Information Fusion 88 (2022) 263–280
264
M.F. Criado et al.
to forget the knowledge from previous tasks. Concept drift, on the con-
trary, is a problem that arises when the model is learning a single task,
but the data distribution is not homogeneous. As a result of this, the
model performance tends to drop dramatically. These inconveniences
occur in any realistic situation that presents a time-evolving nature,
such as FL tasks and many others.
In this work, we present some of the possible scenarios that can arise
when trying to solve a real problem applying FL, and the difficulties
that need to be faced. We classify those scenarios attending to the
statistical heterogeneity of data, combining the federated and continual
settings to visualize the whole problem, and we present a collection of
the most remarkable techniques that have been studied to deal with
some of those issues. We also notice that some real situations that in-
volve both Federated and Continual Learning have not been considered
nor handled so far, and they should be taken into account [
8
–
10
].
The rest of this paper is organized as follows: Section
2
reviews
the state-of-the-art techniques for Federated Learning. In Section
3
, we
present the definition and classification of non-IID data in a federated
environment, and we also discuss the different strategies to deal with it.
In Section
4
, we introduce the Continual Learning framework and the
multiple ways data can evolve over time. In Section
5
, we combine the
different situations of heterogeneous data to show all of the possible
scenarios. In addition, we discuss the strategies used to train a model
under concept drift that are close to the federated learning framework,
and we present a set of restrictions on the data collected that must be
verified to apply appropriate strategies. In Section
6
we empirically
show how the performance of some strategies drop in heterogeneous
settings where the mentioned restrictions are not satisfied. Finally,
Section
7
gathers our main conclusions and unsolved challenges.
Dostları ilə paylaş: