Fig. 1. Classification of the different approaches that are able to solve the problem of spatial heterogeneity in the input spaces.
Table 2 Summary of the datasets employed in the works presented in Section
3.2
. Datasets
marked with an asterisk are modified in different ways, making it impossible to fairly
compare each other. Some of the datasets mentioned were not referenced so far:
Omniglot [
65
], OTB [
66
], VOT2014 [
67
], Shakespeare [
2
] and Fashion MNIST [
68
].
Article
Datasets used in experiments
[
40
]
MNIST;
CIFAR-10
[
45
]
MiniImageNet;
Omniglot
[
46
]
MNIST*;
CIFAR-10*
[
47
]
CASAS
[
49
]
CIFAR-100;
FLICKR-AES
[
50
]
OTB;
VOT2014
[
52
]
MNIST;
FEMNIST;
Shakespeare
[
53
]
MNIST*;
CIFAR-10*;
EMNIST
[
58
]
MNIST*;
CIFAR-100*
[
59
]
MNIST*;
FEMNIST
[
62
]
Fashion MNIST;
EMNIST
[
63
]
Fashion MNIST;
EMNIST
back to
Table 1
, we classify these methods into 2 categories: those who
work with changes in the input space throughout clients, i.e., changes
in 𝑃 (𝑥); and those who work with changes in the behaviour throughout
clients, i.e., changes in 𝑃 (𝑦 |𝑥).
3.3.1. Changes in the input space throughout clients Each participant can collect data from their own input domain.
However, in this Section we assume that their domains will remain
unchanged during training, i.e., they will follow an IID data distribution
over time. In decentralized settings, participants collect data indepen-
dently, so nothing can assure their input domains are the same although
that could be the case for some of them. This is by far the most studied
kind of non-IID data, and strategies developed in this line of research
pay attention to multiple problems that may occur in real-life scenarios.
In this review, we include strategies that are not performed in the FL
framework, but that are prone to be adapted to such settings. These
works are classified into two main categories: (i) Domain Transformation and (ii) Domain Adaptation, which also branch into different approaches
(see
Fig. 1
).
Concerning the first of them, Domain Transformation methods focus
on detecting the particularities and the common parts of the data
domains, and they try to build a new shared input space. After that,
the input spaces perceived are transformed into a common input space.
As far as we are concerned, this strategy has only been performed
in centralized settings. Nonetheless, in a decentralized setting each
