IEEE Circuits and Systems Magazine - Q3 2020 - 7

applications require to stream raw data over unpredictable and lossy wireless channels from end-devices to the
cloud. This strategy is bandwidth consuming and prone
to transmission errors and cyber-attacks, e.g., man-inthe-middle or eavesdropping, so that the transmitted
data may be lost or compromised. These issues are
avoided by performing on-device data processing which
permits to perform fewer transmissions with aggregated
or meaningless data for an attacker.
4) Latency
This is another advantage of performing local processing instead of offloading computing tasks onto the
cloud. State-of-the-art IoT RATs, e.g., Low Power Wide
Area Network (LPWAN) [19], have notably improved
energy efficiency and transmission distances at the
expense of reducing data-rate. This fact leads to very
long message's times-on-air that add undesired delay
to the whole operation. Besides, on-device computation
efficiency of complex ML algorithms is being deeply investigated in order to achieve real-time execution and
notable advances are being achieved [20].
B. Challenges
Given the discussed advantages, integrating ML within
frugal objects also presents important challenges to be
overcome. These are mostly related to device heterogeneity and MCU's constraints.
1) Device Heterogeneity
Regarding this aspect, the enormous range of existing
devices is a notable obstacle for having a general set
of TinyML tools that could be valid for most of MCUbased set-ups. These units may present different power
consumption, processing capabilities, memory, storage
capacity, embedded RATs, peripherals, communication
ports and protocols, etc., which also hinders the design
of generic benchmarking methodologies in TinyML systems [21]. Therefore, the implementation of universal
development and benchmarking frameworks coping
with this heterogeneity is not an easy task although it is
highly necessary for increasing awareness and adoption.
2) MCU Constraints
So far computation and memory constraints have not been
a limitation given the quantity and quality of available resources in modern workstations and data centers. However, up-scaling computing system's resources according to
the processing needs of ML mechanisms is not a valid approach for MCUs [22]. For that reason, TinyML ecosystem
players should coordinate effective efforts for achieving a
functional convergence between ML and frugal objects. In
this line, MCU manufacturers should ensure that their chipTHIRD QUARTER 2020

sets are suitable for straight-through ML integration, which
will permit an easy application development with the support of device-level software. In this regard, it is important
that future firmwares could reach high reliability levels and
tackle key aspects such as in-device data security and optimized transmission operations. In addition, cloud players
should develop enhanced platforms and APIs to allow efficient data exchanges between objects and cloud. Finally,
the adaptation of mainstream data-science frameworks
to MCU requirements is crucial for a wide expansion of
-TinyML. Not only programming libraries but also lifecycle
management tools are needed to simplify solution development, data handling, etc. As discussed in Section III, a first
wave of TinyML frameworks and developing tools is already
available, which will foster the arrival of novel services and
applications to the market by attracting new players.
C. Applications
Many IoT applications are already integrated in our daily
personal and professional activities, although many of
them are prone to be strengthened with local smart processing capabilities (Fig. 3). Most of current IoT devices
make an extensive use of network-related operations,
which is not always the best option as discussed previously. For instance, wearables such as fitness monitors or
"smart" watches are not that smart, as they rely on a Bluetooth-coupled smartphone which assumes the majority
of processing tasks. Integrating ML within frugal objects
permits their independence from a master device, which
will notably enrich current applications landscape.
1) eHealth
This is one of the segments with greater growth potential.
Health monitoring, visual assistance, hearing aids, personal sensing, activity recognition, etc., are just few examples of applications that will exploit truly-smart enddevices. Beyond the health-care perspective, augmented
reality, real-time voice recognition, language translation, context-aware support systems, or precise indoor/
outdoor positioning are other wearable-based services
to come in a near future [23].
2) Smart Spaces
ML-enabled frugal objects will also contribute to the
collective intelligence already present in scenarios such
as smart cities or cognitive buildings, among many
others [24]. Current IoT-based monitoring and surveillance systems, e.g., traffic, pollution, crowdsensing, etc.,
will evolve to smart and autonomous entities capable of
making decentralized and quick decisions. The deployment simplicity and independence from the power grid allow things to be placed in remote and rural areas, hence
creating smart spaces with seamless roaming among them
IEEE CIRCUITS AND SYSTEMS MAGAZINE

IEEE Circuits and Systems Magazine - Q3 2020

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