in vari-ous application areas(see Fig. 3) eg, healthcare, agriculture,
which advocates bringing intelligence to the edge layer. Edge AI evolved industrial IoT and environment. TinyML technology is driven by the
as a response to the limits of cloud-based AI, which is not necessarily necessity to integrate intelligence in a wide range of applications that were
appropriate for real-time applications and devices with limited processing previously not viable owing to the high power and computing needs of
power and bandwidth [41]. The term has become more popular in recent standard ML models as shown in Fig. 4:
years as IoT and the number of linked devices

96896 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 3. Taxonomy of the main TinyML applications.

1) REDUCING LATENCY This is important for applications that need to be deployed in

TinyML models can run on the device itself; thus the response areas with poor or no Internet access.
time is much faster than sending the data to the cloud for pro-
cessing. This is critical for applications that require real-time 3) IMPROVING PRIVACY AND
decisions, such as image and speech recognition. Figure 5 SECURITY TinyML keeps the data on the device; thus,
illustrates a comparison between the latency and accuracy of sensitive information does not have to be sent to the cloud
a ResNet-based model for Image Classification deployed on for processing. This feature drastically changes the AI ethics
two TinyML-supported boards (NDP9120-EVL Board and picture by ensuring the user data privacy, which complies
GAP9 EVK Board), a traditional cloud-based deploy-ment on with data protection regulations.
Google Cloud, and Amazon servers. To conduct the benchmark
test, MLPerf [44] and Stanford DAWN [45] were utilized. The
4) LOW ENERGY CONSUMPTION
results presented in Fig. 5 demonstrate that the deployment
One significant part of decreasing energy usage in TinyML is
of models on TinyML systems significantly reduces latency,
to reduce the quantity of data that has to be transported and
with a range of 0 to 5 ms, as compared to cloud-based
processed. Furthermore, TinyML algorithms are often intended
machine learning models. Furthermore, TinyML systems offer
to be computationally efficient, which can help in lowering the
high accuracy in their models, with only a slight decrease from
power consumption of the device on which they are executed.
95% to 85% due to compression and optimization for use on
Other methods for reducing energy usage in TinyML systems
restricted devices. These findings highlight the potential of
include employing low-power components, designing for low
TinyML to minimize latency while maintaining model accuracy,
voltage and battery-powered operations, and adopting sleep
making it an attractive option for deploying machine learning
or idle modes.
models on resource-constrained devices.

5) REDUCING
2) OFFLINE CAPABILITY COST TinyML models can save on costs associated with
TinyML models can run even when there is no Internet sending data to the cloud for processing and storage, such
connection, whereas cloud ML models require such connectivity. as band-width and storage costs. The low energy consumption is also

VOLUME 11, 2023 96897

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 4. TinyML benefits.

extraction and processing of multi-sensor data. Although fog and edge

computing provide low latency, the major drawback is that the wireless
connection between the IoT device and the gateway or fog node must
remain active for complex operations.

Table 2 compares TinyML with cloud, fog, and edge com-puting in

terms of latency, data privacy, accuracy, power consumption, reliability,
and memory consumption. The val-ues presented in the table are derived
from the benchmark conducted by MLCommons [44], an open engineering
con-sortium dedicated to advancing machine learning innovation and
fostering the development of cutting-edge ML models and systems.
Benchmarking plays a crucial role in allowing ML enthusiasts to measure,
compare, evaluate, and poten-tially enhance learning systems. It is
FIGURE 5. Latency and accuracy comparison. important to note that the TinyML benchmark is still in the early stages of
development, with the latest version being 1.0 [44]. The table shows that
TinyML has advantages over other technologies in terms of latency, privacy,
power consumption, memory consumption and reliability. However, it
an additional factor that implies the reduction of costs of TinyML.
seems to have lower accuracy compared to these technologies due to the
model optimization for deployment on low-constraint devices.
TinyML enables IoT devices to become intelligent by allowing them to
perform data analysis at the edge, leading to faster decision-making.
Additionally, it can be applied to a variety of use cases to provide standalone
ML services. Figure 6 describes the integration of TinyML in IoT workflow.
In the IoT workflow, TinyML refers to the use of ML models that are small
and resource-efficient enough to run on devices with limited computational C. HARDWARE CONSTRAINTS
capabilities, such as MCU found in IoT devices. These models are trained The growth of the Internet of Things industry has been fueled by
to perform specific tasks, such as image recognition, audio classification, advancements in artificial intelligence and machine learn-ing, as well as
or sensor data analysis. The data is collected by sensors on the IoT device improved connectivity. The growing demand for portable devices and the
and passed through the TinyML model for processing. The output of the increasing use of microcon-trollers for edge computing drive the
model is then used for various purposes such as controlling the device, microcontroller Market.
sending data to the edge, fog node or cloud for further analysis. According to verified market research [46], the microcon-troller market
size was valued at USD 348.83 million in 2021 and is expected to reach
USD 992.80 million by 2030.
A microcontroller is a compact integrated circuit designed to control a
There are various techniques proposed in the literature that address the specific function in an embedded system and includes a processor ranging
challenges associated with cloud-based systems, including alternative from small and simple 4-bit processors to complex 32-bit, memory, input/
mechanisms for data analysis such as fog and edge computing. These output (I/O) peripherals, potentially with hardware accelerators, sensors,
technologies have been used to extend cloud services to the local area and connectivity. MCUs are generally small, low-cost and energy-saving.
network, located where data processing occurs at an IoT gateway or a fog
node close to the extreme edge device. Typically, each IoT device is
registered onto a gateway or fog node to enable the The market has seen an emergence of the new 32bits generation of
IoT-ready microcontrollers [46]. This emerged has

96898 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 2. Comparison of TinyML performances against existing technologies.

dustically transformed edge computing. Because of support for single time, and activation sparsity. The GrAI One accelerator pro-cesses network
instruction multiple data (SIMD) and digital sig-nal processing (DSP) events on chip and saves weights in local SRAM. The neurons carry out
instructions, Cortex-M-based devices may now perform previously basic neural and arithmetic operations, and their states are stored in local
unachievable tasks. MCUs also include on-chip SRAM and embedded SRAM. Based on a synapse table, the platform generates events and
Flash; thus models that can fit within the memory limits are free from the neural model values. The sparsity-aware technique is similar to TinyML
costly DRAM accesses that limit classical ML. The broad acceptance and models and has the potential to save energy in neural network calculations
implementation of TinyML relies on the capabilities of these platforms. [23].
While general-purpose MCUs provide flexibility, specialized hardware
provides the best TinyML performance efficiency. These customized Additionally, another technique for optimizing machine learning
devices can achieve performance in the one micro Joule per inference workloads is based on the usage of Tensor Process-ing Units (TPU's). A
level, pushing the limits of machine learning to the ultra-low power end of TPU is a custom-built AI accelerator designed by Google to perform highly
TinyML processors [27]. efficient matrix calculations: the foundation of many machine learning
algorithms.
TPUs may greatly accelerate ML model training and infer-ence, reducing
Deep learning hardware accelerators are specialized chips or circuits the time and expense necessary to create correct models. Edge TPU is a
designed to improve the performance and efficiency of neural networks. form of TPU developed by Google intended exclusively to run TinyML
These hardware accelerators are crucial for deploying deep learning models models with great performance and minimal power consumption. The Edge
on TinyML as they provide parallel processing capabilities and optimized TPU is an ASIC (Application-Specific Integrated Cir-cuit) that can conduct
data flow to reduce computation time, energy consumption, and memory real-time inference in conjunction with edge devices such as the Raspberry
usage. Authors in [47] present a convolu-tional neural network accelerator Pi and Coral Dev Board.
(HBDCA) that uses a high-accuracy block random access memory (BRAM)-
aware FPGA structure. For quantization, the HBDCA incorporates The Edge TPU is a low-cost solution for running TinyML models on devices
TensorFlow Lite (TFL) with 8-bit per-layer activation. The toolchain enables at the network's edge, allowing these devices to do complicated ML tasks
BRAM content to be updated without the need for re-synthesis or re- with minimal energy consuption, and excellent accuracy that is significantly
implementation. The HBDCA supports multiple kernel-level parallelism and better than traditional CPUs, according to the [48]. The edge TPU has
uses spa-tial and temporal mechanisms to minimize memory access. demonstrated the capacity to run cutting-edge mobile vision models such
as mobilenet v2 at around 400 FPS while being power efficient [49].

Keras TensorFlow Lite is used to train the toolchain, and FPGA resources Table 3 compares various TinyML devices, in terms of processor, CPU
are used to find the best hardware con-figuration. The toolchain's final clock frequency, flash memory, SRAM size, power consumption or voltage,
output is a memory map information file, which is generated at the end of connectivity, sensors, product developer, and price. The majority of
the flow. hardware boards use the ARM Cortex processor with CPU clock frequencies
The toolchain is designed for TinyML environments and provides a high- ranging from 100 MHz to 480 MHz. Most of the boards have WiFi and
accuracy workflow that does not require re-implementation. The GrAI One
accelerator which is a platform for sparsity-aware computing in neural Bluetooth connectivity, as well as various on-board sensors such as light
networks that attempts to decrease computation overload and improve sensor, air pressure sensor, microphone, tem-perature sensor, humidity
energy usage. The platform focuses on connection, space, sensor, gyroscope, gesture sensor, accelerometer, air quality sensor and
camera. Some TinyML

VOLUME 11, 2023 96899

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 3. Hardware platforms to support TinyML.

devices also incorporate hardware accelerators, such as the models may be operated on devices with limited process-
Coral Edge TPU ML. ing power and memory. This enables the incorporation of
machine learning capabilities into a wide range of devices,
D. SOFTWARE OPTIMIZATION resulting in new applications and better end-user experiences.
TinyML relies mostly on software, which enables the deploy-
ment of machine learning models on resource-constrained Model compression is a strategy for reducing a machine
hardware. Because software allows for the optimization of learning model's size and processing needs. This method
model size and computational needs, machine learning can result in a 20% to 30% decrease in memory space required

96900 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 6. TinyML In IoT WorkFlow.

for network parameter storing. Several ways to compress a model exist: exactly rebuilt to its original form with no information loss.
Pruning: The Huffman coding works by assigning binary codes to each symbol in the
pruning method begins with training the network and then selecting the data set, with shorter codes for frequently appearing symbols and longer
key links by locating the weights that are greater than a specific threshold. codes for less frequently occurring symbols. The rationale behind this
Weights below this level are subsequently eliminated, resulting in a trimmed method is that symbols that appear more often in the data will take up less
model. This trimmed model may not provide the same accuracy as the space if they are represented by shorter codes.
dense network, however retraining the residual weights can restore the
accuracy. Pruning also aids in the removal of connections and neurons that Knowledge Distillation: In addition to the Model com-pression
have no input or output connections. techniques, Knowledge distillation technique is another impor-tant used in
TinyML. Knowledge distillation is a machine learning approach in which a
smaller, more compact model (referred to as the student model) is trained
Quantization: Quantization is used to reduce the preci-sion of the to mimic the outputs of a larger, more accurate model (known as the teacher
weights and activation is reduced from 32-bit or 64-bit floating-point numbers model). The student model's goal is to approximate the teacher model's
to 8-bit or lower fixed-point numbers. As fixed-point arithmetic is quicker predictions as closely as possible, and the idea behind knowledge distillation
and more energy-efficient than floating-point arithmetic, this reduces the is that the smaller model can learn useful information about the problem
model's memory footprint as well as the amount of pro-cessing required. from the teacher model, even if the student model is not as complex or
Quantization can be done either during or after model training. The accurate as the teacher model. The knowledge distillation training procedure
objective is to strike a compromise between model accuracy and the consists of two steps: (1) train the teacher model on the original training
precision of the weights and activations, as reducing precision too much data, and (2) train the student model with the teacher model's predictions
might result in severe accuracy loss. as the target. The student model is trained using a loss function that takes
into account both the accuracy of its own predictions and the similarity of its
predictions to those of the instructor model.
Low-Rank Factorization: Low-rank factorization is a mathematical
approach for approximating a high-dimensional matrix with a low-dimensional
one while maintaining as much information as possible from the original
matrix. The purpose of low-rank factorization in Machine Learning is to
decompose a dense weight matrix into the product of two lower-dimensional However, the traditional techniques for compressing TinyML models
matrices with lower ranks, thus reducing the matrix's dimensionality while mentioned above can lead to a significant loss of accuracy due to poor
keeping its structure and significant properties. As a result, the model is matrix characteristics result-ing from high compression rates [50]. This
represented more compactly, with fewer parameters and is more compu- prompted the development of Tiny neural networks, which are compact
tively efficient. neural network models with a limited number of parameters designed to
function effectively on embedded devices with low processing resources.
When compared to bigger, more complicated models, tiny neural networks
Huffman Coding: Huffman coding is a lossless compres-sion method, are often trained
which means that compressed data can be

VOLUME 11, 2023 96901

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 4. Tiny neural networks.

on a fraction of a larger dataset and employ a simpler design. strategies, including reinforcement learning, evolutionary
This allows them to be trained and deployed more quickly algorithms, and gradient-based methods. One example of a
and efficiently while retaining a high degree of precision. The NAS framework for TinyML is DNAS. It stands for Differ-entiable
table below 4 illustrates the several Tiny Neural Networks that Neural Architecture Search, which is a type of neural
exist, as well as applications, the number of parameters, and architectural search method for tiny machine learning. DNAS
the size if available. approaches different from traditional NAS methods by using
The technique of autonomously searching for the opti-mum neural gradient-based optimization techniques to search for the best
network design for a particular job is referred network architecture. The idea is to train a network architec-
as big as NAS. NAS can be used in the context of TinyML to ture generator that outputs a candidate architecture, and then
discover compact and efficient neural network architectures use gradient descent to optimize the architecture generator
that can run on small embedded devices with low com- puting resources. parameters such that the accuracy of the resulting architec-ture is
NAS for TinyML seeks neural network improved. [60] Another framework is FPGA-aware
designs that balance accuracy, model size, and computa-tional graph neural architecture search (FGNAS). FGNAS performs
performance in order to satisfy the restrictions of a search for the best GNN architecture while considering
real-time deployment on small embedded devices. NAS algo-rithms hardware constraints. The framework is evaluated on bench-mark
may find structures utilizing a variety of optimization datasets such as Cora, CiteCeer, and PubMed, and the

96902 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

results show that FGNAS has better capability in optimizing the accuracy the need for standard C or C++ libraries or dynamic memory allocation [64].
of GNNs when their hardware implementation is specifically constrained.
[61] Edge Impulse [65] is a cloud-based solution that facilitates the creation
The deployment of models to embedded devices cannot currently and deployment of machine learning models for TinyML devices, from
support model training due to limited resources. collecting data using IoT devices, to extracting features, training models,
Typically, models are trained on the cloud or on a more capa-ble device and finally deploy-ing and optimizing the models for TinyML devices. The
before being distributed to the embedded device. trained models can run on various edge devices, like micro-controllers,
However, applying machine learning algorithms on embedded-ded devices single-board computers, and embedded systems.
creates difficulties, such as limited computing power. There are three
methods for deploying models [62]: hand coding, code generation, and ML It employs the EON compiler for model deployment and also supports
interpreters. Hand coding provides for low-level optimizations but takes TFLM. Edge Impulse utilizes TensorFlow's Model Optimization Toolkit to
time; code generation creates optimized code but has portability difficulties. quantify models, lowering the precision of their weights from float32 to int8
An ML interpreter is a tool used to implement machine learning algorithms with minimal impact on accuracy [66]. The EON compiler [67] compiles the
on embedded devices with limited processing capabilities, including MCUs. neural network model directly into C++ source code, reducing the amount
It is part of a framework that includes tools and software libraries and calls of stored ML operators that are not in use. It has been demonstrated that
individual kernels. the EON compiler can run the same network with 25% to 55% less SRAM
and 35% less flash compared to TFLM [29]. Edge Impulse has also
designed a novel machine learning algorithm named FOMO [68]. It is a ML
Aside from the interpreter, a TinyML framework typically includes TinyML model designed for highly resource-constrained devices. It enables object
libraries and tools for data processing, as well as a Tiny Inference Engine, detection to count objects, determine their location within an image, and
which is a low-level designed soft-ware library or hardware accelerator to track multiple objects in real-time using far less process-ing power and
efficiently perform the computation required for machine learning inferences. memory compared to MobileNet SSD or YOLOv5 [68].
Table 5 shows some TinyML inference engines along with their supported
platform and training library.

Overall, the Tiny Inference Engine provides the computing capabilities

needed to execute the models, while the TinyML Interpreter handles model Arm NN is a Linux-based, open-source software framework-work for
execution. The development tools for constructing and testing machine machine learning inference on embedded devices developed by the
learning models, such as data preparation tools, model training and company Arm. The core of the framework is the CMSIS NN library, which
validation tools, and performance profiling tools, are also included in the is a collection of optimized neural network kernels designed for maximum
frameworks. performance and minimum memory footprint on Cortex-M processor cores.
Arm NN makes use of fixed-point arithmetic, quan-tizing model parameters
A TinyML Framework's goal is to provide a comprehen-sive solution for to either 8-bit or 16-bit integers for deployment to microcontrollers for
constructing and deploying machine learning models on low-power devices, inferencing. The frame-work leverages the processing architectures of
making it easier for developers to create edge computing applications. Arm's popular Cortex-M series of microcontrollers to enhance execution
TinyML Frameworks serve to guarantee that machine learning models
operate quickly and successfully on small, low-power devices by pro-viding
a comprehensive collection of tools and technologies. time [62].

Among these frameworks we can cite: E. CONNECTED TinyML

TensorFlow Lite(TFL) [63] is a Google open-source deep learning TinyML algorithms are intended to be deployed on devices with minimal
framework designed for inference on embedded devices. It is made up of resources, such as micro-controllers. Achieving the connectivity of the tiny
two primary parts: the Converter and the Interpreter. The TensorFlow embedded device, while adhering to the given restrictions of low energy
Converter is used to con-vert TensorFlow code into a compressed flat consumption, may be possible using Low-power wide-area network
buffer (.tflite), shrink the size of the model, and optimize the code with (LPWAN).
minimal accuracy loss. TFL currently supports quantization, pruning and This type of wireless network has various benefits over standard IoT system
clustering. TensorFlow Lite Micro(TFLM) [64] is an extension of TFL approaches such as Bluetooth (BLE), WLANs, IEEE 802.15.4-based
designed to execute machine learn-ing models on 32-bit microcontrollers protocols (Zigbee or 6LoW-PAN), and cellular networks. Fig. 7 compares
with limited memory (just a few kB). It's been successfully ported to a the networking technologies, used by IoT devices, in terms of power con-
variety of processors, including the Arm Cortex-M Series and the ESP32. sumption, bandwidth, and range.
The core runtime just fits in 16 KB on an Arm Cortex M3 and is capable of
executing many basic models. LPWANs employ low-power radio frequencies and operate at low data
speeds. They are based on a cellular-like architecture, in which end-devices
communicate directly with a central gateway or a base station allowing to
It operates independently of an operating system, eliminated handle

VOLUME 11, 2023 96903

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 5. TinyML inference engine.

There are other LPWAN-based wireless communication technologies

available today, but LoraWAN is the most often used in TinyML projects
(Table 1).
LoRaWAN: LoRaWAN is a media access control (MAC) protocol
developed on top of LoRa, a communication technology that uses a
modulation method called Chirp Spread Spectrum (CSS) to send data over
vast distances while consuming minimal power. It is intended for wireless
battery-powered Things in a regional, national, or worldwide network [71].
The utilization of license-free sub-gigahertz radio frequency channels is a
significant characteristic of LoRaWAN. This allows the deployment of a
large number of gateways, enabling extensive coverage and low-power
com-munication with IoT devices. Sanchez-Iborra [72] evaluated the
connectivity of a wearable device within a test region using a pre-existing
FIGURE 7. Different wireless networks used in embedded devices.
LoRaWAN deployment. The results show that the device achieved good
connectivity both indoor and outdoor locations, with a coverage shadow
found in an area with dense vegetation and big buildings. The packet
a large number of devices while consuming little power. delivery ratio (PDR) obtained in the trials was over 92% in all cases, with
Thus, LPWANs are well-suited for IoT applications in which devices must better performance in the uplink direction due to the better noise factor of
run for lengthy periods of time on batteries or other restricted power the gateway radio module as com-pared with that of the end-device. Overall,
sources, and may be scattered across broad regions that sometimes can the results demon-strate that the TinyML device was able to achieve long-
exceed dozens of kilo-meters [69]. LPWANs support applications with low- range connectivity with an efficient transmission technology.
rate requirements where related devices can remain operational in the field
for years without battery charging or replacement.

They are generally considered as a type of Wireless Sensor and Actuator

Networks (WSANs) [70]. LPWANs also offer great scalability and node Sigfox: It deploys its own base stations in multiple countries using its
density, making them suitable for ultra-dense scenarios like highly populated patented ultra-narrowband (UNB) technology in the unlicensed sub-GHz
cities, and their network architecture does not require cooperation among ISM frequencies (eg 915MHz in North America and 868MHz in Europe).
end-devices or the use of routing protocols or node coordi-nation. The End-devices communicate with these base stations through BPSK mod-
research [24] proposed an example of employing LPWAN with TinyML in ulation in a 100 Hz ultra-narrow band with a maximum data rate of 100
a federated learning (FL) scenario, where a variety of edge devices bits/s [73]. Sigfox effectively utilizes the frequency band and has very low

collaborate to develop a global model using only local copies of the data. noise levels by employing ultra-narrowband in the sub-GHz spectrum,
This enables the model to be developed in a decentralized and distributed resulting in low power consumption, great receiver sensitivity, and low-cost
manner, eliminating the requirement for parties to share data. antenna design [74].

96904 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

NB-IoT: To empower the Internet of Things, 3GPP devel-oped 3) PRESSURE PREDICTION

Narrowband Internet of Things (NB-IoT) technology that may be used in The authors in [79] employed Tiny Neural Networks and edge computing
both licensed and unauthorized bands. methodologies to obtain numerical weather forecasts, especially pressure
It is an air interface that is part of the LTE standard but has been reduced forecasts. They employ a Tiny Deep Neural Network, a hybrid of Recurrent
including handover, channel quality measures, carrier aggregation, and Neural Network and Convolutional Neural Network. They evaluated four
dual connection. [75] It is meant to promote long battery life and low-cost distinct network versions and two different cost func-
devices by utilizing one resource block of LTE networks, equivalent to 180
kHz in the frequency range. [76] tions. The model was trained and validated using data from a recognized
weather station; the findings revealed that the forecasts were highly
accurate in matching the ground truth data. The model was then implemented
III. TINYML AND ENVIRONMENT on an STM32 micro-controller using a C-library that was tailored for
TinyML has the potential to play an important role in envi-ronmental simplicity and power economy. The output data was moved to a visual
problem solving. TinyML enables the deployment of intelligent devices platform for analysis. Using edge computing for this task can drastically cut
capable of monitoring and collecting data on numerous environmental costs and compute complexity; it while also allowing for the production of
aspects such as air quality, water quality, and meteorological conditions. online simulations based on atmospheric models.
Indeed, these devices are designed to be resistant to noise and changes in
data distribution. They may also be required to work in tough environments,
such as extreme weather temperatures or places with limited or no access
to power. TinyML devices may also be installed on a wide scale and offer B. HYDROSPHERE-RELATED APPLICATIONS

real-time data on the state of the environment, allowing for a quick 1) WATER QUALITY MONITORING The
identification and resolution of problems. authors in [80] proposed combining machine learning and TinyML-based
devices(Raspberry Pi) to build a method to monitor and evaluate the water
quality. Sensors capture data on numerous water quality variables(eg,
temperature, pH, and chemical material concentration) and communicate
A. ATMOSPHERE-RELATED APPLICATIONS 1) the data to a Raspberry Pi linked to a data center. The data is then analyzed
VEHICULAR EMISSION The by a machine learning model and transferred to the cloud for analysis. The

greenhouse effect, a key contributor to environmental concerns, is mostly authors emphasize the advantages of this technique, such as water saving

generated by automobile emissions. and the capacity to monitor many elements that impact water quality.

These emissions contain significant volumes of carbon dioxide. To solve

this issue, the authors in [77] employed TinyML and the on-board diagnostics
system, available in most auto-mobiles, to monitor the amount of CO2 The authors in [81] created prediction models for water quality utilizing

emitted per liter and send the data to Typicality and Eccentricity Data Ana- multiple characteristics employing a graphi-cal user interface (GUI) and a

lytic(TEDA), an unsupervised anomaly detection algorithm. deep learning algorithm based on Artificial Neural Networks (ANNs). Their
software prop-erly predicted the quantity of toxic compounds in a Malaysian

This method measures the quantity of CO2 created in grams per second; it lake.

saves the data on a microSD card for cloud processing by external systems
via Bluetooth or 4G. The algorithm's recursive structure makes it time-
efficient, with low computation and memory needs. The collected data can 2) WATER MISUSE DETECTION The
be used to guide the execution of public actions to mitigate the greenhouse authors in [82] developed a system for maintaining water reservoirs using a
impact. mix of Raspberry Pi, Arduino IDE, and ANN model. Sensors embedded in
water faucets collect and send data using Raspberry Pi, which analyzes it
with a deep learning application. The application may then determine
2) TEMPERATURE PREDICTION regular water usage as well as any leaks or waste. This has the potential to

TinyML's temperature forecasting is a critical application in the realm of increase water conservation and decrease water

environmental monitoring. Accurate temperature forecasts may help in

waste
making decisions such as a) when to utilize energy in buildings; b) when to
plant crops and c) what type of crops to plant. The authors in [78] proposed
an approch to predict temperature values; it uses an edge computing 3) WATER DISEASE DETECTION AND PREDICTION
technique and a TinyML-based device, such as the Arduino Nano 33 BLE Authors in [83] employed Edge AI to secure water sup-plies from dangerous
Sense. They processed time-series data and made temperature forecasts contaminants such as cholera. They criticized the present Alkaline Peptone
using a multilayer perceptron, a sort of artificial neural network, with three Wate (APW) tech-nique and advocated employing edge computing to
dis-tinct cost functions (Root Mean Square Error(RMSE), Mean Absolute construct an experimental setup to monitor the physicochemical parameters
Error(MAE), and R-squared). of water in order to avoid Cholera outbreaks. Because of the absence of
wireless connectivity in remote regions, they

VOLUME 11, 2023 96905

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

proposed to use an offline model that achieves 96% accuracy while and non-linear activation functions inside and the periphery pixel array. This
conserving battery power. The device was eventually connected with on- demonstrates the need for data transfer between the image sensor and the
tap inference and could monitor water safety metrics. The authors intend to accelerator for Convolutional Neural Networks (CNN) processing, resulting
expand and generalize the model in order to handle other illnesses that in lower energy and bandwidth requirements. The authors in [85] also offer
may contam-inate water. two CNN models that are implemented using PIP (pixel-in-pixel), which
yields considerable compression while reducing data rates and power
consumption.

4) UNDERWATER IMAGING
Traditional underwater imaging systems require active power sources,
which are not accessible in most underwater envi-ronments. These 2) EARTHQUAKE DETECTION
systems can be used to research marine species, climate change, marine Researchers in [86] focuse on earthquake detection as a vital initial step in
geography, aquaculture farms, particulate organic carbon transport, and Earthquake Early Warning (EEW) systems.
maritime archeol-ogy. Recent research has revealed that completely The authors point out that in order to provide real-time alerts, robust EEW
submerged, battery-free cameras and acoustic backscatter can be used for systems must have high detection accuracy, low detection latency, and a
on-site wireless underwater imaging. However, the narrow bandwidth of the high sensor density. They note that classic EEW systems rely on fixed
underwater acoustic channel makes picture acquisition and communication sensor networks or, more recently, networks of mobile phones equipped
energy inefficient. To solve this issue, the authors in [84] created proposed with micro-electromechanical systems (MEMS) accelerometers.
a fish visual wake word (fishVWW) model based on wake word models.
The authors propose a new technique for global-scale earth-quake
detection and warning based on IoT edge devices with TinyML capabilities
A wake word network is a sort of machine learning approach that is used to and always-on, always-connected sta-tionary MEMS accelerometers. They
detect a certain word or phrase (the ''wake word'') in a stream of data. Wake explored and evaluated deep learning ML algorithms for earthquake
word networks are often used in voice assistants such as Amazon's Alexa detection using a limited-resource Arduino Cortex M4 microcontroller (256
or Apple's Siri to allow users to activate the assistant by uttering the ''wake kB of RAM).
word'': ''Alexa'' or ''Hey Siri.'' The goal of a wake word network is to allow
devices to operate on minimal power most of the time. This is accomplished
by activating and processing audio data, only when the wake word is D. BIOSPHERE-RELATED APPLICATIONS
detected.
1) WILDLIFE CONSERVATION
Sensor technology is important in wildlife conservation efforts, especially
Similarly, the fishVWW model was created for battery-free underwater
when it comes to protecting endangered species. Authors in [87] discuss
cameras that can only take, compress, and send images when they detect
the use of ML techniques, including on-animal sensors for tracking
a fish. The fishVWW model was tested on the STM32L476RG, a very low-
movement and bioacoustic sensors, to gather data about wildlife and the
power microcon-troller (with 1 Mb Flash and 128 kbs SRAM). In order to
environment. These techniques can be used to improve our understanding
decrease the amount of needless data transmitted, the authors [84] did
of biodiversity and the environment. However, the paper also notes that
investigate eliminating underwater artifacts from the image.
there are limitations and challenges to be addressed, including with low
latency and low capacity due to the large amount of data and the vastness
of habitats.

C. LITHOSPHERE-RELATED APPLICATIONS
1) HYPERSPECTRAL IMAGING (HSI) Another application in the same scenario was presented in [88]. TinyML
TinyML has the ability to give useful insights in geology by analyzing sensor was used to small payload satellites weigh-ing less than 180 kg, often
data in real time and making it more accessible, usable, and interpretable. known as SmallSats, to meet a specific conservation task. The study
One possible TinyML application in geology is the analysis and classification focuses mainly on the protection of sea turtles, who are under threat from
of var-ious types of rocks and minerals using hyperspectral imaging (HSI). factors such as uncontrolled fishing and sea pollution. The TinyML
HSI collects detailed information on light reluctance at various wavelengths, framework was used to assist conservation efforts by implementing state-of-
which may be utilized to assess the composition and attributes of the the-art real-time vision-based TinyML support. The study's goal was to
materials being photo-tographed. This data may be utilized to pinpoint utilize this technology to track sea turtles in real time, giving real-time data
specific rock formations, mineral deposits, and geologic features. HSI pic- on their position, activity, and population to assist enhance the efficiency
tures, on the other hand, are often high-dimensional and need a significant and effectiveness of conservation operations. This may be utilized not only
amount of power and storage. To address these issues, the authors of the for sea turtles, but also for other conservation duties such as illegal hunting,
research [85] offer an algorithm-hardware co-design solution for 2D TinyML logging, and animal welfare.
workloads. They insert proposing sophisticated calculations like convolutions

Hackster.io and Smart Parks organized an event named ElephantEdge

[89] to encourage members of Hackster.io to

96906 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 8. WASN-based soundscape monitoring system architecture.

develop TinyML models for tracking collars, with the goal of reducing the to anticipate soundscapes using an RF52840 32-bits-microcontroller on a
loss of elephants due to illegal ivory poaching, trophy hunting, human low-cost edge device. The sys-tem's machine learning algorithms were
conflict, and environmental degra-dation. The authors in [90] proposed two able to identify noises into four categories: anthrophony (human sounds),
TinyML models based on two different inputs such as, audio and traffic, biophony (animal sounds), and geophony (environ-mental sounds).
accelerometer data(x,y,z). The first model detects poacher attacks based The final model utilized in the system was a CNN based on mel-frequency
on the sound of kalashnikovs and bush arrows. The second model cepstral coefficients (MFCCs), with an accuracy of 81.6%. The paper
proposes a system architecture using three layers: the edge layer, which
detects poacher attacks based on the elephant's physical activity-ity. Both consists of sensors and a machine learning model running on the Arduino
models were created using Edge Impulse Studio. Nano 33 BLE Sense development board; the fog layer, which consists of
an ESP32 and a LoRa transceiver acting as a gateway; and the cloud layer,
2) BIO-ACTIVITY MONITORING Using which handles global storage as well as web-based application for

TinyML can also help monitor vital biological activi-ties of different species, visualization and analysis using Grafana and Node-Red. The development

providing valuable information about their behavior, movement, and board is equipped with environmental sensors, a MEMS microphone, and

ecosystem. Ultrasonic sensors can be used to track bat activity, as the nRF52840 processor, and is capable of running TinyML and TensorFlow

described in [91]. The sensors capture sound waves and generate a Lite. The sensor mod-ule also includes a real-time clock and a small battery

spectrogram, which is processed by Machine-Learning-based hardware to for retaining date and time. Fig. 8. depicts the architecture of the soundscape

classify the signal and provide information about the bat's location, genus, monitoring system.

and time of recording. This approach can give insights into biodiversity loss
and other important ecological information.

3) SOUNDS CLASSIFICATION IV. TinyML AND HEALTHCARE The

TinyML might be useful in the subject of biophony, which is the study of the development of TinyML has inevitably touched the sector of healthcare,
sounds generated by living creatures in their natural surroundings. It is which is significant in every society worldwide. And since the human body
feasible to analyze and interpret the complex audio patterns created by can be considered as a source of signals emitted by a variety of organs,
different kinds of sources in a studied environment using TinyML mod-els sensors can be deployed to collect these data and use it to mitigate
on small, low-power devices, and potentially utilize this information for dilemmas related to healthcare. The deployment of TinyML will fit to solve
conservation and environmental management. these issues as illustrated in table 6. Pre-vious surveys [20], [23], [93]
indicate that the integration of these edge devices will revolutionize the
TinyML was utilized in the article [92] to build an in-situ soundscape healthcare sector, and enhance the well-being of people. The latter can
monitoring system based on a wireless acoustic sensor network (WASN).
The system was created

VOLUME 11, 2023 96907

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 6. Summary of the applications of TinyML in healthcare.

provide quality and reliable healthcare monitoring; in addition to the A. BLOOD-PRESSURE MONITORING
improvement of many health products. Moreover, the technology has the One way to see how TinyML can find its place in health-care is by
potential to enable new forms of medical monitoring, diagnostics, and investigation [95], where researchers applied the edge device to monitor
therapy, improve the quality of care, and ultimately improve patient high level blood pressure, called ''TinyCare'' [95]. The authors developed
outcomes. a cloud independent TinyML solution that merely relies on data obtained
Especially with the ability of TinyML to process data in real-time and in low from patients. The authors in [95] adopt a systematic procedure to tackle
profile devices. This opens new pos-sibilities for healthcare professionals the problem, starting by preprocessing the data based on physiological
to monitor and treat patients in a more effective and efficient way [94]. The signals and then extracting the features.
analysis of previous works allows us to amalgamate the general process
adopted to apply TinyML in healthcare as illustrated in Fig. 9. The human Their model uses a variety of ML algorithms deployed on three Edge
body is the provider of biological signals that are detected by specialized Devices: Arduino uno, ESP32 Wrover Board, and AdaFruit PyBadge. [95]'s
sensors and later transmitted to the heart of the embedded system that methodology enabled to test a variety of models, not only with respect to
consists of ML or DL algorithms interacting with a cloud environment. This accuracy, but also latency and complexity.
combination allows the device to reveal the type of deficiency, as well as
to monitor, predict or assist the patient. In this section, we explore promised
TinyML-based solutions in the healthcare field, as summarized in table. 6., B. NEURAL SPEECH ENHANCEMENT FOR HEARING AIDS
by shedding light on some of its trending applications. Noise suppression and Hearing Aids are medical solutions used by people
with damaged ears, they help in the decrease of listening difficulties,
especially in noisy environments.
Many models were developed through usage of RNN [96], the authors in
[97] built on these models, but using TinyML.

96908 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 9. Demonstration of the general process adopted to apply TinyML technology in the healthcare sector.

The base of the hardware was chosen to be the Hearing Aid, and the through the detection of epileptic seizure. The experimental setup included
authors used RNNs and pruning techniques to create the model that would an ML model (Random Forest) as a classifi-cation algorithm, that classifies
enhance the hearing of speech. Authors in [97] acknowledge the constraints signals after being deployed on STM32L476 ARM cortex-M4 microcontroller.
imposed by the edge device, since they required to train large neural The entire setup accomplishes up o 40.87 hours of monitoring using a
networks with large data-sets due to the limited storage feature; However, single battery charge (approximately a whole day of work); in addition to
the authors were able to achieve moderate satisfactory rating for their reduced amount of false-alarms that [101] worked on to obtain. are all
model. domains to apply this technology. The authors in [102] and [103] introduce
TinyML as a medical embedded system for the diagnosis of focal liver
lesion (FLL) using a DNN model trained on liver images that were priory
C. IMPROVING WEARABLE AND AMBULATORY SYSTEMS obtained using ultrasound imaging. The entire process was implemented

According to [98], TinyML can generate many embedded solutions and using computer aided diagnosis, which was later transferred in a small

optimization techniques in the domain of wearable devices in Healthcare. memory footprint that runs in an edge platform [102].
Such contributions can be recognized in the detection and recovery from
FoGnin Parkinson's patients, achieving high accuracy of 93.58% by a ML
model deployed on ATMega2560 microcontroller that is able to balance
between power consumption and processing speed [99]. The applications
of TinyML in the healthcare domain can be further extended to the detection D. EMOTION DETECTION
of Cardiac Arrhythmia, an irregular heartbeat caused by abnormal electrical Another substantial application of TinyML comes with emo-tion detection.
activity in the heart. The detection process is boosted by the usage of low- The authors in [104] and [105] aim to build a smart wearable device to
powered microcontrollers as indicated by [100]. The authors in [100] rely achieve the recognition of emotions through physiological signals emitted
on the CMSIS-NN library as a software tool to deploy their convolutional by the human body. For instance, in [104] from the respiratory belt,
neu-ral network. The authors reported a power efficiency of 1.64 GOps/s/ photoplethysmography (a measuring tool of variation of blood in a certain
W achieving a reasonable accuracy of 78.4% (including the implementation organ), and fingertip temperature, mean-while, The authors in [105] used
of CMSIS-NN). The treatment of epilepsy is a typical paradigm for the bioelectrical measurement technologies to measure parameters such as
application of TinyML. The authors in [101] deploy an ultra low powered skin conduc-tance, electroencephalographic, and heart rate signals. After
wearable device in treatment of this neurological deficiency deploying the data collected by sensors on trained ML mod-els, [105]
obtained an accuracy that ranges from 53.6% to 69.9%, in contrast, [104]
achieved an accuracy of 69% to

VOLUME 11, 2023 96909

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

73.08% which proves that these devices can be further har-vested and Agriculture into sustainable, smart, efficient, and eco-friendly agriculture
polished in the identification of the emotional state of a person, which may named Smart Agriculture (SA) or Smart Farming (SF) [110].
contribute in the improvement of ergonomic conditions of people around the
world. IoT plays a vital role in this transformation. Thanks to IoT solutions,
farms can monitor the health of the crop and soil, detect any diseases that
E. IDENTIFICATION OF DEADLY MOSQUITOES can affect the plants, check the growth of plantation with drones and more.
Applications in healthcare also involve the detection of harm-ful species The electronics industry has seen significant advancements, leading to the
that provoke disease and infectious illnesses. availability of high-quality and cost-effective components such as MCUs,
Works already exist before adopting TinyML in the process, such as [106], single-board computers, sensors, and radio transceivers. The newer
where they use CNN to classify audios captured through smartphones. generation of MCU are not only capable of performing standard sensing
However, the authors in [107] introduced TinyML technology by establishing and control tasks, but also support complex operations such as running ML
an approach to detect hazardous mosquitoes using a model sensor that model. Moreover, container-porary radio technology has progressed to the
compiles wing beats to audios, which will be classified by a Tiny Embedded point where long-range transmissions can be achieved with lower energy
System that employs both Machine Learning and Edge Impulse Platform. consumption.
The authors in [107] rely on Arduino Nano BLE 33 Sense-based prototype
that is equipped by a trained ML model capable of classifying collected
audio data of wing beats so as to identify the type of species in which the The SF IoT network employs IoT devices to gather data on soil, crops,
mosquito belongs to. The model of [107] achieves an accuracy of 88.3%, greenhouses, irrigation, and weather via cam-eras or sensors [110]. This
which is a good result especially with respect to the low-cost and low-energy data is transmitted to the cloud through Wireless Sensor Networks (WSN)
consumption. and can be utilized by farmers to monitor crop health and identify diseases
in the plants [110]. By analyzing and interpreting the data, farmers can
understand the relationship between different agricultural factors, such as
soil characteristics and climate variables, which aids in informed decisions
and effective planning [111]. In this context, ML plays a crucial role in
F. TINY RESERVOIR NETWORK FOR THE DETECTION OF modeling the complex patterns present in the data, forming the backbone
PATHOLOGICAL CONDITIONS of decision support systems.
One of the areas where tinyML can be applied is in the processing of
electrocardiograms (ECGs). ECGs are sig-nals used for measuring the
electrical activity of the heart, which provides considerable information There are various architectures used in SF, depending on the specific
about the cardiovas-cular state of the human body. The authors in [108] application. Some common architectures include those with two layers
aim to reduce the complexity of this process by introducing a tinyML based (physical-edge), three layers (physical-edge-cloud) and four layers (physical-
ECGs that supports reservoir computing (eg a machine learning algorithm edge-fog-cloud). Many SF applications use three-layer architecture [112]
that uses a recurrent neural network (RNN) with a fixed architecture). In (see Fig. 10). The physical layer consists of sensors, actuators, and drones
their model, they deployed their model low-power microcontroller unit ability that collect data from the soil, animals, greenhouse, and weather. Data
to process biological signals coming from the ECG for the recognition of a collected are transmitted to edge nodes using WSNs. The edge layer
variety of pathological conditions. Their setup consumed less power, and consists of com-puting devices that interpret and analyze data received
acquired an accuracy of 95.4% with variance over the processed data of from the physical layer. The edge nodes have low or medium com-puting
0.001 [108]. resources. The cloud layer represents the brain and data storage in SF
architecture; it has a high storage capacity to save the data generated by
various sensors. It also aggregates and draws insights to provide ML
The adoption of tinyML-based ECG analysis can unlock many new use models for decision-making.
cases, such as continuous monitoring and real-time feedback for medical
doctors or patients.

V. TinyML AND SMART FARMING Complex ML algorithms require significant computing resources for
According to the United Nations(UN) [109], the world population is expected effective execution, leading to the adoption of cloud computing; it has the
to reach 9.8 billions by 2050. As con-sequence, the requirements of capacity to handle large ML models with millions of parameters, as well as
agricultural products are con-tinually increasing to serve the future the high-speed processors and gigabytes of memory needed to run these
population. However, growing rapidly population, climate change, soil models efficiently. However, this can be challenging in cer-tain locations
degradation and depletion of natural resources are all factors affecting the [113], such as in Africa [18], where internet bandwidth may not be sufficient
food production. The need for an evolutionary agriculture to keep up with to support the quick trans-mission to the cloud. Poor internet connection
growing demand in food production is necessary. causes some issues such as huge latency, data loss and the reliability
issues [112]. Moreover, frequent access to the cloud increases cyber-attack
Farmers, scientists, and agricultural industries turn to new technologies threats and decreases data protection [114].
and solutions such as IoT, drones, ML, big data, cloud, fog and edge
computing to transform traditional

96910 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 10. Smart farming architecture.

case is crop management with 8 papers, smart greenhouse

with 3 papers and smart irrigation with 2 papers. The table 7
summarizes the application of TinyML works in SF in terms
of area, dataset, classification algorithm, target device,
frame-work, power consumption, memory consumption,
latency and accuracy.

A. CROP MANAGEMENT
Crop management refers to the various techniques used to
grow and maintain crops in an efficient and sustainable
FIGURE 11. Number of articles on TinyML in smart farming.
manner. This can include things like proper irrigation and
fertilization, pest and disease management, and optimizing
To overcome the challenges posed by the cloud, some crop yields. TinyML can be used to gather data on factors
specialized techniques have been developed that bring the such as soil moisture and temperature, as well as analyze
computation process closer to end devices, such as fog and this data to make predictions and provide insights that can
edge computing. In recent years, the integration of TinyML in improve crop yields and efficiency. The authors in [115]
SF has also gained attention. TinyML allows sensor devices designed an embedded ML pipeline that helps farmers and
to perform ML tasks, such as monitoring crop health, scientists to monitor the health of the crop and its growth.
detecting plant diseases, and predicting the best crops, This pipeline allows users to create an embedded ML that
without the need for the cloud. TinyML offers several can be used for different plants in labs, greenhouses, farms
advantages over cloud, fog, and edge computing in terms of or gardens. The first step of the pipeline is data collection,
privacy, security, latency, and energy consumption [112]. where the authors proposed best practices to collect data
TinyML holds enormous potential in SF, especially in plants. The next step is training a Convolution Neural Network
Africa where embedded systems and AI are currently under- (CNN) for two cases: a) estimation of Leaf Area Index (LAI)
utilized [18]. PlantVillage [128] is one of these opportunities. and b) prediction of the plant growth stage. After training
It is an open-source project managed by Penn State University. phase, the ML model is compressed and converted to
The team in this project [128] has created the Nuru app to TensorFlow Lite(TFLite) format to be deployed on MCU
aid farmers in identifying and combating plant diseases. By device. For testing, the authors have chosen Sony Spresense
using ML through TensorFlow Lite on mobile phones, the setup as target device.
app provides real-time solutions without internet access, The authors in [33] proposed a TinyML solution to detect
which is crucial for farmers in remote areas. The future drought stress in soybeans. The system is composed of a
development of the system will utilize TinyML and TensorFlow Raspberry Pi zero W and Sony IMX219 camera module. The
lite micro to install sensors across distant farms, resulting in Raspberry device runs CNN model on the captured image to
improved tracking and analysis [24]. detect crop drought stress and then sends prediction to a
In this section, we explore promising TinyML-based web platform. The CNN model was converted to TFLite
solutions in SF field. As shown in Fig. 11, the most deployed format in order to be deployed on limited-resources device.

VOLUME 11, 2023 96911

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 7. Summary of the applications of TinyML in smart farming. Annotation: ''–'' is used to signify that authors did not provide information.

The authors in [116] proposed a low-power and real-time image The authors in [118] developed a TinyML CNN model aims to classify
detector for grape leaf esca diseases based on a compressed CNN model. between infected and healthy coffee plant.
Many compressing techniques such as, CP decomposition, tucker This model aims to monitor the health of the coffee plants and prevent the
decomposition and tSVD are analyzed to choose the method with the best propagation of the epidemic to the oth-ers. The authors compressed a
compression factor and accuracy. CP decomposition is chosen and it model with Qm,n format using X-CUBE-AI tool. The TinyML model is
applied on CNN model. After training and validation of the model; it deployed on STMicroelectronic ''STM32F746G-disco'' board connected to
compressed with post-training quantization using TFLite to generate a an STM32F4DISCAM module; both embedded in a box equipped with
model with 8bits. The compressed model is deployed on OpenMV Cam LEDs which are used to light coffee leaves while taking picture(see Fig. 12).
STM32H7. The device is mounted on an agricultural vehicle moving with a
constant speed through the cultivation field.
The authors in [120] built a prototype to identify and classify fruits and
vegetables from images deployed on MCU

96912 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

The authors in [126] proposed a low-cost device based on the MCU

ESP32-CAM that uses a camera to gather data from numerical water
meters to monitor central pivot irrigation systems. The device runs a TinyML
model to process the images in order to read the water meter; it then
transmitted to a server using LoRaWAN. The TinyML model scored an
accuracy of 88%.

C. SMART GREENHOUSE
A greenhouse is a structure designed to provide controlled environment for
plants to grow. This controlled environment can include factors such as
temperature, humidity, light, and nutrient levels, which can be adjusted to
optimize crop growth. Greenhouses can be used to grow a wide variety of
FIGURE 12. Coffee plant disease prototype [118].
crops, including fruits, vegetables, and flowers, year-round, regardless of
ESP-32. This prototype helps farmers to identify fruits and vegetables in the outdoor weather conditions. TinyML can be used in greenhouse to
real-time. They performed experimentation on three models such as CNN, optimize crop growth by using ML on low-power devices to gather data and
pre-trained MobileNet model v1 and v2 in order to choose a better model make predictions.
in terms of accuracy, inference time, RAM and flash occupation. These
models were converted to TFLite format. Finally, the Mobile Net model v1 The authors in [123] proposed a low+cost device that aims to recognize
was chosen to be deployed on ESP-32. different types of gas (NH3, CH4, N2O). This device helps farmers to
monitor greenhouse gases coming from the soil. They chose an ANN
The authors in [31] proposed a device(STM32F401CC) with a lightweight algorithm to classify three types of gas. It was compressed and deployed
neural network model to predict the best crop to grow based on the on NUCLEO-L476RG device using X-CUBE-AI toolchain.
observed soil parameters, such as nitrogen, phosphorus, potassium, soil
PH, temperature, and humidity of the soil. The embedded ML model The authors in [124] developed a multi-label TinyML model based on
predicts five crops: beans, maize, lentil, peas, and watermelon. The device Multi-Layer Perceptron(MLP) architecture for micro-climate control of a
sends the prediction results to the cloud via GSM to notify farmers. The strawberry agricultural green-house. The model takes five parameters as
model is created and optimized via Edge Impulse platform. input such as, temperature, humidity, soil humidity, solar illuminance and
the air CO2 concentration. And then generates as output a five-dimensional
vector where each binary numerical value is associated with one control
The authors in [94] built a prototype to detect and count the number of action to be performed automatically within the greenhouse. The model
strawberries in the image. This prototype aims to help farmers to know scored an accuracy of 97% on the validation sets, and 96% on test set,
where and when strawberries need attention to harvest them or to apply with a number of 151 trainable parameters.
fertilizer. The authors use the new ML model for constrained devices
developed by Edge Impulse named FOMO (Faster Objects, More Objects).
The authors in [125] proposed a system where IoT devices and edge
They used Edge Impulse cloud platform to design, train, validate, optimize nodes work together to monitor the status of the plants in a greenhouse to
and deploy ML model on Arduino Portenta H7. make decisions about the operation of sprinklers. Each IoT device uses its
The authors in [127] created a TinyML image classification deep neural embedded TinyML model based on Decision Tree(DT) to determine the
network model deployed on smart camera to detect crop diseases. The needs of the plants, and then sends this information to the edge node.
study applies depthwise sepa-rable convolution module instead of standard
convolution and squezze to reduce model parameters and computational The edge node collects decisions made by all the sensors for a specific
complexity. The Sony-CXD5602 MCU device has been used as a target sprinkler and uses an ML model to decide the final action (''no action'',
device. The model achieved a good performance in terms of time inference ''irrigation'', or ''fertigation''). The study uses Arduino Uno board as a target
and accuracy on two publicly available agricultural datasets such as, potato device for end device and edge node. The TinyML model scored an
and tomato datasets. accuracy of 99%.

B. SMART IRRIGATION VI. TinyML AND ANOMALY DETECTION Anomaly

The authors in [32] proposed a system for agro-environmental management detection(AD) or outlier detection refers to the techniques for identifying
employing moisture sensors and real-time video analysis of soil photographs. patterns or observations in data that do not conform to an expected behavior
The VGG-19 model is used to distinguish picture types and calculate the [129]. These techniques are used in variety of IoT smart applications [129]
quantity of water required for irrigation based on the kind of crop planted. such as, smart city, smart monitoring and smart power management. Mostly
of this applications use sensors as an input device. These devices generate
The model is trained on a Kaggle soil structure dataset and evaluated on a a huge volume of data
proprietary dataset.

VOLUME 11, 2023 96913

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

that are transmitted to the cloud servers for analysis, decision making and authors chose for the MicroPython firmware as base firmware for the MCU.
storage [129].
Many detection methods have been introduced in the lit- erature to The authors in [135] compared two types of TinyML models, an
detect anomalies on data such as, geometrical, statistical, and ML [130]. autoencoder and a variational autoencoder for mechanical anomaly
Data-driven approach using ML algorithms and deep learning are becoming detection in washing machines using an MCU device (Arduino Nano 33
more and more popular as a way to identify abnormal situations [131]. It BLE Sense). The MCU device collects accelerometer data (x, y, z) and runs
used to build models from large data generated by sensors to distinguish the ML model to detect an imbalanced laundry load. To determine which
between normal and abnormal classes. ML model should be deployed on the MCU device, the authors evaluate
models using accuracy, precision, and recall.
Cloud-based architectures have been used for many anomaly detection
applications [129]. An example of cloud based anomaly detection is The autoencoder model was chosen due to its performance in terms of
condition monitoring [132]. The study uses cloud computing to monitor the accuracy. Both models were written in Python using TensorFlow library and
health condition of machines using vibration signals. The edge devices then converted to TFLite format to be deployed on Arduino device. The
collect vibration data and transfer them to the cloud for storing. results show around 92% accuracy and 90% precision.

Cloud provides an ML model to detect machine's abnormal behavior.

Analyzing and processing data in the cloud server poses some limitations B. PREDICTIVE MAINTENANCE
such as, latency and security issues. The authors in [136] created a TinyML model deployed on ESP-WROOM-32
MCU device to detect anomalies in ther-mal images; it sends data only
TinyML has the potential to play a big role in anomaly detection by using when it detects an anomaly on a machine to the server using Message
ML at the edge, allowing for real-time monitoring and identification of Queuing Teleme-try Transport(MQTT) protocol. The MCU device runs a
unusual patterns or behavior without a connection to a more powerful CNN model created by Keras and converted using tinyml-gen library from
computing device. In this section, we explore exciting work that integrates EloquentArduino. The study was tested on hydraulic refrigeration system.
TinyML in AD as summarized in table.8 in terms of application, dataset, The results show around 94% accuracy and 92% f1-score.
ML algorithm, target device, framework, accuracy, latency, power and
memory consumption.
The authors in [137] proposed an online learning anomaly detection
model named Deep Echo State Network(DeepESN) to monitor the reliability
of water distribution systems; it adapts itself to the environmental changes
A. CONDITION MONITORING and it can be deployed on a MCU. DeepESN is based on Recurrent Neural
The authors in [133] developed TinyML model deployed in STMicroelectronic Network(RNN). To optimize the complexity of the models, the online learning
STM32H743Z12 to detect anomalies in rotating machinery. The MCU can be made in two different ways: single iteration and batch decomposition.
device acquires vibration signals through an accelerometer, process it to The model was tested on STM32H743ZI Nucleo Arm.
extract features and store data on MCU to train locally an Auto-encoder ML
model. The Auto-encoder model is trained sequentially by providing the
training data in small batches at a time (4 batches of 100 samples). After The authors in [138] proposed a Block based Binary Shallow Echo State
training step, the model is stored in flash memory of MCU to be used during Network (BBS-ESN) model which is a deeply quantized anomaly detector
the inference. To determine whether an anomaly is present or not, unseen of oil leaks that happen in the wind turbines. This TinyML model uses
data is input to the model and the MCU calculates the MSE between the DeepESN, binarized images and one-bit quantization of weights and
sample and its reconstruction. If the MSE is larger than the anomaly activations function. The model was tested on STMicroelec-tronics NUCLEO-
threshold, it is determined that an anomaly. H743ZI device.

C. INTERNET OF INTELLIGENT VEHICLES

The authors in [134] describe a system for detecting anomalies in Authors in [139] proposed a MCU device attached to a vehi-cle to detect
submersible pumps at wastewater management plants using a retrofitting road anomalies such as potholes, bumps and obstacles. The study uses
kit with MCU. The kit includes temperature and vibration sensors, an Arduino Nano 33 IoT that collects accelerometer data and runs an
ESP32DEVKIT MCU, and power line communication equipment installed in unsupervised TinyML algorithm named TEDA. The results on real
the terminal chamber of the pump. The MCU collects data from the sensors, experiments (Asphalt pavement) show an accuracy of 99% and an f1-score
processes it to extract features, and uses this data to train locally an of 0.76.
Isolation Forest model to identify anomalies in the data stream. After training
step, the MCU switches to inference model. Even during this mode, the
MCU continues to process and learn to update the model. The model is VII. NAVIGATING THE CHALLENGES: AN OVERVIEW OF
devel-oped using python 3 and libraries NumPy and SciPy. The THE CHALLENGES FACING TINYML
Investigating the challenges of TinyML technology is a cru-cial part in the
current survey as they define the state of the

96914 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 8. Summary of the applications of TinyML in anomaly detection. Annotation: ''–'' is used to signify that authors do not provide information.

art, open research questions, and possible directions. In spite of the proposed a novel solution called TinyOL (TinyML with Online-
potential benefits and advantages that TinyML offers, as detailed in the Learning) allowing MCUs to learn on stream-ing data. This study is
aforementioned sections, it is important to understand and address the based on the concept of online learning. Online Learning follows a
obstacles and challenges that may impede its progress. From technical process of updating a model's parameters in real-time using new data
limitations to societal concerns, these challenges must be carefully as it becomes available, rather than using a fixed dataset. This allows
considered in order to fully realize the potential of this technology, and the model to adapt to changes in the underlying data distribution and
strengthen its capacities to fulfill further tasks in human society. In this improve its performance over time.
section, we explore the various challenges that are currently facing the
implementation and adoption of TinyML. Other implementations of On-Device learning are pre-sent on this
paper [10]. • Limited Memory:
We delve into the specific issues in more detail in an attempt to provide The memory challenge in TinyML refers to the difficulties that arise from
insight on possible solutions. the limited amount of memory that is available on these devices.
• Evolution of the Environment: The current TinyML solutions are based
on offline learning. ML model is first trained on powerful machine or This challenge remains a major challenge and a trade-off between
cloud and then deployed on tiny edge device. Edge devices cannot the model's performance and memory usage.
adapt themselves to the evolution of the environment because they These devices, such as sensors, wearables, and IoT devices, have
cannot learn from the data. The ML model perfor-mance will limited computational resources and mem-ory, which makes it
accordingly be dropped. This problem is known as concept drift difficult to run complex machine learning models. Larger models tend
[140]. On-device learning facil-itates the transition from offline ML to be more accu-rate, but they also require more memory. This can
model training to update it automatically using real-time data. lead to trade-offs between model accuracy and memory usage, which
Attempts already addressed this issue such as, the authors in [140] can result in models that are less accurate than desired. In addition,
the increase of memory, which is

VOLUME 11, 2023 96915

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

FIGURE 13. Illustration of the security challenges facing TinyML, indicating potential security estimates within the process.

a power-hungry component, yield to heavy power con-sumption researchers wish to exploit this technology by generating
which can affect the battery life of the edge devices. To avoid the experimental complexity, which may also lead to discrepancies
memory constraint, data might be stored in a remote server and between frameworks [18]. • Accuracy Drop: As
accessed on-demand, this can increase the latency and many researchers report a decrease in the accuracy of the ML model
communication overhead. once imported on the edge. Especially that TinyML systems are typi-
Some possible solutions exist via AI Model Efficiency Toolkit (AIMET) cally designed to work low-power devices (eg wear-ables, edge
[141], which blends the AI trained algorithm with compression and devices, IoT devices) the limited available resources of these
quantization which leads to the optimization of the ML or DL model devices can make it challenging and difficult to achieve the same
while maintaining the same accuracy. level of accuracy as larger systems. The limited amount of data that
can be stored and processed on these devices can also lead to lower
• Heterogeneity of Hardwares and Softwares: which mainly includes the accuracy. Additionally, the compression of the algorithm may also
diversity in the devices and systems being used, each with their own yield to the reduction of the accuracy, further reducing accuracy. In
unique software and features [23]. Which creates a challenge when [100] for instance, the team recorded an accuracy dropping from
it comes to managing and coordinating the various systems, as dif- 80% to 78.4% as soon as they deployed it in the edge. In [99], the
ferent software may not be compatible or may require different levels accuracy drop is estimated to be 1.3%. These percentages are
of resources in terms of the operating system and programming significant considering the sensitivity of the healthcare domain. • The
language; in addition to the varieties of levels of memory, processing Privacy Problem: some of the most common sen-sors used in TinyML
power and storage. The heterogeneity of data is also worth are cameras and microphones that may possibly collect sensitive or
mentioning. It mainly refers to the diversity and complexity of data private information about the user without their knowledge which
that is captured and processed by small, low-power devices sensors. obviously implies many privacy concerns. The limited resources
This can be a result of the inevitable noise that interferes with real available in this technology obstruct the ability to secure and protect
data. Format and resolu-tion also take part in this heterogeneity as the collected data and its storage, which makes it hazardous to many
they hinder the ML algorithm to generalize the model. In addition to excerpts such as unauthorized manipulation or access by impostors.
that, Tiny-ML algorithms require an assembly of sys- tems, devices
and software, including microcontrollers, sensors and the cloud
environment, which may obstruct

As indicated by figure 13, the whole edge and computing

96916 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

TABLE 9. Table of acronyms. in the healthcare domain where the life and health of patients is at stake,

and the robustness of the device becomes a necessity. Additionally,

researchers of [143] corroborate the reliability challenge of TinyML,
which consists of the variations in precision of these devices, since they
are vulnerable to variety of extensive factors; in addition to the errors
that may occur during the construction of circuits and assemblage of
wafers. The hardware layer is also exposed to high energy particles that
modifies the output of the algorithm. Finally, aging of the device and the
degradation of its components may provoke other variations in battery-
life, accuracy, and even data collection by sensors.

VIII. CONCLUSION Over

the last couple of years, the topic of TinyML has gained a huge attention from
industry and academia, with the drive of unlocking various new possibilities for
sustainable development technologies. Artificial intelligence has become
nowadays ubiquitous and has demonstrated its capability to bring new
approaches and solutions to various research problems. However, training AI
algorithms need extensive computation as well as specialized (costly) hardware,
which leads to a higher energy consumption and a significant carbon footprint.
To overcome such AI issues, TinyML is the suitable candidate technology,
which will make the future of AI tiny and bright. In this paper, we present the
results of a comprehensive literature survey of all TinyML applications and
related research efforts. Special emphasis has been placed on building a
taxonomy of TinyML techniques that have been used so far to bring new
solutions to various domains, such as healthcare, smart farming, environment,
and anomaly detection. Our survey has shed light on the new insights of
process is susceptible to unethical security criticism, which implies TinyML and how it unlocks new possibilities for sustainable develop-ment. For
unsafety in terms of data collection and data flow. The wide deployment this we discussed how TinyML reduces latency and enables real time
of TinyML may yield to the accumulation of personal data of users with applications to be deployed in the source of data. We also brought to the
respect to the lack of regulation, standardization and specifi-cations of attention of AI researchers how TinyML models can run even when there is no
this technology, and thus exacerbating the privacy issue. Attempts internet connection, and how the processed data do not leave the device,
already addressed this issue such as the TinyMLaaS [23], where they which significantly improves user privacy and thus complies with data protection
build an embedded architecture capable of confining business sensitive regulations. Finally, this survey highlights the remaining challenges and
data within the IoT device boundaries. However, to fulfill this shield of discussed future research directions. We believe that our survey will serve as
protection, the device should be constrained by a narrowband connectivity a guideline for the future research initiatives in TinyML and will motivate further
(NB-IoT) meaning that the device should have a very limited possession discussions in this promising field.
of data transmissions. Another promising model that considers this
important concern is [142], where they established a TinyML model
whose data is processed within the device, which [142] called it privacy-
centric on-device transfer learning, without any interaction or sharing
with the cloud and external servers, Hence solidifying data privacy and
its security. • Product Trustworthiness and Reliability: TinyML is a
relatively new field that deals with the application of machine learning
algorithms on small, low-power devices such as microcontrollers and REFERENCES
sensors. The device will definitely gain many use cases in a variety of [1] D. Perez-Liebana, J. Liu, A. Khalifa, RD Gaina, J. Togelius, and SM Lucas,
sectors and domains; However, ''General video game AI: A multitrack framework for eval- uating agents,
the trustworthiness and reli-ability of such device remains questionable, games, and content generation algorithms,'' IEEE Trans.
Games, vol. 11, no. 3, pp. 195–214, Sep. 2019.
especially
[2] NL Bragazzi, H. Dai, G. Damiani, M. Behzadifar, M. Martini, and J. Wu,
''How big data and artificial intelligence can help better manage the
COVID-19 pandemic,'' Int. J. Environ. Res. Public Health, vol. 17, no. 9,
p. 3176, May 2020.
[3] FE El Aidos, M. Kassab, N. Benamar, and B. Falah, ''A comprehensive
survey on blockchain-based solutions to fight COVID-19 pandemic,'' Int.
J. Comput. Digit. Syst., vol. 11, no. 1, pp. 873–892, Feb. 2022.

VOLUME 11, 2023 96917

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

[4] O. Baclic, M. Tunis, K. Young, C. Doan, and H. Swerdfeger, ''Challenges and opportunities [27] CR Banbury, VJ Reddi, M. Lam, W. Fu, A. Fazel, J. Holle-man, X. Huang, R. Hurtado, D.
for public health made possible by advances in natural language processing,'' Canada Kanter, A. Lokhmotov, D. Patterson, D. Pau, J.-S. Seo, J. Sieracki, U. Thakker, M. Verhelst,
Communicable Disease Rep., vol. 46, no. 6, pp. 161–168, Jun. 2020. and P. Yadav, ''Benchmarking TinyML systems: Challenges and directions,'' 2020,
arXiv:2003.04821.
[5] A. Esteva, K. Chou, S. Yeung, N. Naik, A. Madani, A. Mottaghi, Y. Liu, E. Topol, J. Dean, and
R. Socher, ''Deep learning-enabled medical computer vision,'' NPJ Digit. Med., vol. 4, no. [28] V. Rajapakse, I. Karunanayake, and N. Ahmed, ''Intelligence at the extreme edge: A survey
1, pp. 1–9, Jan. 2021. on reformable TinyML,'' 2022, arXiv:2204.00827.

[6] FA Ozbay and B. Alatas, ''Fake news detection within online social media using supervised [29] SS Saha, SS Sandha, and M. Srivastava, ''Machine learning for microcontroller-class
artificial intelligence algorithms,'' Phys. Ah, Stat. hardware: A review,'' IEEE Sensors J., vol. 22, no. 22, pp. 21362–21390, Nov. 2022.
Mech. Appl., vol. 540, Feb. 2020, Art. no. 123174.
[7] BB Elallid, N. Benamar, AS Hafid, T. Rachidi, and N. Mrani, ''A comprehensive survey on the [30] S. Keele, ''Guidelines for performing systematic literature reviews in software engineering,''
application of deep and reinforcement learning approaches in autonomous driving,'' J. King Dept. Comput. Sci., Univ. Durham, Durham, UK, Tech. Rep. EBSE-2007-01, Ver. 2.3, 2007.
Saud Univ. Comput.
Inf. Sci., vol. 34, no. 9, pp. 7366–7390, Oct. 2022. [31] R. Nalwanga, J. Nsenga, G. Rushingabigwi, and I. Gatare, ''Design of an embedded machine
[8] S. El Hamdani and N. Benamar, ''A comprehensive study of intelligent transportation system learning based system for an environmental-friendly crop prediction using a sustainable
architectures for road congestion avoidance,'' in Ubiquitous Networking. Cham, Switzerland: soil fertility management,'' in Proc.
Springer, 2017, pp. 95–106. IEEE 19th Conf. Res. Develop. (SCOReD), Nov. 2021, pp. 251–256.

[9] E. Strubell, A. Ganesh, and A. McCallum, ''Energy and policy considerations for modern [32] SW Mohammed, NR Soora, N. Polala, and S. Saman, ''Smart water resource management

deep learning research,'' in Proc. AAAI Conf. Artif. by analyzing the soil structure and mois- ture using deep learning,'' in IoT With Smart
Intell., vol. 34, no. 9, 2020, pp. 13693–13696. Systems, J. Choudrie, P. Mahalle, T. Perumal, and A. Joshi, Eds. Singapore: Springer
Nature, 2023, pp. 709–719.
[10] D. Pau and PK Ambrose, ''Automated neural and on-device learning for micro controllers,'' in
Proc. IEEE 21st Medit. Electrotechnical Conf.
(MELECON), Jun. 2022, pp. 758–763. [33] P. Ramos-Giraldo, SC Reberg-Horton, S. Mirsky, E. Lobaton, AM Locke, E. Henriquez, A.
Zuniga, and A. Minin, ''Low-cost smart camera system for water stress detection in crops,''
[11] JR Cheng and M. Gen, ''Accelerating genetic algorithms with GPU computing: A selective
in Proc. IEEE SEN-SORS, Dec. 2020, pp. 1–4.
overview,'' Comput. Ind. Eng., vol. 128, pp. 514–525, Feb. 2019.

[34] D. Evans, ''The Internet of Things. How the next evolution of the internet is changing
[12] Y. Huang, B. Guo, and Y. Shen, ''GPU energy consumption optimization with a global-based
everything, whitepaper,'' Cisco Internet Bus. Solutions Group (IBSG), San Jose, CA, USA,
neural network method,'' IEEE Access, vol. 7, pp. 64303–64314, 2019.
Tech. Rep., 2011.
[35] SGH Soumyalatha, ''Study of IoT: Understanding IoT architecture, applications, issues and
[13] N. Bostrom and E. Yudkowsky, ''The ethics of artificial intelligence,'' in Artificial Intelligence
challenges,'' in Proc. 1st Int. Conf. Innov. Computer. NetWork. (ICICN), CSE, RRCE. Int.
Safety and Security. Boca Raton, FL, USA: Chapman & Hall/CRC, 2018, pp. 57–69.
J. Adv. Netw. Appl., vol. 478, 2016, pp. 1–5.

[14] M. Al-Rubaie and JM Chang, ''Privacy-preserving machine learning: Threats and solutions,''
[36] RT Tiburski, CR Moratelli, SF Johann, MV Neves, ED Matos, LA Amaral, and F. Hessel,
IEEE Secur. Privacy, vol. 17, no. 2, pp. 49–58, Mar. 2019.
''Lightweight security architecture based on embedded virtualization and trust mechanisms
for IoT edge devices,'' IEEE Commun. Mag., vol. 57, no. 2, pp. 67–73, Feb. 2019.
[15] H. Bamoumen, A. Temouden, N. Benamar, and Y. Chtouki, ''How TinyML can be leveraged
to solve environmental problems: A survey,'' in Proc. Int. Conf. Innov. Intel. Informat.,
[37] AS Syed, D. Sierra-Sosa, A. Kumar, and A. Elmaghraby, ''IoT in smart cities: A survey of
Comput., Technol. (ICT), Nov. 2022, pp. 338–343.
technologies, practices and challenges,'' Smart Cities, vol. 4, no. 2, pp. 429–475, Mar. 2021.

[16] A. Curry, ''The internet of animals,'' Nature, vol. 562, no. 7727, pp. 322–326, 2018.
[38] M. Ahmed, R. Mumtaz, SMH Zaidi, M. Hafeez, SAR Zaidi, and M. Ahmad, ''Distributed fog
computing for Internet of Things (IoT) based ambient data processing and analysis,''
[17] TA Wild, M. Wikelski, S. Tyndel, G. Alarcón-Nieto, BC Klump, LM Aplin, M. Meboldt, and HJ
Electronics, vol. 9, no. 11, p. 1756, Oct. 2020.
Williams, ''Internet on animals: Wi-Fi-enabled devices provide a solution for big data
transmission in biologging,'' Methods Ecology Evol., vol. 14, no. 1, pp. 87–102, Jan. 2023.
[39] E. Ahmed, A. Ahmed, I. Yaqoob, J. Shuja, A. Gani, M. Imran, and M. Shoaib, ''Bringing
computation closer toward the user network: Is edge computing the solution?'' IEEE
Commun. Mag., vol. 55, no. 11, pp. 138–144, Nov. 2017.
[18] SO Ooko, MM Ogore, J. Nsenga, and M. Zennaro, ''TinyML in Africa: Opportunities and
challenges,'' in Proc. IEEE Globecom Work-shops (GC Wkshps), Dec. 2021, pp. 1–6.
[40] S. Duan, D. Wang, J. Ren, F. Lyu, Y. Zhang, H. Wu, and X. Shen, ''Dis-tributed artificial
intelligence empowered by end-edge-cloud computing: A survey,'' IEEE Commun. Surveys
[19] H. Han and J. Siebert, ''TinyML: A systematic review and synthesis of existing research,'' in
Tuts., vol. 25, no. 1, pp. 591–624, 1st Quart., 2023.
Proc. Int. Conf. Artif. Intel. Inf. Commun. (ICAIIC), Feb. 2022, pp. 269–274.

[41] L. Greco, G. Percannella, P. Ritrovato, F. Tortorella, and M. Vento, ''Trends in IoT based
[20] V. Tsoukas, E. Boumpa, G. Giannakas, and A. Kakarountas, ''A review of machine learning solutions for health care: Moving AI to the edge,'' Pattern Recognit. Lett., vol. 135, pp. 346–
and TinyML in healthcare,'' in Proc. 25th Pan-Hellenic Conf. Informat., Nov. 2021, pp. 1–7. 353, Jul. 2020.
[42] A. Ignatov, R. Timofte, W. Chou, K. Wang, M. Wu, T. Hartley, and L. Van Gool, ''AI benchmark:
[21] S. Bangalore Lakshman and NU Eisty, ''Software engineering approaches for TinyML based Running deep neural networks on Android smartphones,'' in Proc. Eur. Conf. Comput. Vis.
IoT embedded vision: A systematic lit- erature review,'' 2022, arXiv:2204.08702. (ECCV) Workshops, Sep. 2018, pp. 288–314.

[22] R. Sanchez-Iborra and AF Skarmeta, ''TinyML-enabled frugal smart objects: Challenges and [43] VJ Reddi, B. Plancher, S. Kennedy, L. Moroney, P. Warden, L. Suzuki, A. Agarwal, C.
opportunities,'' IEEE Circuits Syst. Mag., vol. 20, no. 3, pp. 4–18, 3rd Quart., 2020. Banbury, M. Banzi, and M. Bennett, ''Widening access to applied machine learning with
TinyML,'' Harvard Data Sci. Rev., 2022.
[23] PP Ray, ''A review on TinyML: State-of-the-art and prospects,'' J. King Saud Univ. Comput. [44] V1.0 Results. [Online]. Available: https://mlcommons.org/en/inference-
Inf. Sci., vol. 34, no. 4, pp. 1595–1623, Apr. 2022. tiny-10/
[24] N. Schizas, A. Karras, C. Karras, and S. Sioutas, ''TinyML for ultra-low power AI and large [45] Dawnbench. [Online]. benchmark/ Available: https://dawn.cs.stanford.edu/
scale IoT deployments: A systematic review,'' Future Internet, vol. 14, no. 12, p. 363, Dec. CIFAR10/inference.html
2022. [46] Microcontroller Market Size and Forecast. Accessed: Feb. 5, 2023.
[25] DL Dutta and S. Bharali, ''TinyML meets IoT: A comprehensive survey,'' Internet Things, vol. [Online]. Available: https://www.verifiedmarketresearch.com/product/ microcontroller-market/
16, Dec. 2021, Art. no. 100461.
[26] M. Shafique, T. Theocharides, VJ Reddy, and B. Murmann, ''TinyML: Current progress, [47] Z. Li, J. Gao, and J. Lai, ''HBDCA: A toolchain for high-accuracy BRAM-defined CNN
research challenges, and future roadmap,'' in Proc. 58th ACM/ IEEE Design Autom. Conf. accelerator on FPGA with flexible structure,'' IEICE Trans. Inf. Syst., vol. 104, no. 10, pp.
(DAC), Dec. 2021, pp. 1303–1306. 1724–1733, 2021.

96918 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

[48] S. Hosseininoorbin, S. Layeghy, M. Sarhan, R. Jurdak, and M. Portmann, ''Exploring edge [72] R. Sanchez-Iborra, ''LPWAN and embedded machine learning as enablers for the
TPU for network intrusion detection in IoT,'' 2021, arXiv:2103.16295. next generation of wearable devices,'' Sensors, vol. 21, no. 15, p. 5218, 2021.

[49] Edge TPU—Run Inference at the Edge  |  Google Cloud. [73] K. Mekki, E. Bajic, F. Chaxel, and F. Meyer, ''A comparative study of LPWAN
[Online]. Available: https://cloud.google.com/edge-tpu/ [50] U. technologies for large-scale IoT deployment,'' ICT Exp., vol. 5, no. 1, pp. 1–7, Mar.
Thakker, J. Beu, D. Gope, C. Zhou, I. Fedorov, G. Dasika, and M. Mattina, ''Compressing 2019.
RNNs for IoT devices by 15–38x using Kro-necker products,'' 2019, [74] MI Hossain and JI Markendahl, ''Comparison of LPWAN technologies: Cost structure
arXiv:1906.02876. and scalability,'' Wireless Pers. Commun., vol. 121, no. 1, pp. 887–903, Nov. 2021.
[51] S. Zhang, Y. Wu, C. Men, and X. Li, ''Tiny YOLO optimization ori-ented bus
passenger object detection,'' Chin. J. Electron., vol. 29, no. 1, pp. 132–138, Jan. [75] RS Sinha, Y. Wei, and S.-H. Hwang, ''A survey on LPWA technology: LoRa and NB-
2020. IoT,'' ICT Exp., vol. 3, no. 1, pp. 14–21, Mar. 2017. [Online]. Available: https://
[52] B. Wu, A. Wan, F. Iandola, PH Jin, and K. Keutzer, ''SqueezeDet: Uni-fied, small, www.sciencedirect. com/science/article/pii/S2405959517300061 [76] S. Pizzi, C.
low power fully convolutional neural networks for real-time object detection for Suraci, A. Iera, A. Molinaro, and G. Araniti, ''A
autonomous driving,'' in Proc. IEEE Conf. Comput. sidelink-aided approach for secure multicast service delivery: From human-oriented
Vis. Pattern Recognition. Workshops (CVPRW), Jul. 2017, pp. 446–454. multimedia traffic to machine type communications,'' IEEE Trans.
[53] A. Womg, MJ Shafiee, F. Li, and B. Chwyl, ''Tiny SSD: A tiny single-shot detection
deep convolutional neural network for real-time embedded object detection,'' in Broadcast., vol. 67, no. 1, pp. 313–323, Mar. 2021.
Proc. 15th Conf. Comput. Robot Vis. [77] P. Andrade, I. Silva, M. Silva, T. Flores, J. Cassiano, and DG Costa, ''A TinyML soft-
(CRV), May 2018, pp. 95–101. sensor approach for low-cost detection and monitoring of vehicular emissions,''
[54] A. Kusupati, M. Singh, K. Bhatia, A. Kumar, P. Jain, and M. Varma, ''FastGRNN: A Sensors, vol. 22, no. 10, p. 3838, May 2022.
fast, accurate, stable and tiny kilobyte sized gated recur- rent neural network,'' in [78] MF Alati, G. Fortino, J. Morales, JM Cecilia, and P. Manzoni, ''Time series analysis
Proc. Adv. Neural Inf. Process. Syst., vol. 31. Red Hook, NY, USA: Curran for temperature forecasting using TinyML,'' in Proc.
Associates, 2018, pp. 1–23. [Online]. Available: https://proceedings.neurips.cc/ IEEE 19th Annu. Consum. Commun. Netw. Conf. (CCNC), Jan. 2022, pp. 691–694.
paper/2018/file/ab013ca67cf2d50796b0 c11d1b8bc95d-Paper.pdf
[79] F. Alongi, N. Ghielmetti, D. Pau, F. Terraneo, and W. Fornaciari, ''Tiny neural
[55] M. córdova, A. Pinto, H. Pedrini, and R. d. S. Torres, ''Pelee-Text++: A tiny neural networks for environmental predictions: An integrated approach with miosix,'' in
network for scene text detection,'' IEEE Access, vol. 8, pp. 223172–223188, 2020. Proc. IEEE Int. Conf. Smart Computer. (SMARTCOMP), Sep. 2020, pp. 350–355.

[56] C. Zhang, H. Wang, J. Zeng, L. Ma, and L. Guan, ''Tiny-RainNet : A deep convolutional [80] A. Omambia, B. Maake, and A. Wambua, ''Water quality monitoring using IoT &
neural network with bi-directional long short-term memory model for short-term machine learning,'' in Proc. IST-Africa Conf. (IST-Africa), May 2022, pp. 1–8.
rainfall prediction,'' Meteorological Appl., vol. 27, no. 5, p. e1956, Sep. 2020.
[81] NNM Rizal, G. Hayder, and KA Yusof, ''Water quality predictive analysis using an
[57] NP Ghanathe and S. Wilton, ''T-RECX: Tiny-resource efficient convo-lutional neural artificial neural network with a graphical user inter-face,'' Water, vol. 14, no. 8, p.
networks with early-eXit,'' 2022, arXiv:2207.06613. 1221, Apr. 2022.

[58] K. Xu, Y. Li, H. Zhang, R. Lai, and L. Gu, ''EtinyNet: Extremely tiny network for [82] D. Loukatos, K.-A. Lygkoura, C. Maraveas, and KG Arvanitis, ''Enrich-ing IoT modules with

TinyML,'' in Proc. AAAI Conf. Artif. Intell., vol. 36, no. 4, 2022, pp. 4628–4636. edge AI functionality to detect water misuse events in a decentralized manner,'' Sensors,
vol. 22, no. 13, p. 4874, Jun. 2022.

[59] TM Khan, A. Robles-Kelly, and SS Naqvi, ''T-Net: A resource-constrained tiny

convolutional neural network for medical image segmentation,'' in Proc. IEEE/ [83] MM Ogore, K. Nkurikiyeyezu, and J. Nsenga, ''Offline prediction of cholera in rural
CVF Winter Conf. Appl. Comput. Vis. (WACV), Jan. 2022, pp. 1799–1808. communal tap waters using edge AI inference,'' in Proc.
IEEE Globecom Workshops (GC Wkshps), Dec. 2021, pp. 1–6.
[84] N. Naeem, T. Boroushaki, and W. Chen, Efficient Ultra Low Power
[60] H. Liu, K. Simonyan, and Y. Yang, ''DARTS: Differentiable architecture search,''
2018, arXiv:1806.09055. Underwater Imaging.
[85] G. Datta, Z. Yin, A. Jacob, AR Jaiswal, and PA Beerel, ''Toward efficient hyperspectral
[61] Q. Lu, W. Jiang, M. Jiang, J. Hu, S. Dasgupta, and Y. Shi, ''FGNAS: FPGA-aware
image processing inside camera pixels,'' Tech.
graph neural architecture search,'' Tech. Rep., 2020.
Rep., 2022.
[62] S. Soro, ''TinyML for ubiquitous edge AI,'' 2021, arXiv:2102.01255.
[86] T. Clements, ''Earthquake detection with TinyML,'' in Proc. AGU Fall Meeting Abstr.,
[63] TensorFlow. (2022). Tensorflow Lite Guide. [Online]. Available:
Dec. 2021, pp. S15A–0229.
https://www.tensorflow.org/lite/guide
[87] D. Tuia, B. Kellenberger, S. Beery, BR Costelloe, S. Zuffi, B. Risse, A. Mathis, MW
[64] (2022). Tensorflow Lite Micro. [Online]. Available: https://www.
Mathis, F. van Langevelde, and T. Burghardt, ''Perspec-tives in machine learning
tensorflow.org/lite/microcontrollers
for wildlife conservation,'' Nature Commun., vol. 13, no. 1, p. 792, 2022.
[65] Edge Impulse. (2022). Getting Started. [Online]. Available: https://docs.
edgeimpulse.com/docs/
[88] DJ Curnick, AJ Davies, C. Duncan, R. Freeman, DMP Jacoby, HTE Shelley, C.
[66] D. Situnayake. (2021). How Tensorflow Helps Edge Impulse Make ML Accessible to Rossi, OR Wearn, MJ Williamson, and N.
Embedded Engineers. [Online]. Available: https://blog.tensorflow.org/2021/06/how- Pettorelli, ''SmallSats: A new technological frontier in ecology and conservation?''
tensorflow-helps-edge-impulse-make-ml-accessible.html [67] J. Remote Sens. Ecology Conservation, vol. 8, no. 2, pp. 139–150, Apr. 2022.
Jongboom. (2020). Introducing EON:
Neural Networks in up to 55% Less Ram and 35% Less Rom. https://www.edgeimpulse. [89] Hackster IO. (2020). ElephantEdge—Hackster.Io. Accessed: Jan.
com/blog/introducing-eon [68] MKL Moreau. (2020). Announcing Fomo (Faster 13, 2023. [Online]. Available: https://www.hackster.io/ contests/
Objects, More Objects). ElephantEdge
[Online]. Available: https://www.edgeimpulse. com/blog/announcing-fomo-faste [69] J. [90] Manivannan. (2020). Elephant Guardian—TinyMLs to Save the Giant Wild Animals.
Petajajarvi, K. Mikhaylov, A. Roivainen, T. Hanninen, and M. Pettissalo, ''On the Accessed: Jan. 13, 2023. [Online]. Available: https://www.hackster.io/manivannan/
coverage of LPWANs: Range elephant-guardian-tinymls-to-save-the-giant-wild-animal-372306 [91] S. Gallacher,
evaluation and channel attenuation model for LoRa technology,'' in Proc. 14th Int. D. Wilson, A. Fairbrass, D.
Turmukhambetov, M. Firman, S. Kreitmayer, OM Aodha, G. Brostow, and K. Jones,
''Shazam for bats: Internet of Things for continuous real-time biodiversity
Conf. Telecommunication. (ITST), Dec. 2015, pp. 55–59. monitoring,'' IET Smart Cities, vol. 3, no. 3, pp. 171–183, Sep. 2021.
[70] H. Xu, W. Yu, D. Griffith, and N. Golmie, ''A survey on industrial Internet of Things: A
cyber-physical systems perspective,'' IEEE Access, vol. 6, pp. 78238–78259, 2018. [92] N. Karges, J. Staab, J. Rauh, M. Wegmann, and H. Taubenböck, ''Sound-scapes
on edge-the real-time machine learning approach for measuring soundscapes on
[71] What is LoRawan®. [Online]. Available: https://resources.lora- resource-constrained devices,'' in Proc. 24th Int. Congr.
alliance.org/document/what-is-lorawan Acoust., 2022, pp. 128–139.

VOLUME 11, 2023 96919

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

[93] MS Diab and E. Rodriguez-Villegas, ''Embedded machine learning using microcontrollers in [113] G. Singh, A. Singh, and G. Kaur, ''Role of artificial intelligence and the Internet of Things in
wearable and ambulatory systems for health and care applications: A review,'' IEEE Access, agriculture,'' in Artificial Intelligence to Solve Per-vasive Internet of Things Issues.
vol. 10, pp. 98450–98474, 2022. Amsterdam, The Netherlands: Elsevier, 2021, pp. 317–330.

[94] C. Nicolas, B. Naila, and R.-C. Amar, ''TinyML smart sensor for energy saving in Internet of [114] VK Quy, NV Hau, DV Anh, NM Quy, NT Ban, S. Lanza, G. Randazzo, and A. Muzirafuti, ''IoT-
Things precision agriculture platform,'' in Proc. 13th Int. Conf. Ubiquitous Future Netw. enabled smart agriculture: Archi-tecture, applications, and challenges,'' Appl. Sci., vol. 12,
(ICUFN), Jul. 2022, pp. 256–259. no. 7, p. 3396, Mar. 2022.
[95] K. Ahmed and M. Hassan, ''TinyCare: A TinyML-based low-cost contin-uous blood pressure
estimation on the extreme edge,'' in Proc. IEEE 10th Int. Conf. Healthcare Information. [115] D. Sheth, B. Sudharsan, JG Breslin, and MI Ali, ''Embedded ML pipeline for precision
(ICHI), Jun. 2022, pp. 264–275. agriculture,'' in Proc. 21st ACM/ IEEE Int. Conf.
[96] D. Takeuchi, K. Yatabe, Y. Koizumi, Y. Oikawa, and N. Harada, ''Real-time speech Inf. Process. Sensor Netw. (IPSN), May 2022, pp. 527–528.
enhancement using equilibrated RNN,'' in Proc. IEEE Int. [116] L. Falaschetti, L. Manoni, RCF Rivera, D. Pau, G. Romanazzi, O. Silvestroni, V. Tomaselli,
Conf. Acoust., Speech Signal Process. (ICASSP), May 2020, pp. 851–855. and C. Turchetti, ''A low-cost, low-power and real-time image detector for grape leaf esca
[97] I. Fedorov, M. Stamenovic, C. Jensen, L.-C. Yang, A. Mandell, Y. Gan, M. Mattina, and P.N. disease based on a compressed CNN,'' IEEE J. Emerg. Sel. Topics Circuits Syst., vol. 11,
Whatmough, ''TinyLSTMS: Efficient neural speech enhancement for hearing aids,'' Cornell no. 3, pp. 468–481, Sep. 2021.
Univ., Ithaca, NY, USA, Tech. Rep., 2020.
[117] M. Alessandrini, R. Calero Fuentes Rivera, L. Falaschetti, D. Pau, V. Tomaselli, and C.
[98] F. Sabry, T. Eltaras, W. Labda, K. Alzoubi, and Q. Malluhi, ''Machine learning for healthcare Turchetti, ''Avine grape leaves dataset for early detection and classification of esca disease
wearable devices: The big picture,'' J. Healthcare Eng., vol. 2022, pp. 1–25, Apr. 2022. in vineyards through machine learning,'' Data Brief, vol. 35, Apr. 2021, Art. no. 106809.

[99] H. Gokul, P. Suresh, BH Vignesh, RP Kumaar, and V. Vijayaraghavan, ''Gait recovery system [118] F. de Vita, G. Nocera, D. Bruneo, V. Tomaselli, D. Giacalone, and SK Das, ''Quantitative
for Parkinson's disease using machine learning-ing on embedded platforms,'' in Proc. analysis of deep leaves: A plant disease detector on the smart edge,'' in Proc. IEEE Int.
IEEE Int. Syst. Conf. (SysCon), Aug. 2020, pp. 1–8. Conf. Smart Computer. (SMART-COMP), Sep. 2020, pp. 49–56.

[100] A. Faraone and R. Delgado-Gonzalo, ''Convolutional-recurrent neural networks on low-power [119] G. Esgario. (2019). Coffee Dataset. Accessed: Dec. 20, 2022. [Online].
wearable platforms for cardiac arrhythmia detection,'' in Proc. 2nd IEEE Int. Conf. Artif. Available: https://drive.google.com/file/d/15YHebAGrx1Vhv8-naave-R5o3Uo70jsm
Intel. Circuits Syst. (AICAS), Aug. 2020, pp. 153–157.
[120] V. Gutti and R. Karthi, ''Real time classification of fruits and vegetables deployed on low power
[101] R. Zanetti, A. Aminifar, and D. Atienza, ''Robust epileptic seizure detection on wearable embedded devices using tiny ML,'' in Proc. Int.
systems with reduced false-alarm rate,'' in Proc. 42nd Annu. Int. Conf. IEEE Eng. Med. Conf. ImageProcess. Capsule Netw. Cham, Switzerland: Springer, 2022, pp. 347–359.
Biol. Soc. (EMBC), Jul. 2020, pp. 4248–4251.
[121] M. Oltean. (2017). Fruits 360 Dataset: A Dataset of Images Containing Fruits and Vegetables.
[102] CD Caleanu, CL Sîrbu, and G. Simion, ''Deep neural architectures for contrast enhanced Accessed: Dec. 20, 2022. [Online]. Available: https://www.kaggle.com/datasets/moltean/
ultrasound (CEUS) focal liver lesions automated diagnosis,'' in Proc. Int. Symp. Electrons. fruits [122] GJ Miguel. (2021). Crop Recommendation Using
Telecommunication. (ISETC), 2021, pp. 1–9. Machine Learning. Accessed: Dec. 20, 2022. [Online]. Available: https://github.com/gabbygab1233/
Crop-Recommender [123] C. Bruno, A. Licciardello, GAM Nastasi, F. Passaniti, C. Brigante,
[103] TM Ingolfsson, A. Cossettini, X. Wang, E. Tabanelli, G. Tagliavini, P. Ryvlin, L. Benini, and S. F. Sudano, A. Faulisi, and E. Alessi, ''Embedded artificial intelligence
Benatti, ''Towards long-term non-invasive monitoring for epilepsy via wearable EEG approach for gas recognition in smart agriculture applications using low cost MOX gas sensors,'' in
devices,'' in Proc. IEEE Biomed. Circuits Syst. Conf. (BioCAS), Oct. 2021, pp. 1–4. Proc. Smart Syst. Integr. (SSI), Apr. 2021, pp. 1–5.

[104] D. Ayata, Y. Yaslan, and M.E. Kamasak, ''Emotion recognition from multimodal physiological
signals for emotion aware healthcare systems,'' J. Med. Biol. Eng., vol. 40, pp. 149–157,
Jan. 2020. [124] I. Ihoume, R. Tadili, N. Arbaoui, M. Benchrifa, A. Idrissi, and M. Daoudi, ''Developing a multi-
[105] R. Laureanti, M. Bilucaglia, M. Zito, R. Circi, A. Fici, F. Rivetti, R. Valesi, C. Oldrini, L.T. label TinyML machine learning model for an active and optimized greenhouse microclimate
Mainardi, and V. Russo, ''Emotion assessment using machine learning and low-cost control from multivariate sensed data,'' Artif. Intel. Agricult., vol. 6, pp. 129–137, Jan. 2022.
wearable devices,'' in Proc. 42nd Annu.
Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Jul. 2020, pp. 576–579. [125] R. Sanchez-Iborra, A. Zoubir, A. Hamdouchi, A. Idri, and A. Skarmeta, ''Intelligent and efficient
[106] MS Fernandes, W. Cordeiro, and M. Recamonde-Mendoza, ''Detecting aedes aegypti IoT through the cooperation of TinyML and edge computing,'' Informatica, vol. 34, no. 1,
mosquitoes through audio classification with convolu-tional neural networks,'' Comput. pp. 147–168, 2023.
Biol. Med., vol. 129, Feb. 2021, Art. no. 104152. [126] DM Matilla, AL Murciego, DM Jiménez-Bravo, AS Mendes, and VR Leithardt, ''Low-cost edge
computing devices and novel user inter-faces for monitoring pivot irrigation systems based
[107] K. Trivedi and H. Shroff, ''Identification of deadliest mosquitoes using wing beats sound on Internet of Things and LoRaWAN technologies,'' Biosystems Eng., vol. 223, pp. 14–29,
classification on tiny embedded system using machine learning and edge impulse platform,'' Nov. 2022.
in Proc. ITU Kaleidoscope, Connecting Phys. Virtual Worlds (ITUK), 2021, pp. 1–6.
[127] P. Du, T. Polonelli, M. Magno, and Z. Cheng, ''Towards lightweight deep neural network for
[108] N. Abdennadher, D. Pau, and A. Bruna, ''Fixed complexity tiny reser-voir heterogeneous smart agriculture on embedded systems,'' in Proc.
network for on-line ECG learning of anomalies,'' in Proc. IEEE 10th Global Conf. Consum. IEEE Sensors Appl. Symp. (SAS), Aug. 2022, pp. 1–6.
Electrons. (GCCE), Oct. 2021, pp. 233–237. [128] Plant Village. Accessed: Feb. 5, 2023. [Online]. Available: https://plantvillage.psu.edu/

[109] United Nation. (2017). World Population Projected to Reach 9.8 Billion in 2050, and 11.2 [129] J. Manokaran and G. Vairavel, ''Smart anomaly detection using data-driven techniques in IoT
Billion in 2100. Accessed: Dec. 20, 2022. [Online]. Available: https://www.un.org/en/desa/ edge: A survey,'' in Proc. 3rd Int. Conf.
world-population-projected-reach-98- billion-2050-and-112-billion-2100 Commun., Comput. Electrons. Syst. Cham, Switzerland: Springer, 2022, pp. 685–702.

[110] A. Mitra, SLT Vangipuram, AK Bapatla, VKVV Bathalapalli, SP Mohanty, E. Kougianos, and [130] A. Chatterjee and BS Ahmed, ''IoT anomaly detection methods and applications: A survey,''
C. Ray, ''Everything you wanted to know about smart agriculture,'' 2022, arXiv:2201.04754. Internet Things, vol. 19, Aug. 2022, Art. no. 100568.

[111] S. Condran, M. Bewong, MZ Islam, L. Maphosa, and L. Zheng, ''Machine learning in precision [131] AB Nassif, MA Talib, Q. Nasir, and FM Dakalbab, ''Machine learning for anomaly detection: A
agriculture: A survey on trends, appli- cations and evaluations over two decades,'' IEEE systematic review,'' IEEE Access, vol. 9, pp. 78658–78700, 2021.
Access, vol. 10, pp. 73786–73803, 2022.
[132] A. Xenakis, A. Karageorgos, E. Lallas, AE Chis, and H. González-Vélez, ''Towards distributed
[112] Y. Kalyani and R. Collier, ''A systematic survey on the role of cloud, fog, and edge computing IoT/Cloud based fault detection and maintenance in industrial automation,'' Proc. Comput.
combination in smart agriculture,'' Sensors, vol. 21, no. 17, p. 5922, Sep. 2021. Sci., vol. 151, pp. 683–690, Jan. 2019.

96920 VOLUME 11, 2023

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

[133] A. Mostafavi and A. Sadighi, ''A novel online machine learning approach for HATIM BAMOUMEN (Student Member, IEEE) is currently a
real-time condition monitoring of rotating machines,'' in Proc. 9th RSI Int. highly motivated student with Al Akhawayn University, Ifrane.
Conf. Robots. Mechatronics (ICRoM), Nov. 2021, pp. 267–273. He has a passion for cutting-edge technologies and has
[134] M. Antonini, M. Pincheira, M. Vecchio, and F. Antonelli, ''A TinyML approach demonstrated a strong commitment to research and innovation.
to non-repudiable anomaly detection in extreme industrial environments,''
in Proc. IEEE Int. Workshop Metrology Ind. 4.0 IoT (MetroInd4.0IoT), Jun. As a member of several research projects, he has contributed
2022, pp. 397–402. to the advancement of the field of TinyML and its applications
[135] M. Lord and A. Kaplan, ''Mechanical anomaly detection on an embedded
in various sectors, typically in the environment. As he
microcontroller,'' in Proc. Int. Conf. Comput. Sci. Comput. Intel. (CSCI),
elucidated the advantages of applying TinyML for sustainable
Dec. 2021, pp. 562–568.
environmental development. He has also served as a member
[136] VM Oliveira and AH Moreira, ''Edge AI system using a thermal camera for
for the IEEE Student Branch and has earned accredited certifications as a peer tutor in several
industrial anomaly detection,'' in International Summit Smart City 360ÿ
STEM subjects, including applied mechanics and electrical engineer-ing. His research interests
. Cham, Switzerland: Springer, 2022, pp. 172–187.
[137] D. Pau, A. Khiari, and D. Denaro, ''Online learning on tiny micro-controllers include beyond engineering and into the realm of economics, where he has conducted

for anomaly detection in water distribution systems,'' in Proc. IEEE 11th Int. numerous research analysis projects.
Conf. Consum. Electrons. (ICCE-Berlin), Nov. 2021, pp. 1–6.
One notable project involved the application of machine learning models to predict African
[138] M. Cardoni, DP Pau, L. Falaschetti, C. Turchetti, and M. Lattuada, ''Online economies, highlighting his multidisciplinary approach to problem-solving. With his diverse skill
learning of oil leak anomalies in wind turbines with block-based binary set and passion for innovation, he is poised to make a significant impact in the field of
reservoir,'' Electronics, vol. 10, no. 22, p. 2836, Nov. 2021. engineering and beyond.
[139] P. Andrade, I. Silva, G. Signoretti, M. Silva, J. Dias, L. Marques, and DG
Costa, ''An unsupervised TinyML approach applied for pave-ment
anomalies detection under the Internet of Intelligent vehicles,'' in Proc.
IEEE Int. Workshop Metrology for Ind. 4.0 IoT (MetroInd4.0IoT), Jun. 2021,
pp. 642–647.
[140] H. Ren, D. Anicic, and TA Runkler, ''TinyOL: TinyML with online-learning on
NABIL BENAMAR received the master's and Ph.D. degrees
microcontrollers,'' in Proc. Int. Joint Conf. Neural Netw.
(IJCNN), Jul. 2021, pp. 1–8. from Moulay Ismail University, Meknes, Morocco, in 2001 and
[141] State of the Tinyautoml Market 2022, TinyML Found., Altos, CA, USA, 2004, respectively.
2022. He is currently a Professor in computer science with the
[142] K. Kopparapu, E. Lin, JG Breslin, and B. Sudharsan, ''TinyFedTL: Federated School of Technology, Moulay Ismail University, and an
transfer learning on ubiquitous tiny IoT devices,'' in Proc. Adjunct Faculty Member of Computer Science with Al
IEEE Int. Conf. Pervasive Comput. Commun. Workshops Affiliated Events Akhawayn University, Ifrane, Morocco. He is the author of
(PerCom Workshops), Mar. 2022, pp. 79–81. several jour-nal articles and IETF standard documents. His
[143] M. Shafique, A. Marchisio, RVW Putra, and MA Hanif, ''Towards energy-
efficient and secure edge AI: A cross-layer framework,'' in Proc. 40th IEEE/ research interests include future generation net-works,
ACM Int. Conf. Comput.-Aided Design (ICCAD). Ithaca, NY, USA: Cornell
autonomous driving, the IoT, and TinyML. He is a member of the Tiny Machine Learning Open
University, 2021, pp. 1–9.
Education Initiative (TinyMLedu). He is also serving as an Associate Editor for the IEEE
ACCESS journal and the Journal of King Saud University—Computer and Information Sciences
(IF 8.8). He is a TPC Member of highly ranked IEEE Flagship Conferences (GLOBECOM, ICC,
PIMRC, and WCNC). He served as the Chair for IEEE MenaComm'20 Conference and a
member of the Organizing Committee for IEEE WCNC'2019 and IWCMC'23. He is an expert
in internet governance and he was an ISOC Ambassador to IGF (2012 and 2013), Google
panelist in the first Arab-IGF, an ISOC Fellow to IETF'89&92&95&99&103, and an ICANN'50&54
Fellow. Among his international commitments, he is currently serving as the Chair for the Task
YOUSSEF ABADADE received the M.Sc. degree in information Force for Arabic Script IDNs, a team of people working on the implementation of the Arabic
systems engineering from the Fac-ulty of Sciences Semlalia,
script in the DNS root zone. He is also chairing the UASG measurement WG promoting the
Cadi Ayyad University, Marrakesh, Morocco, in 2015. He is
universal acceptance of all valid domain names and e-mail addresses.
currently pursuing the Ph.D. degree in TinyML with the Sys-
tem Engineering Laboratory, National School of Applied
Sciences, Ibn Tofail University, Kenitra, Morocco. He is also
the Technical Lead of Java with KARAVEL and the Leader of
Tourism Market in Paris, France.

YOUSRA CHTOUKI received the bachelor's degree in

computer science from UCO, USA, in 2001, the master's
degree in database man-agement from OCU, USA, in 2005,
and the Ph.D. degree (Hons.) in web services applications in
e-Learning from Ecole Mohammadia, Rabat, in 2017. She has
ANAS TEMOUDEN is currently pursuing the degree in artificial been a full-time Professor and a Researcher with Al Akhawayn
University (AUI), since 2008. She has coauthored two books
intelligence and robotization with Al Akhawayn University,
Ifrane, Morocco. and contributed to one book on technology in edu-cation
He has been interested in research since his first year of related topics. She has published multiple journals and

university and he has published his first paper at the age of conference papers in well ranked journals. She is currently
18. His research interests include artificial intelligence, the Founder of the EFP Educational Farm Project with AUI, where she is also an Advisor of
embedded systems in general, and tiny machine learning in the first sustainability club during which time the club earned the PRME winner 2022 Award for
particular. promoting sustainability in education. She created many

VOLUME 11, 2023 96921

Machine Translated by Google
Y. Abadade et al.: Comprehensive Survey on TinyML

courses with the Computer Science Department implementing different peda-gogies related ABDELHAKIM SENHAJI HAFID received the MS and Ph.D.
to computer science education. She is also a Main Contributor to the launch of the degrees in computer science.
undergraduate research program with AUI through which undergraduate students published He spent several years as a Senior Research Scientist with
papers specific to data analysis, machine learning, and TinyML. She has been an active Bell Communications Research
member of the EduSummIT UNESCO sponsored research group, since 2013; a member of (Bellcore), Piscataway, NJ, USA, working in the context of
the AI Committee with AUI; and a member of the TinyML Foundation. She is also working on major research projects on the management of next-
the implementation of the first living laboratory with AUI, a multidisciplinary research space to generation networks. He was also an Assistant Professor with
learn and innovate mainly aimed at making use of technologies, such as TinyML to solve Western University (WU), London, ON, Canada, the Research
environment and agriculture related topics. She has developed multiple theories related to Director of the Advance Communication Engineering Center
teaching computer science and programming courses such as ''Content: Less is More'' ''Pace (venture established by WU, Bell Canada, and Bay Networks),
Management in Introductory Programming.'' London, a Researcher with CRIM, Montreal, QC, Canada, a Visiting Sci-entist with GMD-
Fokus, Berlin, Germany, and a Visiting Professor with the University of Évry, Évry-
Courcouronnes, France. He is currently a Full Professor with the University of Montreal,
Montreal. He is also the Founding Director of the Network Research Laboratory, Montreal, QC,
Canada, and the Montreal Blockchain Laboratory, Montreal. He is also a Research Fellow with
CIRRELT, Montreal. He co-founded Tipot Technologies Inc., Ottawa, ON, Canada (research
and development platform for IoT). He consulted for a number of telecommunication companies
and startups in North America.

He has extensive academic and industrial research experience in the area of the management
and design of next-generation networks. He has supervised to graduate over 50 graduate and
postgraduate students. He has authored or coauthored over 250 journals and conference
papers. He also holds three US patents. His current research interests include the IoT, fog/
edge computing, blockchain, and intelligent transport systems. He also gave talks/keynotes at
a number of international conferences.

96922 VOLUME 11, 2023