IEEE Computational Intelligence Magazine - February 2021 - 58

min
w

2
L (n0) ` w D (n0) j + b t / 1(k) w - w (nk) ; (2)
N
g
k

learning system can accommodate large group members and
ensure more frequent model updates, thereby improving the
group learning performance.

(0)

where L k is the learning loss function of the learning agent n
(0)
for a classification task given its personalized dataset D k and
(k)
the learning model w n of the higher-level groups [33]. Also,
other knowledge transfer techniques such as multi-task regularizer [21] or knowledge distillation regularizer [22] can be used
to define GK in the PLO of learning agents.
Specialized learning problem of specialized groups:
In case of group members who do not have adequate learning
capabilities (e.g., IoT sensing devices with low computational
capabilities), or learning agents who are required to solve a
coordination task. For example, in practical IoT applications, a
given learning system may not always be able to guarantee the
participation of all clients in every communication round due
to intermittent communication, battery drainage, or hardware
ailments [34]. Therefore, a specialized group learning problem
can be performed at the edge servers, and/or fog nodes at the
network's edge as done, for example, in the In-Edge AI framework [35]. The goal of this problem is to fit the collective datasets of all group members, where the group behaves as a virtual
agent that solves the learning problem. Furthermore, the specialized group learning can also have special decentralized
learning structures (e.g., sharing of critic network in multiagent deep reinforcement learning (DRL) [36], meta-training
phase in MLF [9]). Similar to the personalized learning problem, specialized group learning needs to be extended by leveraging the generalized knowledge from the higher-level
specialized groups. Next, in order to achieve joint energylearning efficiency given the limited communication and
computation resources, the design of group learning problem
needs to consider a synergy of resource allocation, device
scheduling, and learning perfor mance. In doing so, the

Personalized
Learning Direction

D. Generalization Mechanism

The generalization mechanism aims to collectively construct
the hierarchical generalized knowledge from all existing specialized groups or learning agents, as illustrated in Fig. 6(b).
Accordingly, the generalized knowledge extends the generalization capability of the Dem-AI system to learn new tasks or
deal with environment changes more efficiently. For this purpose, we propose four strategies that are suitable for different
levels: direct knowledge sharing, indirect knowledge sharing, election,
and union, which can be fixed or can be realized from a categorical distribution of these strategies. At the lowest-level specialized groups, the direct knowledge sharing of learning agents
is possible due to the similarity in the learning task to be performed. At the higher-level specialized groups, the indirect
knowledge sharing among subgroups (i.e., transferred knowledge, meta-knowledge) becomes more probable due to the
huge differences among specialized groups and the characteristics of learning tasks. Throughout the learning process, the
groups become more and more specialized to efficiently solve
different complex learning tasks. Consequently, the generalized
knowledge of the specialized groups becomes very different at
a higher level. Thus, an election mechanism based on voting
can help in reaching consensus among specialized groups. To
this end, a union mechanism is designed as an ensemble of the
collection of highly-specialized groups. This is a possible way to
maintain the diversity of potential groups for the entire learning system. Basically, the diversity of potential groups plays a
vital role in the learning system. It allows the preservation of
ineffective specialized groups who have fewer members or
those who show low performance in the training setting, but

Generalization

Group Learning
Direction

Union
Learning Models

Feedback

Specialized Groups

Generalized Level 3

Learning Agents
Specialized Learning
(a)

Election
Generalized Level 2
Knowledge
Sharing
Generalized Level 1

Personalized Level

(b)
FIGURE 6 Specialized learning and generalization mechanism in the Dem-AI system. (a) Specialized learning mechanism. (b) Hierarchical generalization mechanism.

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2021

IEEE Computational Intelligence Magazine - February 2021

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