-
concept that has been introduced in 2017 (the
corresponding algorithm already in 2016) is to
parallelize computing for machine learning in
order to decrease the latency [
4
].
This development can be put into the context of
decentralized autonomous organizations [
5
] since
it is comparable to the fundamental idea of
crossfunctional working in a BPM (Business Process
Management) context. The often mentioned prime
example is the parallelized machine learning for
image recognition on smart phone agents [
4
]. This
short paper discusses major topics of centralized
vs. decentralized AI in more detail and displays
further application areas.
There are several accompanying developments
that encourage decentralization: for example
“softwarization” and “disappearing internet of
things” [
2
]. The idea is roughly that devices are
more becoming like terminals as end-nodes of a
network, thus spanning a seamless continuum
from the devices to the cloud [
2
]. Here in
particular a lot of computation already takes place
at the devices, so that computation is moving
“toward the edge”, hence enforcing Edge and Fog
Computing (here Fog Computing covers the gap
between Cloud and Edge Computing) [
2
]. As
simple computations can typically evolve into
machine learning routines, there seems to be a
similar development for decentralized AI
solutions in a broader context (AI at the edge
instead of simple computations at the edge) and
not only in the aforementioned image recognition
example. So the decentralized devices can not only
be seen as contributors to a consolidated learning
procedure, but in fact learning for their own
purpose as well as exchanging information and/or
synchronizing when it seems useful. In short, AI is
no longer purely centralized and also no longer
purely a result of one single participant in learning.
So, there are two distinct conceptual setups: either
several agents contribute to a consolidated
learning in a hierarchical structure without
substantially learning for themselves (for example
when they only perform a limited number of
iterations [
6
]) or several agents synchronize on the
same hierarchy level meaning that they produce an
aggregated model that is shared between all of
them.
2
Relevant technological aspects
There are several relevant (technological) aspects
with accompanying challenges that have to be
understood and mastered in order to decide on the
parameters of central/decentral AI approaches in a
meaningful manner:
1. Speed: As already mentioned, multiple agents
can be regarded as a kind of parallelized
computing mechanism in order to use
available time windows more efficiently. This
objective is in particular achieved via
delegating subtasks in a hierarchical setup to
various agents hence improving “virtually” the
computing power. Furthermore, as the
computational procedures are performed at the
local clients at the edge (in the sense of Edge
Computing), several data transmissions and
information exchanges with the cloud become
obsolete [
6
]. In summary, speed is not only
increased by higher computational power due
to the number of participating devices, but also
by reduction of now unnecessary back-forth
transmissions of the training data which is
stored locally [
6
]. Additionally, a reduction of
(where applicable) communication costs can
occur [
6
]. So under business process
considerations, federated learning can
constitute one method to avoid bottlenecks and
speed up critical process steps.
2. Quality: Apart from “just” parallelizing tasks
for ML procedures, where the volume of the
data that has to be processed is very large, the
learning of several agents can be used to
improve the overall quality of the respective
ML routine [
6
]. Typically, this is done via
ensemble methods of machine learning (for
instance via “bagging” or “boosting” to name
a few examples) [
7
]. One might question
whether quality is strictly improved by
consolidating the inputs of various learning
agents (as they might provide misleading
adjustments). The quality improvement cannot
be strictly guaranteed and is dependent on the
method used for aggregation, but typically,
more input from agents should generate the
potential for superior quality in learnings [
7
].
Bagging and boosting may be the most widely
known procedures for this, but of course there
exist many different variants with their own
nuances / strengths and this topic is very much
an area of active research [
7
]. Detailed analysis
is needed what specific algorithmic procedure
in what setup best leverages the already gained
insights of the single distributed learners (to
illustrate the nuances of such a procedure, it
shall be mentioned that the newly developed
federated learning concept also considers
which subparts of the training set are sent to
the distributed clients) [
4
].
3. Compatibility: The agents that participate in
hierarchical decentralized learning or same
level synchronization must adhere to the same
machine learning method / model as a
necessary condition in order to facilitate
consolidation and exchange of information
(for example weight functions). Of course, as
there exists a huge variety of machine learning
procedures which are all performed
differently, these methods need to be aligned.
The same applies to questions of
implementation, since, for example, in a smart
factory various proprietary devices with
underlying architectures and software
solutions have to work together [
8
].
4. Security/Privacy: The security challenges
arise when a lot of not fully trustworthy agents
contribute to the decentralized learning (either
hierarchical or same level) [
6
]. One can refer
here to discussions about security in
serviceoriented process environments - there are
similar challenges (for example, asking for a
trustful registry). It could be that information
supplied for consolidation is "manipulated"
intentionally to distort the quality of learning
[
6
]. Here lies the corresponding challenge of
suppressing such occurrences. Apart from that
challenge, decentralized learning is commonly
regarded as a way to increase privacy since
confidential data can be kept locally and only
processed (anonymized) information, like
weightings of ML routines, is then shared [
6
].
Considering these technological aspects, it can be
discerned that there are several potential
advantages which can be gained from the
application of decentralized learning scenarios.
However, also respective challenges need to be
addressed in order to fully realize these potential
advantages.
3
Commercial application areas and use cases
After understanding technological aspects,
business questions and typical use cases that build
on the decentralized learning idea shall be
addressed. Generally, decentralized artificial
intelligence can be applied to any field where
currently artificial intelligence solutions are
deployed since there are typically multiple devices
in use which are up to now not connected regularly
in a hierarchical or same level relationship. In the
following, two selected areas are outlined since
they play a major role in the ongoing discussion
[
9
] in science and business practice: the
“embedded AI solutions” and applications in the
smart factory setting.
3.1
Embedded AI Solutions
There is one related thinking approach by C. E.
Bouee [
10
] that deserves special attention as he
was the first to address the decentralization issue
of AI in full force (in particular with respect to its
commercial dimension). He envisions an enforced
development and usage of “portable AI” solutions
which are decentralized AI solutions that can be
integrated (embedded) into all kinds of devices,
typically resembling a highly individualized smart
assistant [
10
]. Furthermore, he calls into doubt that
the large digital platforms that dominate the tech
sector today will dominate the market for portable
AI solutions in the future and even believes that
portable AI could end their monopoly [
10
].
He provides several examples for such emerging
companies that are active in the segment of
embedded AI solutions, thereby illustrating the
parameter decision of (de)centralization [
10
]:
1. Snips (from France) attempts to put “AI in
every device”, hence creating “a unified voice
strategy with a complete range of interface
offerings - from simple voice commands to
comprehensive natural language voice
recognition” [
11
]. So they are exactly
addressing the softwarization topic already
described.
2. Arago (from Germany) focuses on the B2B
sector and its solutions are meant “to automate
enterprise IT and business operations” via
machines “with human problem-solving
skills” that underwent supervised training
[
12
]. Hence, they essentially support the
automation of business processes and show the
potential of decentralized AI solutions therein.
3. SenseTime (from China) is active in many
deep learning and supercomputing areas, in
particular with its independently developed
large-scale training system it also employs
decentralized machine learning [
13
].
Therefore, they act as a kind of orchestrator for
decentralized learning.
From this exemplary listing one can already get an
idea in what fields entrant companies try to capture
market share from the incumbent large tech
companies. Foremost, they are supposed to have a
competitive advantage over the incumbent
companies since they are not burdened with
having negative publicity concerning data privacy
and security issues. Furthermore, it is conceivable
that they gain advantages from more appealing
product offerings with higher individualization
and “tailor-made” fit [
10
]. One further argument
of Bouee for the competitive edge of these
emerging firms is that data becomes obsolete
faster and that therefore the advantages which
incumbent companies possess from their already
amassed data consequently diminishes very
quickly [
10
]. However, on the contrary it can be
argued that the large incumbent companies with
their already existing wide outreach and
computational power do still retain an advantage
in this respect since bigger amounts of data are
created faster and hence require much more
computational power for assessing and evaluating
these large volumes. This is certainly the case with
respect to potent cloud offerings which heavily
rely on the magnitude of computational strength.
The aspect of highly individualized smart
assistants is also very interesting, as this is
somehow the opposite of a smart assistant
provided by a large technology company that
could be sometimes prone “to lose its voice” (as
featured in a famous commercial) [
14
].
Highly individualized smart assistants are then AI
solutions in their most decentralized form as these
solutions do not just complement technical
devices, but also every single end user [
10
]. Bouee
elaborates that basically new business models will
arise from this as these individualized smart
assistants can act as much more strict gatekeepers
for personal decisions than common smart
assistants currently do [
10
]. An interesting
immediately arising question is, whether
companies can and will deliberately set up barriers
such that personalized smart assistants cannot be
used to full extent (much like the already existing
bot-barriers for travel websites).
The topic of decentralization is of particular
relevance in a likely “second wave” of
digitization. While in the first (B2C
marketingoriented) wave large (U.S.) companies use
centrally stored personalized data, manufacturing
processes and B2B transactions will then be
digitized in a second wave. Here, the tendencies to
centralize AI will likely not take place due to
already established physical equipment (with
various therein embedded software standards)
which is difficult to merge. Furthermore, the
possible fear of process participants of losing
control over their data encourages
decentralization.
3.2 Smart Factory
The other large application field that should be
discussed in more detail is the smart factory setting
which was originally envisioned by the German
Ministry of Education and Research (BMBF) [
15
].
“Smart factory” refers to manufacturing processes
(including resources) in which relevant involved
machines and devices are all equipped with
sensors and perform typical smart tasks, like
Condition Monitoring, Diagnosis for Maintenance
and Optimization of Processes, that can be
improved via machine learning [
16
]. It is certainly
conceivable that an enterprise has multiple smart
factories in use that either share the full
infrastructure or at least key components. This is
then the “factory network” scenario that strives to
optimize network performance [
17
]. This typically
addresses the four basic layers of the “Life Cycle
Value Stream” of the RAMI 4.0 framework:
Asset, Integration, Communication and
Information [
18
].
As data basically is the “lifeblood” of smart
factories, an evident approach for maximizing
network performance would be to harness the
available data of the whole network in a more
efficient way by centralizing the decentralized
learnings of specific similar tasks.
Here one has to carefully distinguish according to
the degree of similarity. To give two extreme
examples: Concerning autonomous driving it
makes a lot of sense to align and share learnings,
since increasing the security of autonomous
driving has highest priority and conditions of
streets are highly standardized. However, if
natural language processing is concerned, it seems
recommendable not to mix and interchange
training data or learning weights of different
dialects within a language (although they do
represent the same language) as this rather tends
to create confusion instead of resolving it.
In the same manner one has to proceed in the smart
factory setting, since there might be tasks and
routines that are more similar to each other than to
the rest (depending on the type of smart factory).
Certainly, such training and synchronization
procedures can be applied to the same type of task
within the same type of factory. But generalization
to similar tasks is, as always, accompanied by
certain challenges in deciding whether the
similarity is sufficient enough. Here it is
conceivable that differences for Condition
Monitoring and Diagnosis for Maintenance
concern certain types of machines, whereas
Optimization of Processes might depend on the
level in a multi-level production process.
However, if a sufficient similarity can be
guaranteed, such (de)centralized learning concepts
make a lot of sense. One could argue that different
tasks are never similar enough, however,
considering that a factory is a highly standardized
setting, it is conceivable that there might be a
smart transformation/translation of corresponding
use cases.
But not only the question “whether” it does make
sense, also the question “how often or regular”
such a synchronization shall take place has to be
answered. In a mimimum scenario, the
synchronization can take place in typical inactive
periods like production pauses or overnight as not
to interfere with the regular business operations
(thereby reducing process speed). This is in
particular apt for devices and components used in
process steps that require fast decisions and
actions. However, if devices and components are
operating on a much slower time scale (for
example registering inbound commodity flows or
outbound logistics of finished products) the
synchronization could take place continuously.
Similar questions are relevant when integrating
partners in SCM (Supply Chain Management) or
PLM (Product Lifecycle Management) processes.
However, in modern value networks which rely on
data exchange for smart/reliable decisions,
wellknown process standards (e.g., EDI, flat XML)
may not be sufficient.
According to experts of “Internet of Business” the
10 smart factory trends to watch in 2019 are the
following [
19
]:
1. “Collaborative robots will augment
workforces”
2. “Cloud robotics & APIs will give
manufacturers greater control”
3. “Robotics-as-a-service will make robotics
viable for smaller manufacturers”
4. “5G & Multi-access Edge Computing (MEC)
will help keep factory workers informed”
5. “Edge computing will see new use cases”
6. “Cybersecurity will be given greater priority”
7. “AI & advanced analytics will become
nearubiquitous”
8. “Digital twins will be employed more widely
across manufacturing and the supply chain”
9. “Additive manufacturing will be used to create
final products”
10. “Wearables will become commonplace on the
factory floor”
In this list there are two trends that have a
particular high relevance for our discussed topic,
namely “Edge computing will see new use cases”
and “AI & advanced analytics will become
nearubiquitous”. By design, the components and
devices of a smart factory are meant to make
autonomous decisions. So in time-critical
processes it is wanted that more computation is
moving towards the edge in order to increase the
speed and decrease the latency. Also crucial is the
rising importance and spread of AI & advanced
analytics, since this first ensures that local AI
solutions are present and, furthermore, that there
exists potential which can be realized via their
alignment. It is certainly highly dependent on the
particular use case whether the cooperation of the
distributed AI solutions takes place on the same
level or in a “master-slave” setup.
But it has to be noted that the aforementioned
centralization procedures for AI solutions are not
yet to be anticipated in full force for the current
year, hence pointing out potentials for realizations
that still have to find their place on the respective
agendas. The strength of such shared learnings
should in particular become apparent in cases
where there is a lack of suitable training material,
for example when whole processes in smart
factory are redesigned according to machine
learning insights. Although this would be a
desirable feature of the smart factory, this redesign
of processes is certainly regarded as a much more
complex task than just checking the need for
maintenance of a machine since much more
factors (like implications for following processes
and business partners) have to be considered and
evaluated.
4