Goal -
Autoencoche ?
-
-
/
-
Decompassed
compressed Reconstructed
Limage
D
Priginal
image
2 = Wich ,
tw 12 TWsWs-- - >
--- Neurous
Funch (2).
Activat
1
Filters
a ques
I kame
m -
learn wob
ANN a
-
parame
LI ·! Detecting all features
element
CNN
kee
eul
-
emage
-]
W2
We
us IIa
xz
RI
us won
I unde
N5
rog
+ b)
~
morks---
que
E
,
,
not product of
WoR
# sum
ANN weighted
our
input
-
filter/keemd
Nuron
elements
a
of filters
connects
weights
-
do Activation
Diff o
- op
ANN Single
Malin
Map
it
E
X
ANN
-
flatten
- -
·
-
-
-
-
123
n
Encode
Is-28
-
Featu
- #
op
layer
-
14 32)
(14
.
a
come N/
,
①
filters Feature
32 Maps
- 353
Kernel
=2
strides
no
of parameters
Filter ,
Per
- +
!
=
paramo
323
- bias Params
wits 320
-
10 params
Filters Filters
64
32 2
layer
-
2D
come maps
② feature
132 this layer
il p 2)
for
Filte
6-
Filter
e
Per 283
parms 288
+1 =
(32-383)
=
+ 1
7 64
open
Rosinas
.
.
Filter = ↓
ju
&
642289 72706
Y
layer as -
Plate
③
=
layer &
64
⑨ Dense o(p
bias3137
params
+t
3136 won
-3 137
params
per
- 6S
-
neurons 2007
by I
= -
we w
w
↓
-
y
D 9
9
Dog
Cat
Feature
Detecting rector
op a
"Compressed
Encoder
&
j
e
))
: 25 IA
:;
25
o 255
-
Ba fee s
FM
olp 32
14)
of ( ,
Kemel
H
O
zaza32
Decoder
-
2D 64e313
Dense 64)
(2 ,
2 ,
Reshape XI
14 14 ,
,
18
es ,
T
H
1128 . ,
28
olp
becoder %
j
X =>
- O
)831
① De ( +
6553136
- (2 ,
2
sus
,
69)
Reshape -upsampling
② b
-
(3 3054564)
+
Transpose
21
come +
ass
③ -
- 64 ↓ Filte
Fr
Filter us
&(3 3)
kennel
,
ine
olp a
14 14 .
64 no of featur /
(
,
Maps
Filter beight Ip
chance this
(filter
widsu
a
a olp channels
t 32
6432
(303) ↓ as
Filter
=
18464
watermark
Removing image
dij from
Remove
Autofill
Autoencodel Reconstructed
-
compressed-
recher >1 Decode
~
>
- Snope
writ
-
- ↓ Nov
with
pinel
Image
Reconstructed
↑ Eage
~
↑
·
-
Red
- Remov y
Pined
y
30
is by i >
- Yourput
j Denoising &
180s
gos
mags
·
Noise
pic
↓
Do y -
y
⑧ Denois
&
Isos
J
90s
-
·
IO
Distated = Ip
zuge
Image
N
Distanted surge
# & b
-
Data
&
Image
-
Test
-
Data a
NewCata
I,
olp
--
New
Data
· Date op
Det
Traine
2/p
AB
nage
e -
bistes
Fi
-
in
over
Ma
New Images
Generation of
-
Models Deepfale
Generational
separat
- or
- Filters
Dall-E & zust
Commer
midy It new
smage
wat
n[
avesmir) variation
Diffusion
VAE Model
GAN ↓
H tive
General
NIw
GANS vichos
= Audio
New Images
Generate
Goal -
right O
I
D
O Enemy I
-
~ fle/ Real
derate mage #
S
S Enemy
Adeensaufe -
& Every
Models
2
ht
① Generater ② Discriminator
=>
DiscriminateReal
Generate e
e
Fake
Images Belen Pale
Images
W
Discriminate
-
Take Train
·
nemeter
Discerate
Generate
*
GenerateDe
He
Discriminator
T
Real
Fake
-
-
DD ,
=
G H
Du Train D
pain wab
b
We
winner ,
, lose
&
unchanged
·
remain
winner model 3, 3
update 3 3
3
.
,
,
lose
complete
mode
collapse
-
Partial
- 3/119
-
T
-
Intermittent E Q]
100
As
Diversifying
s
&
a Discriminator
-
Game
-
