IEEE Computational Intelligence Magazine - February 2021 - 54

A. Definitions, Goal, and Concepts

Definition and goal: Democratized Learning (Dem-AI in
short) focuses on the study of dual (coupled and working
together) specialized-generalized processes in a self-organizing
hierarchical structure of large-scale distributed learning systems. The specialized and generalized processes must operate
jointly towards an ultimate learning goal identified as performing collective learning from biased learning agents, who
are committed to learn from their own data using their limited
learning capabilities.
As such, the ultimate learning goal of the Dem-AI system is
to establish a mechanism for collectively solving common
(single or multiple) complex learning tasks from a large
number of learning agents. In case of a common single
learning task setting, the Dem-AI system aims to improve
the aggregation accuracy of all agents, as done in federated
learning. The ultimate goal in a conventional federated
learning system is to minimize the average model loss of all
agents for a single learning task. Moreover, for different
learning settings and applications, the goal of Dem-AI systems can be derived from following specific designs of learning objectives. The learning agents can also collectively
contribute to solving common multiple tasks as an ultimate
goal in Dem-AI system. For example, with multi-tasks learning setting [20], the goal is to attain the overall learning performance for multiple tasks, simultaneously. In meta-learning
[9], the learning goal is to construct a generalized knowledge
that can efficiently deal with similar new learning tasks. Similarly, in reinforcement learning, the coordination task is
defined to maximize the cumulative rewards for the joint
actions taken by individual agents according to their partial
observation [19].
Democracy in learning: The Dem-AI system features a
unique characterization of participation in the learning process,
and consequently develops the notion of democracy in learning
whose principles include the following:
❏❏ According to the differences in their characteristics, learning
agents are divided into suitable groups that can be specialized for the learning tasks. These specialized groups are selforganized in a hierarchical structure to mediate voluntary
contributions from all members in the collaborative learning
for solving multiple complex tasks.
❏❏ The shared generalized learning knowledge supports specialized groups and learning agents to improve their learning performance by reducing individual biases during
participation. In particular, the learning system allows new
group members to: a) speed up their learning process with
the existing group knowledge and b) incorporate their new
learning knowledge in expanding the generalization capability of the whole group.
❏❏ Learning agents are free to join suitable learning groups and
exhibit equal power in constructing their groups' generalized learning model. The group power can be represented
by the number of its members, which can vary over the
training time.

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2021

Dem-AI Meta-Law: We define a meta-law as a mechanism that can be used to manipulate the transition between
the dual specialized-generalized processes of our Dem-AI
system. This meta-law is driven by two coincident primary
forces: 1) a stability force, and 2) a plasticity force. Throughout
the learning time, the transition mechanism adjusts the
importance weight between these forces to empower the
plasticity or the stability in the specialized learning and generalization as well as the hierarchical structure of the DemAI system. The Dem-AI meta-law also provides the necessary
information to regulate the self-organizing hierarchical
structuring mechanism.
Specialized Process: This process is used to leverage the
specialized learning capabilities in the learning agents and specialized groups by exploiting their collected data. This process also
drives the hierarchical structure of specialized groups with
many levels of relevant generalized knowledge to become stable and well-separated. Thus, with the addition of higher levels
of generalized knowledge created by the generalization of all
specialized group members, the learning agents can exploit
their local datasets so as to reduce biases during personalized
learning for a single learning task. Thus, the personalized learning objective has two goals in its learning problem: 1) to perform specialized learning, and 2) to reuse the available hierarchical
generalized knowledge. Besides, generalized knowledge can be
incorporated by the regularizers for the personalized learning
objectives. Moreover, the specialized learning can be performed
as group learning formation when the members do not have
learning capabilities on their own or when they are required to
solve a coordination task. Considerably, over time, the generalized knowledge becomes less important compared to the specialized learning goal and a more stable hierarchical structure of
specialized groups will form. These transitions are the direct
consequence of the stability force characterized by the specialized knowledge exploitation and knowledge consistency. However, it becomes stronger over time in the meta-law design of
our Dem-AI principle.
Generalized Process: This process is used to regulate the
generalization mechanism for all existing specialized groups as
well as the plasticity level of all groups. Here, group plasticity pertains to the ease with which learning agents can change their
groups. The generalization mechanism encourages group
members to become close together when doing a similar
learning task and sharing knowledge. This process enables the
sharing of knowledge among group members and the construction of a hierarchical level of the generalized knowledge
from all of the specialized groups. Thereafter, the generalized
knowledge helps the Dem-AI system to maintain the generalization ability for efficiently dealing with environment changes
or new learning tasks. Hence, knowledge sharing is the mechanism to construct the generalized knowledge from similar and
correlated learning tasks such as model averaging in FL [12],
sharing knowledge mechanisms in multi-tasks learning [21],
and knowledge distillation [22]. Moreover, to resolve the conflict among excessively different specialized groups, an election

IEEE Computational Intelligence Magazine - February 2021

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