<s>
A	O
generative	B-Algorithm
adversarial	I-Algorithm
network	I-Algorithm
(	O
GAN	O
)	O
is	O
a	O
class	O
of	O
machine	O
learning	O
frameworks	O
designed	O
by	O
Ian	O
Goodfellow	O
and	O
his	O
colleagues	O
in	O
June	O
2014	O
.	O
</s>
<s>
Two	O
neural	B-Architecture
networks	I-Architecture
contest	O
with	O
each	O
other	O
in	O
the	O
form	O
of	O
a	O
zero-sum	O
game	O
,	O
where	O
one	O
agent	O
's	O
gain	O
is	O
another	O
agent	O
's	O
loss	O
.	O
</s>
<s>
Though	O
originally	O
proposed	O
as	O
a	O
form	O
of	O
generative	O
model	O
for	O
unsupervised	B-General_Concept
learning	I-General_Concept
,	O
GANs	O
have	O
also	O
proved	O
useful	O
for	O
semi-supervised	B-General_Concept
learning	I-General_Concept
,	O
fully	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
and	O
reinforcement	O
learning	O
.	O
</s>
<s>
The	O
core	O
idea	O
of	O
a	O
GAN	O
is	O
based	O
on	O
the	O
"	O
indirect	O
"	O
training	O
through	O
the	O
discriminator	O
,	O
another	O
neural	B-Architecture
network	I-Architecture
that	O
can	O
tell	O
how	O
"	O
realistic	O
"	O
the	O
input	O
seems	O
,	O
which	O
itself	O
is	O
also	O
being	O
updated	O
dynamically	O
.	O
</s>
<s>
Typically	O
,	O
the	O
generative	O
network	O
learns	O
to	O
map	O
from	O
a	O
latent	B-Algorithm
space	I-Algorithm
to	O
a	O
data	O
distribution	O
of	O
interest	O
,	O
while	O
the	O
discriminative	O
network	O
distinguishes	O
candidates	O
produced	O
by	O
the	O
generator	O
from	O
the	O
true	O
data	O
distribution	O
.	O
</s>
<s>
Typically	O
,	O
the	O
generator	O
is	O
seeded	O
with	O
randomized	O
input	O
that	O
is	O
sampled	O
from	O
a	O
predefined	O
latent	B-Algorithm
space	I-Algorithm
(	O
e.g.	O
</s>
<s>
Independent	O
backpropagation	B-Algorithm
procedures	O
are	O
applied	O
to	O
both	O
networks	O
so	O
that	O
the	O
generator	O
produces	O
better	O
samples	O
,	O
while	O
the	O
discriminator	O
becomes	O
more	O
skilled	O
at	O
flagging	O
synthetic	O
samples	O
.	O
</s>
<s>
When	O
used	O
for	O
image	O
generation	O
,	O
the	O
generator	O
is	O
typically	O
a	O
deconvolutional	O
neural	B-Architecture
network	I-Architecture
,	O
and	O
the	O
discriminator	O
is	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
.	O
</s>
<s>
Compared	O
to	O
fully	O
visible	O
belief	O
networks	O
such	O
as	O
WaveNet	B-Application
and	O
PixelRNN	O
and	O
autoregressive	O
models	O
in	O
general	O
,	O
GANs	O
can	O
generate	O
one	O
complete	O
sample	O
in	O
one	O
pass	O
,	O
rather	O
than	O
multiple	O
passes	O
through	O
the	O
network	O
.	O
</s>
<s>
Compared	O
to	O
Boltzmann	B-Algorithm
machines	I-Algorithm
and	O
nonlinear	O
ICA	B-Algorithm
,	O
there	O
is	O
no	O
restriction	O
on	O
the	O
type	O
of	O
function	O
used	O
by	O
the	O
network	O
.	O
</s>
<s>
Since	O
neural	B-Architecture
networks	I-Architecture
are	O
universal	B-Algorithm
approximators	I-Algorithm
,	O
GANs	O
are	O
asymptotically	O
consistent	O
.	O
</s>
<s>
Variational	B-Algorithm
autoencoders	I-Algorithm
might	O
be	O
universal	B-Algorithm
approximators	I-Algorithm
,	O
but	O
it	O
is	O
not	O
proven	O
as	O
of	O
2017	O
.	O
</s>
<s>
In	O
most	O
applications	O
,	O
is	O
a	O
deep	O
neural	B-Architecture
network	I-Architecture
function	O
.	O
</s>
<s>
developed	O
the	O
same	O
idea	O
of	O
reparametrization	O
into	O
a	O
general	O
stochastic	O
backpropagation	B-Algorithm
method	O
.	O
</s>
<s>
Among	O
its	O
first	O
applications	O
was	O
the	O
variational	B-Algorithm
autoencoder	I-Algorithm
.	O
</s>
<s>
In	O
practice	O
,	O
the	O
generator	O
has	O
access	O
only	O
to	O
measures	O
of	O
form	O
,	O
where	O
is	O
a	O
function	O
computed	O
by	O
a	O
neural	B-Architecture
network	I-Architecture
with	O
parameters	O
,	O
and	O
is	O
an	O
easily	O
sampled	O
distribution	O
,	O
such	O
as	O
the	O
uniform	O
or	O
normal	O
distribution	O
.	O
</s>
<s>
Similarly	O
,	O
the	O
discriminator	O
has	O
access	O
only	O
to	O
functions	O
of	O
form	O
,	O
a	O
function	O
computed	O
by	O
a	O
neural	B-Architecture
network	I-Architecture
with	O
parameters	O
.	O
</s>
<s>
Further	O
,	O
even	O
if	O
an	O
equilibrium	O
still	O
exists	O
,	O
it	O
can	O
only	O
be	O
found	O
by	O
searching	O
in	O
the	O
high-dimensional	O
space	O
of	O
all	O
possible	O
neural	B-Architecture
network	I-Architecture
functions	O
.	O
</s>
<s>
The	O
standard	O
strategy	O
of	O
using	O
gradient	B-Algorithm
descent	I-Algorithm
to	O
find	O
the	O
equilibrium	O
often	O
does	O
not	O
work	O
for	O
GAN	O
,	O
and	O
often	O
the	O
game	O
"	O
collapses	O
"	O
into	O
one	O
of	O
several	O
failure	O
modes	O
.	O
</s>
<s>
To	O
improve	B-Algorithm
the	O
convergence	O
stability	O
,	O
some	O
training	O
strategies	O
start	O
with	O
an	O
easier	O
task	O
,	O
such	O
as	O
generating	O
low-resolution	O
images	O
or	O
simple	O
images	O
(	O
one	O
object	O
with	O
uniform	O
background	O
)	O
,	O
and	O
gradually	O
increase	O
the	O
difficulty	O
of	O
the	O
task	O
during	O
training	O
.	O
</s>
<s>
For	O
example	O
,	O
a	O
GAN	O
trained	O
on	O
the	O
MNIST	B-General_Concept
dataset	I-General_Concept
containing	O
many	O
samples	O
of	O
each	O
digit	O
might	O
only	O
generate	O
pictures	O
of	O
digit	O
0	O
.	O
</s>
<s>
So	O
for	O
example	O
,	O
if	O
during	O
GAN	O
training	O
for	O
generating	O
MNIST	B-General_Concept
dataset	I-General_Concept
,	O
for	O
a	O
few	O
epochs	O
,	O
the	O
discriminator	O
somehow	O
prefers	O
the	O
digit	O
0	O
slightly	O
more	O
than	O
other	O
digits	O
,	O
the	O
generator	O
may	O
seize	O
the	O
opportunity	O
to	O
generate	O
only	O
digit	O
0	O
,	O
then	O
be	O
unable	O
to	O
escape	O
the	O
local	O
minimum	O
after	O
the	O
discriminator	O
improves	O
.	O
</s>
<s>
They	O
also	O
proposed	O
using	O
the	O
Adam	B-Algorithm
stochastic	I-Algorithm
optimization	I-Algorithm
to	O
avoid	O
mode	O
collapse	O
,	O
as	O
well	O
as	O
the	O
Fréchet	B-Algorithm
inception	I-Algorithm
distance	I-Algorithm
for	O
evaluating	O
GAN	O
performances	O
.	O
</s>
<s>
In	O
such	O
case	O
,	O
the	O
generator	O
cannot	O
learn	O
,	O
a	O
case	O
of	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
.	O
</s>
<s>
Intuitively	O
speaking	O
,	O
the	O
discriminator	O
is	O
too	O
good	O
,	O
and	O
since	O
the	O
generator	O
cannot	O
take	O
any	O
small	O
step	O
(	O
only	O
small	O
steps	O
are	O
considered	O
in	O
gradient	B-Algorithm
descent	I-Algorithm
)	O
to	O
improve	B-Algorithm
its	O
payoff	O
,	O
it	O
does	O
not	O
even	O
try	O
.	O
</s>
<s>
One	O
important	O
method	O
for	O
solving	O
this	O
problem	O
is	O
the	O
Wasserstein	B-Algorithm
GAN	I-Algorithm
.	O
</s>
<s>
GANs	O
are	O
usually	O
evaluated	O
by	O
Inception	B-General_Concept
score	I-General_Concept
(	O
IS	O
)	O
,	O
which	O
measures	O
how	O
varied	O
the	O
generator	O
's	O
outputs	O
are	O
(	O
as	O
classified	O
by	O
an	O
image	O
classifier	O
,	O
usually	O
Inception-v3	B-General_Concept
)	O
,	O
or	O
Fréchet	B-Algorithm
inception	I-Algorithm
distance	I-Algorithm
(	O
FID	O
)	O
,	O
which	O
measures	O
how	O
similar	O
the	O
generator	O
's	O
outputs	O
are	O
to	O
a	O
reference	O
set	O
(	O
as	O
classified	O
by	O
a	O
learned	O
image	O
featurizer	O
,	O
such	O
as	O
Inception-v3	B-General_Concept
without	O
its	O
final	O
layer	O
)	O
.	O
</s>
<s>
Another	O
evaluation	O
method	O
is	O
the	O
Learned	O
Perceptual	O
Image	O
Patch	O
Similarity	O
(	O
LPIPS	O
)	O
,	O
which	O
starts	O
with	O
a	O
learned	O
image	O
featurizer	O
,	O
and	O
finetunes	O
it	O
by	O
supervised	B-General_Concept
learning	I-General_Concept
on	O
a	O
set	O
of	O
,	O
where	O
is	O
an	O
image	O
,	O
is	O
a	O
perturbed	O
version	O
of	O
it	O
,	O
and	O
is	O
how	O
much	O
they	O
differ	O
,	O
as	O
reported	O
by	O
human	O
subjects	O
.	O
</s>
<s>
For	O
example	O
,	O
for	O
generating	O
images	O
that	O
look	O
like	O
ImageNet	B-General_Concept
,	O
the	O
generator	O
should	O
be	O
able	O
to	O
generate	O
a	O
picture	O
of	O
cat	O
when	O
given	O
the	O
class	O
label	O
"	O
cat	O
"	O
.	O
</s>
<s>
In	O
2017	O
,	O
a	O
conditional	O
GAN	O
learned	O
to	O
generate	O
1000	O
image	O
classes	O
of	O
ImageNet	B-General_Concept
.	O
</s>
<s>
In	O
the	O
original	O
paper	O
,	O
the	O
authors	O
demonstrated	O
it	O
using	O
multilayer	B-Algorithm
perceptron	I-Algorithm
networks	O
and	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
.	O
</s>
<s>
Self-attention	O
GAN	O
(	O
SAGAN	O
)	O
:	O
Starts	O
with	O
the	O
DCGAN	O
,	O
then	O
adds	O
residually-connected	O
standard	O
self-attention	B-General_Concept
modules	I-General_Concept
to	O
the	O
generator	O
and	O
discriminator	O
.	O
</s>
<s>
Variational	B-Algorithm
autoencoder	I-Algorithm
GAN	O
(	O
VAEGAN	O
)	O
:	O
Uses	O
a	O
variational	B-Algorithm
autoencoder	I-Algorithm
(	O
VAE	O
)	O
for	O
the	O
generator	O
.	O
</s>
<s>
Transformer	B-Algorithm
GAN	O
(	O
TransGAN	O
)	O
:	O
Uses	O
the	O
pure	O
transformer	B-Algorithm
architecture	O
for	O
both	O
the	O
generator	O
and	O
discriminator	O
,	O
entirely	O
devoid	O
of	O
convolution-deconvolution	O
layers	O
.	O
</s>
<s>
Hinge	B-Algorithm
loss	I-Algorithm
GAN:Least	O
squares	O
GAN:where	O
are	O
parameters	O
to	O
be	O
chosen	O
.	O
</s>
<s>
The	O
Wasserstein	B-Algorithm
GAN	I-Algorithm
modifies	O
the	O
GAN	O
game	O
at	O
two	O
points	O
:	O
</s>
<s>
An	O
adversarial	O
autoencoder	B-Algorithm
(	O
AAE	O
)	O
is	O
more	O
autoencoder	B-Algorithm
than	O
GAN	O
.	O
</s>
<s>
The	O
idea	O
is	O
to	O
start	O
with	O
a	O
plain	O
autoencoder	B-Algorithm
,	O
but	O
train	O
a	O
discriminator	O
to	O
discriminate	O
the	O
latent	O
vectors	O
from	O
a	O
reference	O
distribution	O
(	O
often	O
the	O
normal	O
distribution	O
)	O
.	O
</s>
<s>
The	O
idea	O
of	O
InfoGAN	O
is	O
to	O
decree	O
that	O
every	O
latent	O
vector	O
in	O
the	O
latent	B-Algorithm
space	I-Algorithm
can	O
be	O
decomposed	O
as	O
:	O
an	O
incompressible	O
noise	O
part	O
,	O
and	O
an	O
informative	O
label	O
part	O
,	O
and	O
encourage	O
the	O
generator	O
to	O
comply	O
with	O
the	O
decree	O
,	O
by	O
encouraging	O
it	O
to	O
maximize	O
,	O
the	O
mutual	O
information	O
between	O
and	O
,	O
while	O
making	O
no	O
demands	O
on	O
the	O
mutual	O
information	O
between	O
.	O
</s>
<s>
The	O
standard	O
GAN	O
generator	O
is	O
a	O
function	O
of	O
type	O
,	O
that	O
is	O
,	O
it	O
is	O
a	O
mapping	O
from	O
a	O
latent	B-Algorithm
space	I-Algorithm
to	O
the	O
image	O
space	O
.	O
</s>
<s>
This	O
naturally	O
leads	O
to	O
the	O
idea	O
of	O
training	O
another	O
network	O
that	O
performs	O
"	O
encoding	O
"	O
,	O
creating	O
an	O
autoencoder	B-Algorithm
out	O
of	O
the	O
encoder-generator	O
pair	O
.	O
</s>
<s>
,	O
the	O
latent	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
Applications	O
of	O
bidirectional	O
models	O
include	O
semi-supervised	B-General_Concept
learning	I-General_Concept
,	O
interpretable	O
machine	O
learning	O
,	O
and	O
neural	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
.	O
</s>
<s>
The	O
BigGAN	O
is	O
essentially	O
a	O
self-attention	O
GAN	O
trained	O
on	O
a	O
large	O
scale	O
(	O
up	O
to	O
80	O
million	O
parameters	O
)	O
to	O
generate	O
large	O
images	O
of	O
ImageNet	B-General_Concept
(	O
up	O
to	O
512	O
x	O
512	O
resolution	O
)	O
,	O
with	O
numerous	O
engineering	O
tricks	O
to	O
make	O
it	O
converge	O
.	O
</s>
<s>
In	O
such	O
cases	O
,	O
data	B-General_Concept
augmentation	I-General_Concept
can	O
be	O
applied	O
,	O
to	O
allow	O
training	O
GAN	O
on	O
smaller	O
datasets	O
.	O
</s>
<s>
Naïve	O
data	B-General_Concept
augmentation	I-General_Concept
,	O
however	O
,	O
brings	O
its	O
problems	O
.	O
</s>
<s>
Consider	O
the	O
original	O
GAN	O
game	O
,	O
slightly	O
reformulated	O
as	O
follows:Now	O
we	O
use	O
data	B-General_Concept
augmentation	I-General_Concept
by	O
randomly	O
sampling	O
semantic-preserving	O
transforms	O
and	O
applying	O
them	O
to	O
the	O
dataset	O
,	O
to	O
obtain	O
the	O
reformulated	O
GAN	O
game:This	O
is	O
equivalent	O
to	O
a	O
GAN	O
game	O
with	O
a	O
different	O
distribution	O
,	O
sampled	O
by	O
,	O
with	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
is	O
the	O
distribution	O
of	O
images	O
in	O
ImageNet	B-General_Concept
,	O
and	O
samples	O
identity-transform	O
with	O
probability	O
0.5	O
,	O
and	O
horizontal-reflection	O
with	O
probability	O
0.5	O
,	O
then	O
is	O
the	O
distribution	O
of	O
images	O
in	O
ImageNet	B-General_Concept
and	O
horizontally-reflected	O
ImageNet	B-General_Concept
,	O
combined	O
.	O
</s>
<s>
For	O
example	O
,	O
it	O
would	O
generate	O
images	O
that	O
look	O
like	O
they	O
are	O
randomly	O
cropped	O
,	O
if	O
the	O
data	B-General_Concept
augmentation	I-General_Concept
uses	O
random	O
cropping	O
.	O
</s>
<s>
The	O
solution	O
is	O
to	O
apply	O
data	B-General_Concept
augmentation	I-General_Concept
to	O
both	O
generated	O
and	O
real	O
images:The	O
authors	O
demonstrated	O
high-quality	O
generation	O
using	O
just	O
100-picture-large	O
datasets	O
.	O
</s>
<s>
The	O
StyleGAN-2-ADA	O
paper	O
points	O
out	O
a	O
further	O
point	O
on	O
data	B-General_Concept
augmentation	I-General_Concept
:	O
it	O
must	O
be	O
invertible	O
.	O
</s>
<s>
Continue	O
with	O
the	O
example	O
of	O
generating	O
ImageNet	B-General_Concept
pictures	O
.	O
</s>
<s>
If	O
the	O
data	B-General_Concept
augmentation	I-General_Concept
is	O
"	O
randomly	O
rotate	O
the	O
picture	O
by	O
0	O
,	O
90	O
,	O
180	O
,	O
270	O
degrees	O
with	O
equal	O
probability	O
"	O
,	O
then	O
there	O
is	O
no	O
way	O
for	O
the	O
generator	O
to	O
know	O
which	O
is	O
the	O
true	O
orientation	O
:	O
Consider	O
two	O
generators	O
,	O
such	O
that	O
for	O
any	O
latent	O
,	O
the	O
generated	O
image	O
is	O
a	O
90-degree	O
rotation	O
of	O
.	O
</s>
<s>
The	O
solution	O
is	O
to	O
only	O
use	O
invertible	O
data	B-General_Concept
augmentation	I-General_Concept
:	O
instead	O
of	O
"	O
randomly	O
rotate	O
the	O
picture	O
by	O
0	O
,	O
90	O
,	O
180	O
,	O
270	O
degrees	O
with	O
equal	O
probability	O
"	O
,	O
use	O
"	O
randomly	O
rotate	O
the	O
picture	O
by	O
90	O
,	O
180	O
,	O
270	O
degrees	O
with	O
0.1	O
probability	O
,	O
and	O
keep	O
the	O
picture	O
as	O
it	O
is	O
with	O
0.7	O
probability	O
"	O
.	O
</s>
<s>
This	O
way	O
,	O
the	O
generator	O
is	O
still	O
rewarded	O
to	O
keep	O
images	O
oriented	O
the	O
same	O
way	O
as	O
un-augmented	O
ImageNet	B-General_Concept
pictures	O
.	O
</s>
<s>
Discrete	O
case	O
:	O
Invertible	O
stochastic	B-Algorithm
matrices	I-Algorithm
,	O
when	O
is	O
finite	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
is	O
the	O
set	O
of	O
four	O
images	O
of	O
an	O
arrow	O
,	O
pointing	O
in	O
4	O
directions	O
,	O
and	O
the	O
data	B-General_Concept
augmentation	I-General_Concept
is	O
"	O
randomly	O
rotate	O
the	O
picture	O
by	O
90	O
,	O
180	O
,	O
270	O
degrees	O
with	O
probability	O
,	O
and	O
keep	O
the	O
picture	O
as	O
it	O
is	O
with	O
probability	O
"	O
,	O
then	O
the	O
Markov	O
kernel	O
can	O
be	O
represented	O
as	O
a	O
stochastic	B-Algorithm
matrix	I-Algorithm
:	O
and	O
is	O
an	O
invertible	O
kernel	O
iff	O
is	O
an	O
invertible	O
matrix	O
,	O
that	O
is	O
,	O
.	O
</s>
<s>
More	O
examples	O
of	O
invertible	O
data	B-General_Concept
augmentations	I-General_Concept
are	O
found	O
in	O
the	O
paper	O
.	O
</s>
<s>
SinGAN	O
pushes	O
data	B-General_Concept
augmentation	I-General_Concept
to	O
the	O
limit	O
,	O
by	O
using	O
only	O
a	O
single	O
image	O
as	O
training	O
data	O
and	O
performing	O
data	B-General_Concept
augmentation	I-General_Concept
on	O
it	O
.	O
</s>
<s>
The	O
StyleGAN	B-Application
family	O
is	O
a	O
series	O
of	O
architectures	O
published	O
by	O
Nvidia	O
's	O
research	O
division	O
.	O
</s>
<s>
Here	O
,	O
the	O
functions	O
are	O
image	O
up	O
-	O
and	O
down-sampling	O
functions	O
,	O
and	O
is	O
a	O
blend-in	O
factor	O
(	O
much	O
like	O
an	O
alpha	B-Algorithm
in	O
image	O
composing	O
)	O
that	O
smoothly	O
glides	O
from	O
0	O
to	O
1	O
.	O
</s>
<s>
StyleGAN-1	O
is	O
designed	O
as	O
a	O
combination	O
of	O
Progressive	O
GAN	O
with	O
neural	B-Algorithm
style	I-Algorithm
transfer	I-Algorithm
.	O
</s>
<s>
The	O
key	O
architectural	O
choice	O
of	O
StyleGAN-1	O
is	O
a	O
progressive	O
growth	O
mechanism	O
,	O
similar	O
to	O
Progressive	O
GAN	O
.	O
</s>
<s>
Each	O
style	O
block	O
applies	O
a	O
"	O
style	O
latent	O
vector	O
"	O
via	O
affine	O
transform	O
(	O
"	O
adaptive	O
instance	O
normalization	O
"	O
)	O
,	O
similar	O
to	O
how	O
neural	B-Algorithm
style	I-Algorithm
transfer	I-Algorithm
uses	O
Gramian	B-Algorithm
matrix	I-Algorithm
.	O
</s>
<s>
First	O
,	O
run	O
a	O
gradient	B-Algorithm
descent	I-Algorithm
to	O
find	O
such	O
that	O
.	O
</s>
<s>
This	O
is	O
called	O
"	O
projecting	O
an	O
image	O
back	O
to	O
style	O
latent	B-Algorithm
space	I-Algorithm
"	O
.	O
</s>
<s>
StyleGAN-2	O
improves	O
upon	O
StyleGAN-1	O
,	O
by	O
using	O
the	O
style	O
latent	O
vector	O
to	O
transform	O
the	O
convolution	O
layer	O
's	O
weights	O
instead	O
,	O
thus	O
solving	O
the	O
"	O
blob	O
"	O
problem	O
.	O
</s>
<s>
This	O
was	O
updated	O
by	O
the	O
StyleGAN-2-ADA	O
(	O
"	O
ADA	O
"	O
stands	O
for	O
"	O
adaptive	O
"	O
)	O
,	O
which	O
uses	O
invertible	O
data	B-General_Concept
augmentation	I-General_Concept
as	O
described	O
above	O
.	O
</s>
<s>
It	O
also	O
tunes	O
the	O
amount	O
of	O
data	B-General_Concept
augmentation	I-General_Concept
applied	O
by	O
starting	O
at	O
zero	O
,	O
and	O
gradually	O
increasing	O
it	O
until	O
an	O
"	O
overfitting	O
heuristic	O
"	O
reaches	O
a	O
target	O
level	O
,	O
thus	O
the	O
name	O
"	O
adaptive	O
"	O
.	O
</s>
<s>
StyleGAN-3	O
improves	O
upon	O
StyleGAN-2	O
by	O
solving	O
the	O
"	O
texture	O
sticking	O
"	O
problem	O
,	O
which	O
can	O
be	O
seen	O
in	O
the	O
official	O
videos	O
.	O
</s>
<s>
To	O
solve	O
this	O
,	O
they	O
proposed	O
imposing	O
strict	O
lowpass	B-Algorithm
filters	I-Algorithm
between	O
each	O
generator	O
's	O
layers	O
,	O
so	O
that	O
the	O
generator	O
is	O
forced	O
to	O
operate	O
on	O
the	O
pixels	O
in	O
a	O
way	O
faithful	O
to	O
the	O
continuous	O
signals	O
they	O
represent	O
,	O
rather	O
than	O
operate	O
on	O
them	O
as	O
merely	O
discrete	O
signals	O
.	O
</s>
<s>
The	O
resulting	O
StyleGAN-3	O
is	O
able	O
to	O
solve	O
the	O
texture	O
sticking	O
problem	O
,	O
as	O
well	O
as	O
generating	O
images	O
that	O
rotate	O
and	O
translate	O
smoothly	O
.	O
</s>
<s>
GANs	O
can	O
also	O
be	O
used	O
to	O
inpaint	B-Algorithm
photographs	O
or	O
create	O
photos	O
of	O
imaginary	O
fashion	O
models	O
,	O
with	O
no	O
need	O
to	O
hire	O
a	O
model	O
,	O
photographer	O
or	O
makeup	O
artist	O
,	O
or	O
pay	O
for	O
a	O
studio	O
and	O
transportation	O
.	O
</s>
<s>
In	O
2020	O
,	O
Artbreeder	B-Application
was	O
used	O
to	O
create	O
the	O
main	O
antagonist	O
in	O
the	O
sequel	O
to	O
the	O
psychological	O
web	O
horror	O
series	O
Ben	O
Drowned	O
.	O
</s>
<s>
GANs	O
can	O
improve	B-Algorithm
astronomical	O
images	O
and	O
simulate	O
gravitational	O
lensing	O
for	O
dark	B-Application
matter	I-Application
research	O
.	O
</s>
<s>
They	O
were	O
used	O
in	O
2019	O
to	O
successfully	O
model	O
the	O
distribution	O
of	O
dark	B-Application
matter	I-Application
in	O
a	O
particular	O
direction	O
in	O
space	O
and	O
to	O
predict	O
the	O
gravitational	O
lensing	O
that	O
will	O
occur	O
.	O
</s>
<s>
GANs	O
have	O
been	O
proposed	O
as	O
a	O
fast	O
and	O
accurate	O
way	O
of	O
modeling	O
high	O
energy	O
jet	O
formation	O
and	O
modeling	O
showers	O
through	O
calorimeters	B-Algorithm
of	O
high-energy	O
physics	O
experiments	O
.	O
</s>
<s>
In	O
2018	O
,	O
GANs	O
reached	O
the	O
video	O
game	O
modding	O
community	O
,	O
as	O
a	O
method	O
of	O
up-scaling	B-Algorithm
low-resolution	O
2D	O
textures	O
in	O
old	O
video	O
games	O
by	O
recreating	O
them	O
in	O
4k	B-Architecture
or	O
higher	O
resolutions	O
via	O
image	O
training	O
,	O
and	O
then	O
down-sampling	O
them	O
to	O
fit	O
the	O
game	O
's	O
native	O
resolution	O
(	O
with	O
results	O
resembling	O
the	O
supersampling	B-Algorithm
method	O
of	O
anti-aliasing	B-Algorithm
)	O
.	O
</s>
<s>
Known	O
examples	O
of	O
extensive	O
GAN	O
usage	O
include	O
Final	B-Application
Fantasy	I-Application
VIII	I-Application
,	O
Final	B-Application
Fantasy	I-Application
IX	I-Application
,	O
Resident	B-Application
Evil	I-Application
REmake	I-Application
HD	O
Remaster	O
,	O
and	O
Max	B-Application
Payne	I-Application
.	O
</s>
<s>
State-of-art	O
transfer	B-General_Concept
learning	I-General_Concept
research	O
use	O
GANs	O
to	O
enforce	O
the	O
alignment	O
of	O
the	O
latent	B-Algorithm
feature	I-Algorithm
space	I-Algorithm
,	O
such	O
as	O
in	O
deep	O
reinforcement	O
learning	O
.	O
</s>
<s>
GANs	O
that	O
produce	O
photorealistic	B-Application
images	O
can	O
be	O
used	O
to	O
visualize	O
interior	O
design	O
,	O
industrial	O
design	O
,	O
shoes	O
,	O
bags	O
,	O
and	O
clothing	O
items	O
or	O
items	O
for	O
computer	O
games	O
 '	O
scenes	O
.	O
</s>
<s>
Such	O
networks	O
were	O
reported	O
to	O
be	O
used	O
by	O
Facebook	B-Application
.	I-Application
</s>
<s>
GANs	O
can	O
reconstruct	B-Algorithm
3D	I-Algorithm
models	I-Algorithm
of	I-Algorithm
objects	I-Algorithm
from	I-Algorithm
images	I-Algorithm
,	O
generate	O
novel	O
objects	O
as	O
3D	O
point	O
clouds	O
,	O
and	O
model	O
patterns	O
of	O
motion	O
in	O
video	O
.	O
</s>
<s>
GANs	O
can	O
also	O
be	O
used	O
to	O
inpaint	B-Algorithm
missing	O
features	O
in	O
maps	O
,	O
transfer	O
map	O
styles	O
in	O
cartography	O
or	O
augment	O
street	O
view	O
imagery	O
.	O
</s>
<s>
In	O
1991	O
,	O
Juergen	O
Schmidhuber	O
published	O
generative	O
and	O
adversarial	O
neural	B-Architecture
networks	I-Architecture
that	O
contest	O
with	O
each	O
other	O
in	O
the	O
form	O
of	O
a	O
zero-sum	O
game	O
,	O
where	O
one	O
network	O
's	O
gain	O
is	O
the	O
other	O
network	O
's	O
loss	O
.	O
</s>
<s>
The	O
second	O
network	O
learns	O
by	O
gradient	B-Algorithm
descent	I-Algorithm
to	O
predict	O
the	O
reactions	O
of	O
the	O
environment	O
to	O
these	O
patterns	O
.	O
</s>
<s>
Adversarial	B-General_Concept
machine	I-General_Concept
learning	I-General_Concept
has	O
other	O
uses	O
besides	O
generative	O
modeling	O
and	O
can	O
be	O
applied	O
to	O
models	O
other	O
than	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
In	O
control	O
theory	O
,	O
adversarial	B-General_Concept
learning	I-General_Concept
based	O
on	O
neural	B-Architecture
networks	I-Architecture
was	O
used	O
in	O
2006	O
to	O
train	O
robust	O
controllers	O
in	O
a	O
game	O
theoretic	O
sense	O
,	O
by	O
alternating	O
the	O
iterations	O
between	O
a	O
minimizer	O
policy	O
,	O
the	O
controller	O
,	O
and	O
a	O
maximizer	O
policy	O
,	O
the	O
disturbance	O
.	O
</s>
<s>
Faces	O
generated	O
by	O
StyleGAN	B-Application
in	O
2019	O
drew	O
comparisons	O
with	O
Deepfakes	B-Application
.	O
</s>
<s>
In	O
May	O
2020	O
,	O
Nvidia	O
researchers	O
taught	O
an	O
AI	O
system	O
(	O
termed	O
"	O
GameGAN	O
"	O
)	O
to	O
recreate	O
the	O
game	O
of	O
Pac-Man	B-Application
simply	O
by	O
watching	O
it	O
being	O
played	O
.	O
</s>
