IEEE Circuits and Systems Magazine - Q3 2020 - 6

opened by this emerging hot-topic. A dissection of the
potential services that will benefit from the introduction
of ML-powered mechanisms into frugal objects is also
presented. Currently available frameworks that permit to embed ML schemes into MCUs are explored and
reviewed. Although some performance evaluations of
certain frameworks have been reported [15], [16], to the
authors' knowledge there is not any prior work providing a comprehensive survey focused on these valuable
tools. Besides, we elaborate on a specific use case which
is the intelligent communication interface selection in a
constrained multi-Radio Access Technology (RAT) device, by considering the system status as well the characteristics of the data to be sent. To this end, an architectural design of the proposed frugal object and a real
on-device implementation of a ML-based decision-making mechanism are finally showed. Therefore, the main
contributions of this work are the following:
■■ A wide discussion of the TinyML concept as well
as the range of potentially benefited applications.
■■ A survey of the principal available frameworks for
integrating ML algorithms within MCUs.
■■ Insights regarding the real implementation of
ML-based schemes over frugal objects to solve
the problem of intelligent interface selection in a
multi-RAT set up.
The rest of the paper is organized as follows. Section II provides a comprehensive overview of the TinyML
landscape. Section III presents a survey of the available
frameworks for integrating ML within frugal objects.
Section IV focuses on the specific use case of devices
with multiple communication interfaces, presenting the
proposed architecture, the implementation process of
Processing Power
High
Simplicity

Medium

Memory

Low
Low-Cost

Storage

Multi-RAT

Portability
Energy Efficiency

Microcontroller-Based Device
Microprocessor-Based Device
Figure 2. Feature comparison between MCU and microprocessor-powered devices.

ML algorithms within a MCU powered device, and the
attained results. Finally, Section V summarizes the most
important findings and draws future research lines.
II. Machine Learning in Frugal Objects
The society digitalization achieved during recent times
has democratized the daily use of a plethora of frugal
objects, e.g., wearables, environmental data collectors,
actuators, etc. The integration of intelligence within
these small devices brings a series of advantages
that outweigh the possible cons. In the following, we discuss these aspects aiming at exploring the possibilities
opened by TinyML from a realistic perspective.
A. Benefits
1) Energy Efficiency
This is the first positive point that comes to mind when adopting MCU-based solutions. Whereas powerful processors and
Graphics Processing Units (GPUs) demand great amount of
power, MCUs present a reduced energy consumption, even
working at their maximum workload level. This permits
MCUs to rely on batteries or even energy harvesting.
Therefore, smart frugal objects can be placed almost everywhere without the need of being plugged to the power grid,
which opens the door for novel cognitive itinerant applications [17]. Furthermore, having a scarce power consumption
allows the smart unit to be coupled with larger battery-powered devices, hence converting them into connected smart
entities, e.g., personal mobility devices such as scooters or
segways. Fig. 2 compares the main characteristics of devices
powered either by microcontroller or microprocessor.
2) Low Cost
The simplicity and hardware constraints of frugal objects
leads to a moderate cost per unit. Clearly, this has been
one of the main reasons for the quick and fruitful exploitation of IoT architectures in a large range of segments:
industry, agriculture, e-health, or entertainment, among
others [18]. These deployments consist of a large number
of frugal objects specialized on performing varied tasks.
The heterogeneity of employed end-devices, as well as
their reprogrammability bring endless opportunities for
enriching current deployments with intelligence, hence
enabling their evolution into smart and flexible systems.
3) System Reliability and Data Security
As stated above, wireless transmissions require much
energy in comparison with processing tasks. This
-involves a real need for limiting communication activities in frugal objects. Although this could be seen as a
negative aspect, it actually allows to increase system's
reliability as well as data security and privacy. Big data

IEEE CIRCUITS AND SYSTEMS MAGAZINE

THIRD QUARTER 2020

IEEE Circuits and Systems Magazine - Q3 2020

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2020