IEEE Computational Intelligence Magazine - February 2021 - 59

are potentially able to deal with the changes in the training
environment or new tasks. Therefore, the union and election
mechanisms of Dem-AI are related to the diversity maintenance of the biological species through natural selection and
non-competitive processes (i.e., symbiosis) in the evolution
process or the robustness of decentralized systems [37]. In addition to the measurement of efficiency, the robustness or diversity of the Dem-AI system can be measured and controlled
throughout the training time following a validation procedure.
Hierarchical learning model parameters averaging
for knowledge sharing: In the case of a shared generalized
model among all group members, the direct knowledge sharing can be designed with the hierarchical averaging of the generalized model parameters (GMP) at each level k as follows:

GMP ^N (gk)h = (1 - c t) GMP ^N (gk)h
+ ct

i!S

N (gk, i- 1)
^ (k - 1) h,
(k) GMP N g, i
Ng

(3)

where S is the set of subgroups of a specialized group, N (gk, i- 1)
is the number of agents in subgroup i, and N (gk) is the total
number of agents in the current specialized group at level k.
The model averaging implementation is a typical aggregation
mechanism adopted in several FL algorithms [12], [13].
Parameter c t controls the update frequency of the generalized
knowledge whose value decreases over time as the members
become well-specialized in their learning knowledge. Accordingly, the model parameters of the subgroups which have
more numbers of agents become more important in the generalized model.
Knowledge distillation and knowledge transfer: For
multiple complex tasks, the Dem-AI framework allows knowledge transfer across tasks in different domains by leveraging
collaboration amongst learning agents in the hierarchical structure. In this regard, multi-task learning enables generalization
by solving multiple relevant tasks simultaneously [20, 21, 38].
The work in [38] studied the relationship between jointly
trained tasks and proposed a framework for task grouping in
MTL setting. Accordingly, the authors have analyzed learning
task compatibility in computer vision systems by evaluating
task cooperation and competition. For example, a shared
encoder and representation is learned through training highlycorrelated tasks together such as in Semantic Segmentation,
Depth Estimation, and Surface Normal Prediction. However,
this framework is limited to analyzing multiple learning tasks
for a single agent. Whereas, in Dem-AI systems, a group of
agents can train similar tasks in the low-level groups. Meanwhile, highly related tasks can be jointly trained together in the
high-level groups.
Furthermore, the latent representations across different
devices or groups are supported by adopting existing techniques of knowledge distillation, transfer learning, meta-knowledge construction, and specialized knowledge transfer.
Knowledge distillation [22] and knowledge transfer among
multiple tasks [10, 11] are important techniques to extend the

capabilities of knowledge sharing. For example, in [22], knowledge distillation mechanisms such as exchanging model parameters, model outputs, and surrogate data are incorporated in
distributed machine learning frameworks. Meanwhile, knowledge transfer has been recently studied in the federated MTL
setting using different types of MTL regularization such as
cluster structure, probabilistic priors, and graphical models [10].
Moreover, the work [11] forms a Bayesian network and uses
variational inference methods with the lateral connections
design between the server and client models to transfer knowledge among tasks. Different from the recent works, the conventional organizational knowledge creation theory [39] introduced
a promising paradigm in which the new knowledge of an
organization is articulated from the knowledge of individuals
and self-organized in a hierarchical structure. Thus, the shared
knowledge can be in an abstract form or an explicit combination of the individual's knowledge through the conceptualization and crystallization process. In doing so, together with the
hierarchical learning model parameters averaging, we can
develop suitable knowledge sharing approaches for the generalization mechanism in our Dem-AI systems.
E. Example of Dem-AI Systems

More recently, the use of personalized applications, such as virtual assistants that could adhere to users' personality, has gained
significant attraction. The goal of such intelligent systems is to
learn the unique features and personalized characteristics during daily activities and make appropriate decisions for each
user, then enhance user interest. However, the main problem is
the extraction of personalized features to perform knowledge
transfer with limited local data. The Dem-AI system allows
end-users and service providers to take part in a win-win solution, that is, the service providers exploit user's knowledge
to scale up their services, and the end-users can collectively
improve their personalized performance through knowledge
sharing in a suitable group. For example, Google has provided a
personalized virtual assistant (i.e., Google Now [40]) which can
respond to user's questions with more relevant answers. Such
reactive response systems can be extended to provide intelligent
personalized recommendation services in a proactive manner.
In this application, the hierarchical recommendation models
can be constructed following Dem-AI mechanisms by leveraging the shared features from different domains, and users/
groups at different levels.
In addition, we present a novel multi-language handwriting recognition system based on our Dem-AI reference
design, as shown in Fig. 7. A typical handwriting language
recognition application has an embedded virtual assistant to
improve the capability for understanding human written
texts in various languages. However, to realize such systems,
we need separate recognition models for each language (e.g.,
English and Korean). Using our Dem-AI reference design,
agents undergo self-organization to form appropriate hierarchical regional/social groups so as to share the similarity in
the characteristics of their languages. By exploiting such

FEBRUARY 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - February 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2021