I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
14
,
No
.
2
,
A
p
r
il 2
0
2
5
,
p
p
.
1
3
1
1
~
1
3
2
5
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
2
.
p
p
1
3
1
1
-
1
3
2
5
1311
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Accuracy
of n
eur
a
l net
wo
rks i
n br
a
in wav
e diag
no
sis
of
schizo
phrenia
Su
k
em
i
1
,
G
a
briel
E
k
o
pu
t
ra
H
a
rt
o
no
Ca
hy
a
di
2
,
Sa
m
s
ury
a
di
3
,
M
uh
a
m
m
a
d Ag
un
g
Akba
r
4
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
y
st
e
m,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
s
S
c
i
e
n
c
e
,
S
r
i
w
i
j
a
y
a
U
n
i
v
e
r
si
t
y
,
P
a
l
e
m
b
a
n
g
,
I
n
d
o
n
e
s
i
a
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
y
s
t
e
ms
,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
S
r
i
w
i
j
a
y
a
U
n
i
v
e
r
s
i
t
y
,
P
a
l
e
m
b
a
n
g
,
I
n
d
o
n
e
si
a
3
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
c
En
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
S
r
i
w
i
j
a
y
a
U
n
i
v
e
r
si
t
y
,
P
a
l
e
m
b
a
n
g
,
I
n
d
o
n
e
s
i
a
4
N
u
r
si
n
g
S
t
u
d
y
P
r
o
g
r
a
m
,
S
T
I
K
e
s Al
-
M
a
'
a
r
i
f
,
B
a
t
u
r
a
j
a
,
I
n
d
o
n
e
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
2
2
,
2
0
2
3
R
ev
is
ed
Oct
1
8
,
2
0
2
4
Acc
ep
ted
No
v
1
5
,
2
0
2
4
Th
is
re
se
a
rc
h
e
x
p
lo
re
s
th
e
a
p
p
li
c
a
ti
o
n
o
f
a
m
o
d
if
ied
d
e
e
p
lea
rn
in
g
m
o
d
e
l
fo
r
e
lec
tro
e
n
c
e
p
h
a
lo
g
ra
p
h
y
(EE
G
)
sig
n
a
l
c
las
sifica
ti
o
n
in
t
h
e
c
o
n
tex
t
o
f
sc
h
izo
p
h
re
n
ia
d
iag
n
o
sis.
Th
is
st
u
d
y
a
ims
to
u
ti
li
z
e
t
h
e
tem
p
o
ra
l
a
n
d
sp
a
ti
a
l
c
h
a
ra
c
teristics
o
f
EE
G
d
a
ta
to
im
p
ro
v
e
c
las
sifica
ti
o
n
a
c
c
u
ra
c
y
.
F
o
u
r
p
o
p
u
la
r
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(CNN
)
a
rc
h
it
e
c
tu
re
s,
n
a
m
e
ly
LeN
e
t
-
5
,
Ale
x
Ne
t,
VG
G
1
6
,
a
n
d
Re
sN
e
t
-
1
8
,
a
re
a
d
a
p
ted
to
h
a
n
d
le
1
D
EE
G
sig
n
a
ls.
I
n
a
d
d
it
i
o
n
,
a
h
y
b
ri
d
a
rc
h
it
e
c
tu
re
o
f
CNN
-
g
a
ted
re
c
u
rre
n
t
u
n
i
t
(
G
RU)
a
n
d
CNN
-
lo
n
g
sh
o
r
t
-
term
m
e
m
o
ry
(
LS
TM
)
is
p
r
o
p
o
se
d
t
o
c
a
p
tu
re
s
p
a
ti
a
l
a
n
d
tem
p
o
ra
l
d
y
n
a
m
ics
.
Th
e
m
o
d
e
l
wa
s
e
v
a
lu
a
ted
o
n
a
d
a
tas
e
t
c
o
n
sistin
g
o
f
EE
G
re
c
o
rd
in
g
s
fr
o
m
1
4
p
a
ti
e
n
ts
wit
h
p
a
ra
n
o
id
sc
h
izo
p
h
re
n
i
a
a
n
d
1
4
h
e
a
lt
h
y
c
o
n
tro
ls.
Th
e
re
su
lt
s
sh
o
w
h
i
g
h
a
c
c
u
ra
c
y
a
n
d
F
1
sc
o
r
e
s
fo
r
a
l
l
m
o
d
ifi
e
d
m
o
d
e
ls,
with
CNN
-
LS
TM
a
n
d
CNN
-
G
RU
a
c
h
iev
in
g
t
h
e
h
ig
h
e
st
p
e
rfo
rm
a
n
c
e
with
sc
o
re
s
o
f
0
.
9
6
a
n
d
0
.
9
7
,
re
sp
e
c
ti
v
e
ly
.
Re
c
e
iv
e
r
o
p
e
ra
ti
n
g
c
h
a
ra
c
teristic
(ROC)
c
u
rv
e
s
d
e
m
o
n
stra
te
th
e
m
o
d
e
l'
s
a
b
il
it
y
t
o
d
i
stin
g
u
ish
b
e
twe
e
n
h
e
a
lt
h
y
c
o
n
tro
ls
a
n
d
sc
h
izo
p
h
re
n
ia
p
a
ti
e
n
ts.
Th
e
p
ro
p
o
s
e
d
m
o
d
e
l
o
ffe
rs
a
p
ro
m
isin
g
a
p
p
r
o
a
c
h
fo
r
a
u
to
m
a
ted
sc
h
izo
p
h
re
n
ia
d
iag
n
o
sis
b
a
se
d
o
n
EE
G
sig
n
a
ls,
p
o
ten
ti
a
ll
y
a
ss
is
ti
n
g
c
li
n
icia
n
s
i
n
e
a
rly
d
e
tec
ti
o
n
a
n
d
in
terv
e
n
t
io
n
.
F
u
tu
re
wo
r
k
will
fo
c
u
s
o
n
larg
e
r
d
a
ta
se
ts
a
n
d
e
x
p
lo
re
tran
sfe
r
lea
rn
in
g
tec
h
n
i
q
u
e
s t
o
imp
r
o
v
e
t
h
e
g
e
n
e
ra
li
z
a
ti
o
n
a
b
il
it
y
o
f
th
e
m
o
d
e
l.
K
ey
w
o
r
d
s
:
C
o
n
v
o
lu
tio
n
al
Dee
p
lear
n
in
g
Diag
n
o
s
is
E
lectr
o
en
ce
p
h
al
o
g
r
a
p
h
y
Neu
r
al
n
etwo
r
k
Sig
n
al
class
if
icatio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Su
k
em
i
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Sy
s
t
em
,
Facu
lty
o
f
C
o
m
p
u
ter
Scie
n
ce
,
Sriwijay
a
Un
iv
er
s
ity
Ma
s
jid
Al
Gaz
ali
St.
,
B
u
k
it L
a
m
a,
I
lir
B
ar
.
I
,
Palem
b
a
n
g
C
ity
,
So
u
th
Su
m
at
r
a
3
0
1
2
8
,
I
n
d
o
n
esia
E
m
ail: su
k
em
i@
u
n
s
r
i.a
c.
id
1.
I
NT
RO
D
UCT
I
O
N
S
c
h
i
z
o
p
h
r
e
n
i
a
i
s
a
s
e
v
e
r
e
a
n
d
o
f
t
e
n
c
h
r
o
n
i
c
p
s
y
c
h
i
a
t
r
i
c
d
i
s
o
r
d
e
r
t
h
a
t
c
a
u
s
e
s
s
u
b
s
t
a
n
t
ia
l
p
er
s
o
n
a
l
a
n
d
s
o
c
i
a
l
b
u
r
d
e
n
o
f
l
o
n
g
-
t
e
r
m
d
i
s
a
b
i
l
i
t
y
[
1
]
.
S
c
h
i
z
o
p
h
r
e
n
i
a
is
a
s
e
v
e
r
e
m
e
n
t
a
l
d
is
o
r
d
e
r
c
h
a
r
a
c
t
e
r
i
z
e
d
b
y
p
o
s
i
t
i
v
e
s
y
m
p
t
o
m
s
s
u
c
h
as
d
e
l
u
s
i
o
n
s
a
n
d
h
a
l
l
u
c
i
n
a
t
i
o
n
s
,
n
e
g
at
i
v
e
s
y
m
p
t
o
m
s
i
n
cl
u
d
i
n
g
a
m
o
t
i
v
a
t
i
o
n
a
n
d
s
o
c
i
al
w
i
t
h
d
r
a
w
al
,
a
n
d
c
o
g
n
i
ti
v
e
s
y
m
p
t
o
m
s
s
u
c
h
a
s
w
o
r
k
i
n
g
m
e
m
o
r
y
d
e
f
i
c
i
ts
a
n
d
c
o
g
n
i
ti
v
e
f
l
e
x
i
b
i
li
t
y
[
2
]
.
W
o
r
l
d
w
i
d
e
,
s
c
h
i
z
o
p
h
r
e
n
i
a
i
s
o
n
e
o
f
t
h
e
t
o
p
2
0
c
a
u
s
es
o
f
d
is
a
b
i
li
t
y
a
n
d
t
h
e
l
i
f
e
ti
m
e
p
r
e
v
a
l
e
n
ce
o
f
s
c
h
i
z
o
p
h
r
e
n
i
a
i
s
e
s
t
i
m
at
e
d
t
o
b
e
a
r
o
u
n
d
0
.
7
%
[
3
]
.
E
a
r
ly
d
e
t
e
c
ti
o
n
a
n
d
d
i
a
g
n
o
s
i
s
o
f
s
c
h
iz
o
p
h
r
e
n
i
a
is
c
h
al
l
e
n
g
i
n
g
d
u
e
t
o
t
h
e
m
a
n
y
c
o
m
o
r
b
i
d
i
t
i
e
s
as
s
o
c
ia
t
e
d
w
i
t
h
s
c
h
i
z
o
p
h
r
e
n
i
a
,
s
u
b
o
p
t
i
m
a
l
p
a
t
ie
n
t
m
a
n
a
g
e
m
e
n
t
a
n
d
p
o
t
e
n
t
i
all
y
l
i
m
it
i
n
g
p
o
s
i
t
i
v
e
o
u
t
c
o
m
e
s
[
4
]
.
T
h
e
m
a
i
n
p
r
o
b
l
e
m
i
n
d
e
t
e
c
t
i
n
g
s
c
h
iz
o
p
h
r
e
n
i
a
i
s
t
h
a
t
t
h
e
r
e
a
r
e
n
o
d
i
a
g
n
o
s
ti
c
t
e
s
ts
o
r
b
i
o
m
a
r
k
e
r
s
a
v
a
i
l
a
b
l
e
a
s
t
h
e
y
a
r
e
b
a
s
e
d
o
n
t
h
e
p
a
ti
e
n
t'
s
h
is
t
o
r
y
a
n
d
t
h
ei
r
o
v
e
r
a
l
l
m
e
n
t
a
l
s
t
a
t
e
[
5
]
.
B
ec
a
u
s
e
s
c
h
i
z
o
p
h
r
e
n
i
a
s
h
a
r
e
s
m
a
n
y
c
li
n
i
c
a
l
f
e
at
u
r
e
s
w
i
t
h
o
t
h
e
r
m
e
n
t
a
l
d
is
o
r
d
e
r
s
,
t
r
a
d
i
t
i
o
n
a
l
c
li
n
i
c
al
m
e
t
h
o
d
s
a
r
e
u
n
r
e
l
i
a
b
l
e
a
n
d
l
e
s
s
a
c
c
u
r
a
t
e
[
6
]
.
T
o
o
v
e
r
c
o
m
e
s
u
ch
l
i
m
i
t
at
i
o
n
s
,
a
n
a
u
t
o
m
a
t
e
d
,
r
el
i
a
b
l
e
a
n
d
r
e
p
r
o
d
u
c
i
b
l
e
a
p
p
r
o
ac
h
o
f
b
r
a
i
n
i
m
a
g
i
n
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
2
,
Ap
r
il 2
0
2
5
:
1
3
1
1
-
1
3
2
5
1312
m
o
d
a
l
i
t
i
es
is
r
e
q
u
i
r
e
d
[
7
]
.
A
l
t
h
o
u
g
h
s
c
h
i
z
o
p
h
r
e
n
i
a
is
u
s
u
al
l
y
d
e
t
e
c
t
e
d
b
y
s
p
e
c
i
al
i
s
t
s
,
i
n
r
e
c
en
t
t
i
m
es
,
t
o
o
ls
s
u
c
h
a
s
el
e
c
t
r
o
e
n
c
e
p
h
a
l
o
g
r
a
p
h
y
(
E
E
G
)
h
a
v
e
b
e
e
n
u
s
e
d
t
o
a
u
t
o
m
a
t
ic
a
l
l
y
d
e
te
c
t
s
c
h
i
z
o
p
h
r
e
n
i
a
.
As
E
E
G
is
c
h
e
a
p
e
r
a
n
d
m
o
r
e
p
r
a
c
t
i
c
a
l
,
t
h
e
u
s
e
o
f
E
E
G
i
n
s
c
h
i
z
o
p
h
r
e
n
i
a
d
e
te
c
t
i
o
n
is
wi
d
e
l
y
p
r
e
f
e
r
r
e
d
[
8
]
.
EEG
/
m
ag
n
eto
en
ce
p
h
alo
g
r
ap
h
y
(
ME
G)
ca
n
b
e
u
s
ed
f
o
r
n
o
n
-
in
v
asiv
e
s
tu
d
y
o
f
b
r
ain
elec
tr
ical
ac
tiv
ity
.
Scalp
p
o
te
n
tial
d
if
f
er
en
ce
s
f
r
o
m
elec
tr
ic
f
ield
s
d
r
i
v
en
b
y
n
e
u
r
al
cu
r
r
en
ts
ar
e
m
ea
s
u
r
ed
u
s
in
g
E
E
G
[
9
]
.
E
E
G
is
a
p
h
y
s
io
lo
g
ical
tech
n
iq
u
e
th
at
r
e
co
r
d
s
s
p
o
n
tan
eo
u
s
elec
tr
ical
b
r
ain
ac
tiv
it
y
o
r
ig
in
atin
g
f
r
o
m
n
eu
r
o
n
s
with
h
ig
h
r
eso
lu
tio
n
t
h
r
o
u
g
h
elec
tr
o
d
es
co
n
n
ec
te
d
t
o
th
e
s
ca
lp
[
1
0
]
.
Ho
we
v
er
,
E
E
G
is
a
s
ig
n
al
with
a
v
er
y
s
m
all
a
m
p
litu
d
e;
t
h
er
ef
o
r
e,
r
ec
o
g
n
izin
g
em
o
tio
n
s
f
r
o
m
E
E
G
is
a
v
er
y
ch
allen
g
in
g
task
.
Nev
er
th
eless
,
m
an
y
r
esear
ch
e
r
s
h
av
e
attem
p
ted
to
o
v
er
c
o
m
e
th
is
p
r
o
b
lem
b
y
ad
o
p
tin
g
ad
v
an
ce
d
tec
h
n
iq
u
es,
in
clu
d
in
g
d
ee
p
lear
n
in
g
[
1
1
]
.
Un
lik
e
o
t
h
er
cl
ass
if
icatio
n
m
eth
o
d
s
,
d
ee
p
lear
n
in
g
d
o
es
n
o
t
r
eq
u
ir
e
a
s
ep
ar
ate
alg
o
r
ith
m
to
m
an
u
ally
ex
tr
ac
t
f
ea
tu
r
es
f
r
o
m
d
ata.
T
h
e
u
s
e
o
f
d
ee
p
lear
n
i
n
g
h
as
in
cr
ea
s
ed
d
u
e
to
its
au
to
m
atic
ex
tr
ac
tio
n
o
f
d
esire
d
f
ea
tu
r
es a
n
d
its
g
o
o
d
p
er
f
o
r
m
a
n
ce
in
class
if
icatio
n
[
1
2
]
.
R
ec
en
tly
,
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
h
as
r
ec
eiv
ed
a
lo
t
o
f
atten
tio
n
an
d
ac
h
i
ev
ed
g
r
ea
t
s
u
cc
ess
in
th
e
v
is
u
al
f
ield
d
u
e
to
its
ab
ilit
y
to
au
to
m
atica
lly
ex
tr
ac
t
s
tr
o
n
g
f
ea
tu
r
es
[
1
3
]
.
T
h
e
o
u
tp
u
t
o
f
E
E
G
s
ig
n
al
is
tim
e
s
er
ies
d
ata,
th
er
ef
o
r
e,
C
NN
ca
n
a
u
to
m
atica
lly
d
is
co
v
er
a
n
d
e
x
tr
ac
t
th
e
in
ter
n
al
s
tr
u
ctu
r
e
o
f
th
e
in
p
u
t
tim
e
s
er
ies
to
g
en
er
ate
d
ee
p
f
ea
tu
r
es
f
o
r
class
if
icatio
n
[
1
4
]
.
Var
io
u
s
ty
p
es
o
f
C
N
N
ar
ch
itectu
r
es
h
av
e
also
b
ee
n
im
p
lem
en
ted
i
n
v
ar
io
u
s
E
E
G
s
ig
n
al
class
if
icatio
n
s
y
s
tem
s
.
Acc
o
r
d
in
g
Yıld
ır
ı
m
et
a
l.
[
1
5
]
,
a
n
ew
one
-
d
im
e
n
s
io
n
al
(
1D
)
C
NN
m
o
d
el
is
p
r
o
p
o
s
ed
f
o
r
au
to
m
atic
r
ec
o
g
n
itio
n
o
f
n
o
r
m
al
an
d
ab
n
o
r
m
al
E
E
G
s
ig
n
als.
T
h
e
d
ev
elo
p
ed
m
o
d
el
r
esu
lted
in
o
n
ly
2
0
.
6
6
%
m
is
class
if
icatio
n
r
ate
in
clas
s
if
y
in
g
n
o
r
m
al
a
n
d
ab
n
o
r
m
al
E
E
G
s
ig
n
als.
T
h
e
m
o
d
el
p
r
o
p
o
s
ed
b
y
Oh
et
a
l
.
[
1
6
]
r
esu
lted
in
class
if
icatio
n
ac
cu
r
ac
ies
o
f
9
8
.
0
7
%
an
d
8
1
.
2
6
% f
o
r
n
o
n
-
s
u
b
ject
-
b
ased
test
in
g
an
d
s
u
b
ject
-
b
ased
test
in
g
,
r
esp
ec
tiv
ely
.
I
n
th
e
m
eth
o
d
p
r
o
p
o
s
ed
b
y
Sh
ey
k
h
iv
a
n
d
et
a
l.
[
1
7
]
,
r
aw
E
E
G
s
ig
n
als ar
e
ap
p
lied
to
C
NN
-
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
n
etwo
r
k
s
,
with
o
u
t
in
v
o
lv
in
g
f
ea
tu
r
e
ex
tr
ac
tio
n
/s
elec
tio
n
.
Simu
la
tio
n
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
f
o
r
two
-
s
tag
e
class
if
icatio
n
(
n
eg
ativ
e
an
d
p
o
s
itiv
e)
an
d
th
r
ee
-
s
tag
e
class
if
icatio
n
(
n
eg
ativ
e,
n
e
u
tr
al
,
an
d
p
o
s
itiv
e)
o
f
em
o
tio
n
s
f
o
r
1
2
ac
tiv
e
c
h
an
n
els s
h
o
wed
ac
cu
r
ac
ies o
f
9
7
.
4
2
% a
n
d
9
6
.
7
8
%
an
d
Kap
p
a
c
o
ef
f
icien
ts
o
f
0
.
9
4
a
n
d
0
.
9
3
,
r
esp
ec
tiv
ely
.
Sh
o
eib
i
et
a
l
[
1
8
]
u
tili
ze
d
v
ar
io
u
s
in
tellig
en
t
d
ee
p
lear
n
in
g
b
ased
m
eth
o
d
s
f
o
r
a
u
to
m
atic
s
ch
izo
p
h
r
e
n
ia
d
iag
n
o
s
is
th
r
o
u
g
h
E
E
G
s
ig
n
als.
C
l
ass
if
icatio
n
o
f
E
E
G
s
ig
n
als
was
f
ir
s
t
p
er
f
o
r
m
ed
b
y
co
n
v
en
tio
n
al
m
ac
h
in
e
l
ea
r
n
in
g
m
eth
o
d
s
,
e.
g
.
,
s
u
p
p
o
r
t
v
ec
t
o
r
m
ac
h
in
e
,
k
-
n
ea
r
est
n
eig
h
b
o
r
,
d
ec
is
io
n
tr
ee
,
n
aïv
e
B
ay
es,
r
an
d
o
m
f
o
r
est,
h
ig
h
ly
r
an
d
o
m
ized
tr
ee
s
,
an
d
b
ag
g
in
g
.
Var
io
u
s
d
ee
p
lear
n
i
n
g
m
o
d
els
wer
e
p
r
o
p
o
s
ed
,
n
am
el
y
,
L
STM
,
C
NN,
an
d
C
NN
-
L
STM
.
T
h
e
r
esu
lts
s
h
o
w
th
at
C
NN
-
L
STM
h
as
th
e
b
est
p
er
f
o
r
m
a
n
ce
.
T
h
e
p
r
o
p
o
s
ed
C
NN
-
L
STM
m
o
d
el
h
as
ac
h
iev
ed
an
ac
cu
r
ac
y
p
er
ce
n
tag
e
o
f
9
9
.
2
5
%,
b
etter
t
h
an
p
r
e
v
io
u
s
m
ac
h
in
e
lear
n
in
g
m
o
d
els.
Acc
o
r
d
in
g
to
Yıld
ır
ım
et
a
l.
[
1
5
]
,
a
n
ew
1
D
-
C
NN
m
o
d
el
was
p
r
o
p
o
s
ed
f
o
r
a
u
to
m
atic
r
ec
o
g
n
itio
n
o
f
n
o
r
m
al
an
d
ab
n
o
r
m
al
E
E
G
s
ig
n
als
with
9
7
%
ac
cu
r
ac
y
.
R
esear
ch
b
y
Sh
ey
k
h
iv
a
n
d
et
a
l.
[
1
7
]
r
ec
o
r
d
ed
E
E
G
s
ig
n
als
f
r
o
m
1
4
s
u
b
jects
with
m
u
s
ical
s
tim
u
latio
n
f
o
r
th
e
p
r
o
ce
s
s
.
Simu
latio
n
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
f
o
r
two
-
s
tag
e
class
i
f
icatio
n
(
n
eg
ativ
e
an
d
p
o
s
itiv
e)
an
d
th
r
ee
-
s
tag
e
class
if
icatio
n
(
n
e
g
ativ
e,
n
eu
tr
al
,
an
d
p
o
s
itiv
e)
o
f
em
o
tio
n
s
f
o
r
1
2
ac
tiv
e
ch
a
n
n
els
s
h
o
wed
ac
cu
r
ac
ies
o
f
9
7
.
4
2
%
an
d
9
6
.
7
8
%
,
an
d
Kap
p
a
co
ef
f
icien
ts
o
f
0
.
9
4
an
d
0
.
9
3
,
r
esp
ec
tiv
ely
.
R
esear
ch
b
y
Na
g
ab
u
s
h
an
am
et
a
l.
[
1
9
]
an
aly
ze
d
d
ee
p
lear
n
in
g
f
o
r
E
E
G
s
ig
n
al
class
if
icatio
n
.
T
h
e
im
p
r
o
v
ed
L
STM
an
d
n
eu
r
a
l
n
etwo
r
k
wer
e
p
r
o
p
o
s
ed
f
o
r
b
etter
p
er
f
o
r
m
an
c
e
with
7
1
.
3
%
an
d
7
8
.
9
%
ac
cu
r
ac
y
in
E
E
G
class
if
ica
tio
n
,
r
esp
ec
tiv
ely
.
with
co
m
p
ar
ativ
e
ex
p
er
im
en
ts
,
u
s
in
g
Op
en
B
C
I
to
co
llect
E
E
G
ac
ti
o
n
id
ea
s
d
u
r
in
g
s
tatic
ac
tio
n
an
d
d
y
n
am
ic
ac
tio
n
an
d
u
s
in
g
C
o
n
v
1
D
-
GR
U
E
E
G
r
ec
o
g
n
itio
n
m
o
d
el
t
o
tr
ain
a
n
d
r
ec
o
g
n
ize
ac
tio
n
s
,
r
esp
ec
tiv
ely
.
T
h
e
ex
p
er
im
en
tal
r
esu
lts
s
h
o
w
th
at
th
e
b
r
ain
wav
e
ac
tio
n
id
ea
is
ea
s
ier
to
r
ec
o
g
n
ize
in
th
e
s
tatic
s
tate.
T
h
e
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
o
f
th
e
b
r
ain
wa
v
e
ac
tio
n
id
ea
in
th
e
d
y
n
am
ic
s
tate
is
o
n
ly
7
2
.
2
7
%,
an
d
th
e
r
ec
o
g
n
iti
o
n
ac
cu
r
ac
y
o
f
th
e
b
r
ain
wav
e
ac
tio
n
id
ea
in
t
h
e
s
tatic
s
ta
te
is
9
9
.
9
8
%.
An
elev
en
-
lay
er
C
NN
m
o
d
el
is
p
r
o
p
o
s
ed
to
d
etec
t
s
ch
izo
p
h
r
en
ia
u
s
in
g
E
E
G
s
ig
n
als.
Hig
h
class
if
icatio
n
ac
cu
r
ac
ie
s
o
f
9
8
.
0
7
%
an
d
8
1
.
2
6
%
w
er
e
o
b
tain
e
d
f
o
r
n
o
n
-
s
u
b
ject
-
b
ased
test
in
g
a
n
d
s
u
b
ject
-
b
ased
test
in
g
,
r
esp
ec
ti
v
ely
,
alth
o
u
g
h
t
h
e
d
ata
s
et
was sm
all
[
1
6
]
.
I
n
th
is
p
ap
er
,
we
will
test
th
e
d
iag
n
o
s
is
o
f
E
E
G
s
ig
n
als
to
p
r
ed
ict
s
ch
izo
p
h
r
en
ia
d
is
ea
s
e
u
s
in
g
C
NN
m
eth
o
d
s
with
ex
is
tin
g
ar
ch
itectu
r
es
(
L
eNe
t
-
5
,
Alex
Net,
VGG
-
1
6
,
an
d
R
esNet
-
1
8
)
m
o
d
if
ied
to
b
e
ad
ap
ted
to
1
D
E
E
G
s
ig
n
al
d
atasets
.
T
h
e
n
ex
t
C
NN
m
o
d
els
ar
e
C
NN
-
L
STM
an
d
C
NN
-
g
ated
r
ec
u
r
r
en
t
u
n
it
(
GR
U
)
.
T
h
en
co
m
p
ar
e
t
h
e
ac
cu
r
ac
y
o
f
ea
ch
ar
ch
itectu
r
e
u
s
ed
an
d
an
aly
ze
th
e
in
cr
ea
s
e
an
d
d
ec
r
e
ase
in
ac
cu
r
ac
y
i
n
class
if
y
in
g
E
E
G
s
ig
n
als to
d
iag
n
o
s
e
p
eo
p
le
with
s
ch
izo
p
h
r
e
n
ia.
2.
M
E
T
H
O
D
2
.
1
.
E
E
G
s
ig
na
l
E
E
G
is
a
n
o
n
-
in
v
asiv
e
m
ea
s
u
r
em
en
t
o
f
th
e
b
r
ai
n
'
s
elec
tr
i
c
f
ield
s
.
E
lectr
o
d
es
p
lace
d
o
n
th
e
s
ca
lp
r
ec
o
r
d
v
o
ltag
e
p
o
te
n
tials
r
esu
ltin
g
f
r
o
m
cu
r
r
en
t
f
lo
w
in
a
n
d
ar
o
u
n
d
n
eu
r
o
n
s
[
2
0
]
.
E
E
G
is
a
m
ea
s
u
r
e
o
f
th
e
elec
tr
ic
f
ield
s
g
en
er
ated
b
y
an
ac
tiv
e
b
r
ain
,
is
a
b
r
ain
m
ap
p
in
g
an
d
n
eu
r
o
im
ag
in
g
tech
n
iq
u
e
wid
ely
u
s
ed
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
cc
u
r
a
cy
o
f n
e
u
r
a
l n
etw
o
r
ks in
b
r
a
in
w
a
ve
d
ia
g
n
o
s
is
o
f sch
i
z
o
p
h
r
en
ia
(
S
u
ke
mi)
1313
an
d
o
u
ts
id
e
th
e
clin
ical
d
o
m
ain
[
2
1
]
.
E
E
G
is
wid
ely
u
s
ed
in
r
esear
ch
in
v
o
lv
in
g
n
eu
r
al
en
g
in
ee
r
in
g
,
n
eu
r
o
s
cien
ce
,
an
d
b
io
m
e
d
ica
l
en
g
in
ee
r
in
g
.
T
h
e
o
u
t
p
u
t
o
f
an
E
E
G
s
ig
n
al
is
tim
e
s
er
i
es
d
ata.
T
h
er
ef
o
r
e,
co
m
p
ar
in
g
th
e
o
u
tp
u
t
d
ata
wit
h
th
e
s
tan
d
ar
d
E
E
G
s
ig
n
al
ca
n
d
eter
m
in
e
t
h
e
d
is
ea
s
e
o
r
p
r
o
b
lem
[
2
2
]
.
T
h
e
E
E
G
s
ig
n
al
v
is
u
aliza
tio
n
in
th
is
s
tu
d
y
r
ep
r
esen
ts
d
ata
r
ec
o
r
d
e
d
o
v
er
a
d
u
r
atio
n
o
f
8
0
0
to
9
0
0
s
ec
o
n
d
s
,
as
in
d
icate
d
b
y
th
e
r
an
g
e
o
n
th
e
X
-
ax
is
o
f
th
e
g
r
ap
h
.
T
h
e
E
E
G
s
ig
n
als
co
n
s
is
t
o
f
1
9
ch
an
n
els
r
ec
o
r
d
ed
f
r
o
m
s
tan
d
ar
d
elec
tr
o
d
e
lo
ca
tio
n
s
o
n
th
e
s
ca
lp
,
with
th
e
Y
-
ax
is
r
ep
r
esen
tin
g
ea
ch
E
E
G
c
h
an
n
el.
E
ac
h
s
ig
n
al
r
ef
lects
th
e
b
r
ain
'
s
elec
tr
ical
ac
tiv
ity
,
wit
h
v
a
r
y
in
g
p
atter
n
s
an
d
am
p
litu
d
es
d
e
p
en
d
in
g
o
n
th
e
c
o
n
d
it
io
n
o
f
th
e
s
u
b
ject.
T
h
e
s
ig
n
als
ar
e
d
is
p
lay
ed
wi
th
a
s
ep
ar
atio
n
o
f
f
s
et
to
d
is
tin
g
u
is
h
in
d
iv
id
u
al
ch
an
n
els
clea
r
ly
.
T
h
e
y
-
ax
is
v
alu
es,
r
an
g
in
g
f
r
o
m
-
5
0
0
to
4
,
5
0
0
,
ar
e
lik
ely
n
u
m
er
ical
r
e
p
r
esen
tatio
n
s
o
f
th
e
E
E
G
s
ig
n
al
am
p
litu
d
es.
T
h
is
v
is
u
aliza
tio
n
s
ty
le
is
ty
p
ical
in
MA
T
L
AB
,
wh
er
e
th
e
s
ca
le
a
d
ap
ts
au
to
m
atica
lly
b
ased
o
n
t
h
e
s
ig
n
al
v
alu
es
to
f
ac
ilit
ate
clea
r
o
b
s
er
v
atio
n
o
f
tem
p
o
r
al
p
atter
n
s
an
d
v
a
r
iatio
n
s
with
in
th
e
E
E
G
d
ata.
Fig
u
r
e
1
s
h
o
ws
E
E
G
Sig
n
als
f
r
o
m
n
o
r
m
al
s
u
b
ject
(
Fig
u
r
e
1
(
a)
)
an
d
p
atien
ts
with
s
ch
izo
p
h
r
en
ia
(
Fig
u
r
e
1
(
b
)
)
.
(
a)
(
b
)
Fig
u
r
e
1
.
E
E
G
s
ig
n
als f
r
o
m
(
a
)
n
o
r
m
al
s
u
b
ject
an
d
(
b
)
p
atien
ts
with
s
ch
izo
p
h
r
en
ia
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
2
,
Ap
r
il 2
0
2
5
:
1
3
1
1
-
1
3
2
5
1314
2
.
2
.
Co
nv
o
lutio
na
l
neura
l net
wo
rk
T
h
e
ter
m
d
ee
p
lear
n
in
g
o
r
d
e
ep
n
eu
r
al
n
etwo
r
k
r
e
f
er
s
to
an
ar
tific
ial
n
eu
r
al
n
etwo
r
k
(
A
NN)
[
2
3
]
with
m
u
ltip
le
lay
er
s
.
On
e
o
f
th
e
m
o
s
t
p
o
p
u
lar
d
ee
p
n
eu
r
al
n
etwo
r
k
s
is
th
e
C
NN
.
On
e
o
f
th
e
m
o
s
t
p
o
p
u
lar
d
ee
p
n
eu
r
al
n
etwo
r
k
s
is
th
e
C
NN.
T
h
e
m
o
s
t
b
en
ef
icial
a
s
p
ec
t
o
f
C
NN
is
th
at
it
r
e
d
u
ce
s
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
in
th
e
ANN.
T
h
e
m
o
s
t
im
p
o
r
tan
t
ass
u
m
p
tio
n
ab
o
u
t
th
e
p
r
o
b
lem
s
o
lv
ed
b
y
C
NN
s
h
o
u
ld
n
o
t
h
av
e
s
p
atially
d
ep
en
d
en
t
f
ea
t
u
r
es
[
2
4
]
.
C
NN
ar
ch
itectu
r
e
is
in
s
p
ir
ed
b
y
v
is
u
al
p
e
r
ce
p
tio
n
[
2
5
]
.
On
e
o
f
th
e
m
ain
d
if
f
er
en
ce
s
is
th
at
th
e
n
eu
r
o
n
lay
er
s
in
a
C
NN
co
n
s
is
t
o
f
n
e
u
r
o
n
s
o
r
g
an
ized
in
to
th
r
ee
d
i
m
en
s
io
n
s
,
th
e
in
p
u
t
s
p
atial
d
im
en
s
io
n
s
(
h
eig
h
t a
n
d
wid
th
)
an
d
d
ep
th
[
2
6
]
.
C
NNs
ar
e
g
en
er
ally
d
esig
n
ed
to
o
p
er
ate
ex
clu
s
iv
ely
o
n
2
D
d
ata
s
u
ch
as
im
ag
es
an
d
v
i
d
e
o
s
.
T
h
is
is
wh
y
th
ey
a
r
e
o
f
ten
r
ef
er
r
ed
to
as,
''
2
D
C
NN
s
''
.
Alter
n
ativ
ely
,
a
m
o
d
if
ied
v
e
r
s
io
n
o
f
2
D
C
NNs
ca
lled
1
D
C
NNs
h
as
r
ec
en
tly
b
ee
n
d
ev
elo
p
e
d
.
I
n
t
h
e
r
ec
en
t
s
tu
d
ies
m
en
tio
n
ed
ab
o
v
e,
co
m
p
a
ct
1
D
C
NNs
h
av
e
s
h
o
wn
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
o
n
ap
p
licatio
n
s
th
at
h
a
v
e
lim
ite
d
lab
eled
d
ata
a
n
d
h
ig
h
s
ig
n
al
v
ar
iatio
n
o
b
tain
e
d
f
r
o
m
d
if
f
e
r
en
t
s
o
u
r
ce
s
(
i.e
.
,
p
atien
t
E
E
G,
ci
v
il,
m
ec
h
an
ical
o
r
ae
r
o
s
p
ac
e
s
tr
u
ct
u
r
es,
h
ig
h
p
o
wer
cir
c
u
its
,
an
d
en
g
in
e
o
r
m
o
to
r
p
o
wer
)
as sh
o
wn
in
Fig
u
r
e
2
[
2
7
]
.
Fig
u
r
e
2
.
E
x
am
p
le
o
f
1
D
C
NN
ar
ch
itectu
r
e
w
ith
3
C
o
n
v
1
D
an
d
2
d
en
s
e
lay
er
s
[
2
7
]
2
.
3
.
L
o
ng
s
ho
rt
-
t
er
m
m
emo
ry
R
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
(
R
NNs)
ar
e
a
g
r
o
u
p
o
f
d
ee
p
lear
n
in
g
m
o
d
els
u
s
ed
in
s
p
ee
ch
r
e
co
g
n
itio
n
,
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
,
an
d
b
i
o
m
ed
ical
s
ig
n
al
p
r
o
ce
s
s
in
g
.
C
NN
m
o
d
els
ar
e
o
f
th
e
f
ee
d
-
f
o
r
war
d
ty
p
e.
Ho
wev
er
,
R
NNs
h
av
e
a
f
ee
d
b
ac
k
lay
er
,
w
h
er
e
th
e
n
etwo
r
k
o
u
tp
u
t
r
etu
r
n
s
to
th
e
n
etwo
r
k
alo
n
g
with
th
e
n
ex
t
in
p
u
t.
Sin
ce
it
h
as
an
in
ter
n
al
m
em
o
r
y
,
t
h
e
R
NN
r
em
em
b
e
r
s
p
r
ev
io
u
s
in
p
u
ts
an
d
u
s
es
th
em
to
p
r
o
ce
s
s
th
e
in
p
u
t
s
eq
u
en
ce
.
Simp
le
R
NNs,
L
STM
s
,
an
d
GR
U
n
etwo
r
k
s
ar
e
th
r
ee
im
p
o
r
tan
t
g
r
o
u
p
s
o
f
R
NNs
[
1
8
]
.
T
h
e
s
p
ec
ialized
m
em
o
r
y
ce
ll
ar
c
h
i
tectu
r
e
in
L
STM
s
m
ak
es
it
ea
s
y
to
s
to
r
e
in
f
o
r
m
atio
n
f
o
r
lo
n
g
p
e
r
io
d
s
o
f
tim
e.
T
h
e
ce
ll
s
tr
u
ctu
r
e
h
as
b
ee
n
m
o
d
if
ied
b
y
m
an
y
p
e
o
p
le
s
in
ce
th
en
,
b
u
t
th
e
s
tan
d
a
r
d
f
o
r
m
u
latio
n
o
f
a
s
in
g
le
L
STM
ce
ll c
an
b
e
g
iv
e
n
b
y
(
1
)
-
(
6
)
[
2
8
]
.
=
(
∙
[
ℎ
−
1
,
]
+
)
(
1
)
=
(
∙
[
ℎ
−
1
,
]
+
)
(
2
)
~
=
ℎ
(
∙
[
ℎ
−
1
,
]
+
)
(
3
)
=
∗
(
−
1
)
+
∗
~
(
4
)
=
(
∙
[
ℎ
−
1
,
]
+
)
(
5
)
ℎ
=
∗
ℎ
(
)
(
6
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
cc
u
r
a
cy
o
f n
e
u
r
a
l n
etw
o
r
ks in
b
r
a
in
w
a
ve
d
ia
g
n
o
s
is
o
f sch
i
z
o
p
h
r
en
ia
(
S
u
ke
mi)
1315
R
NNs
ar
e
f
o
u
n
d
to
b
e
an
e
f
f
e
ctiv
e
to
o
l
f
o
r
a
p
p
r
o
x
im
atin
g
d
y
n
am
ic
s
y
s
tem
s
th
at
d
ea
l
with
tim
e
an
d
s
eq
u
en
ce
-
d
e
p
en
d
e
n
t
d
ata
s
u
ch
as
v
id
eo
,
a
u
d
io
,
an
d
o
th
er
s
.
L
STM
is
p
ar
t
o
f
an
R
NN
wit
h
s
tate
m
em
o
r
y
an
d
m
u
ltil
ay
er
ce
ll
s
tr
u
ctu
r
e.
Har
d
war
e
ac
ce
ler
atio
n
o
f
L
STM
s
u
s
in
g
m
em
r
is
to
r
cir
cu
its
is
an
em
er
g
in
g
to
p
ic
o
f
s
tu
d
y
.
Fig
u
r
e
3
s
h
o
ws
th
e
ex
am
p
le
o
f
L
STM
,
Fig
u
r
e
3
(
a)
s
h
o
ws
th
e
e
x
am
p
le
o
f
o
r
ig
in
al
L
STM
u
n
it
ar
ch
itectu
r
e:
m
em
o
r
y
ce
ll
a
n
d
two
g
ates
,
Fig
u
r
e
3
(
b
)
s
h
o
ws
L
STM
ce
ll
with
f
o
r
g
et
g
ate
,
an
d
Fig
u
r
e
3
(
c
)
s
h
o
ws
m
o
d
er
n
r
ep
r
esen
tatio
n
o
f
L
STM
with
f
o
r
g
et
g
ate
[
2
9
]
.
(
a)
(
b
)
(
c)
Fig
u
r
e
3
.
T
y
p
e
o
f
L
STM
ar
ch
i
tectu
r
e
:
(
a)
e
x
a
m
p
le
o
f
o
r
ig
i
n
a
l L
STM
u
n
it a
r
ch
itectu
r
e:
m
e
m
o
r
y
ce
ll a
n
d
two
g
ates
,
(
b
)
L
STM
ce
ll with
f
o
r
g
et
g
ate
,
an
d
(
c)
m
o
d
er
n
r
ep
r
es
en
tatio
n
o
f
L
STM
with
f
o
r
g
et
g
ate
[
2
9
]
2
.
4
.
G
a
t
ed
re
curr
ent
un
it
I
n
tr
o
d
u
ce
d
in
2
0
1
4
,
GR
Us
ar
e
s
im
ilar
to
L
STM
s
b
u
t
h
av
e
f
ewe
r
p
ar
am
eter
s
.
T
h
e
y
also
h
av
e
g
ated
u
n
its
lik
e
L
STM
s
th
at
co
n
tr
o
l
th
e
f
lo
w
o
f
i
n
f
o
r
m
atio
n
with
in
th
e
u
n
it
b
u
t
with
o
u
t
h
av
in
g
s
ep
ar
ate
m
em
o
r
y
ce
lls
.
Un
lik
e
L
STM
s
,
GR
U
s
d
o
n
o
t
h
av
e
o
u
tp
u
t
g
ates,
th
u
s
d
is
p
lay
in
g
th
ei
r
f
u
ll
c
o
n
ten
t.
T
h
e
f
o
r
m
u
latio
n
o
f
GR
U
ca
n
b
e
g
iv
en
b
y
(
7
)
-
(
9
)
[
2
8
]
.
=
(
+
ℎ
ℎ
−
1
+
)
(
7
)
=
(
+
ℎ
ℎ
−
1
+
)
(
8
)
ℎ
=
ℎ
(
ℎ
+
ℎℎ
(
⊙
ℎ
−
1
)
+
ℎ
)
(
9
)
ℎ
=
⊙
ℎ
−
1
+
(
1
−
)
⊙
ℎ
(
1
0
)
GR
U
is
a
p
o
p
u
lar
n
e
u
r
al
n
et
wo
r
k
n
o
wad
ay
s
,
esp
ec
ially
f
e
ed
-
f
o
r
war
d
n
etwo
r
k
s
,
m
em
o
r
y
n
etwo
r
k
s
,
an
d
g
r
a
p
h
n
etwo
r
k
s
.
I
n
a
f
ee
d
-
f
o
r
war
d
n
etwo
r
k
,
ea
c
h
n
e
u
r
o
n
is
o
n
a
d
if
f
er
e
n
t
lay
er
.
E
a
ch
lay
er
o
f
n
e
u
r
o
n
s
ca
n
g
et
s
ig
n
als
f
r
o
m
th
e
p
r
ev
i
o
u
s
lay
er
o
f
n
eu
r
o
n
s
an
d
o
u
tp
u
t
s
ig
n
als
to
th
e
n
ex
t
lay
er
.
T
h
e
o
u
tp
u
t
is
n
o
t
o
n
ly
r
elate
d
to
th
e
c
u
r
r
e
n
t
in
p
u
t
an
d
th
e
n
etwo
r
k
weig
h
ts
b
u
t
also
r
elate
d
to
th
e
p
r
ev
io
u
s
in
p
u
t
s
.
I
t
is
m
ain
ly
u
s
ed
in
th
e
f
ield
s
o
f
tim
e
s
er
ies
d
ata
an
d
tex
t
d
ata
[
3
0
]
.
GR
U
h
as
a
less
co
m
p
licated
s
tr
u
ctu
r
e
co
m
p
ar
ed
t
o
L
STM
.
I
t
h
as
n
o
o
u
tp
u
t
g
ates
b
u
t
h
as
an
u
p
d
ate
z
an
d
a
r
eset
g
ate
r
.
T
h
ese
g
ates
ar
e
v
ec
to
r
s
th
at
d
ec
id
e
wh
at
in
f
o
r
m
atio
n
s
h
o
u
ld
b
e
p
ass
ed
to
th
e
o
u
tp
u
t.
T
h
e
r
eset
g
ate
d
ef
in
es
h
o
w
to
co
m
b
in
e
th
e
n
ew
in
p
u
t
with
th
e
p
r
ev
io
u
s
m
e
m
o
r
y
.
T
h
e
d
ef
in
iti
o
n
o
f
h
o
w
m
u
ch
o
f
th
e
last
m
e
m
o
r
y
s
to
r
e
d
is
d
o
n
e
b
y
th
e
u
p
d
ate
[
3
1
]
.
Fig
u
r
e
4
s
h
o
ws e
x
am
p
le
d
iag
r
a
m
o
f
h
o
w
th
e
GR
U
wo
r
k
s
.
2
.
5
.
P
r
o
po
s
ed
a
rc
hite
ct
ure
T
h
e
C
NN
-
1
D
s
tr
u
ctu
r
e
u
s
ed
in
th
is
r
esear
ch
is
an
im
p
lem
en
tatio
n
o
f
th
e
ex
is
tin
g
C
NN
-
2
D
ar
ch
itectu
r
e
with
th
e
ad
d
itio
n
o
f
p
ar
am
eter
s
s
o
th
at
it
ca
n
wo
r
k
well
in
p
r
o
ce
s
s
in
g
1
D
d
ata
in
th
e
f
o
r
m
o
f
s
ch
izo
p
h
r
en
ia
E
E
G
s
ig
n
als.
T
h
e
C
NN
ar
ch
itectu
r
es
in
clu
d
e
VGG
-
1
6
,
R
esNet
-
1
8
,
L
eNe
t
-
5
,
an
d
Alex
Net.
T
h
e
s
tr
u
ctu
r
e
ar
r
a
n
g
em
e
n
t
is
b
y
r
ep
lacin
g
th
e
2
D
co
n
v
o
lu
tio
n
a
l
lay
er
with
a
1
D
co
n
v
o
lu
tio
n
al
lay
er
.
Seco
n
d
ly
,
th
e
p
o
o
lin
g
lay
er
m
u
s
t
also
b
e
r
ep
lace
d
with
a
1
D
p
o
o
lin
g
lay
er
.
T
h
ir
d
,
th
e
f
u
lly
c
o
n
n
ec
t
ed
lay
er
s
h
o
u
ld
b
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
2
,
Ap
r
il 2
0
2
5
:
1
3
1
1
-
1
3
2
5
1316
r
em
o
v
ed
an
d
r
ep
lace
d
with
a
1
D
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
lay
er
.
T
h
e
o
u
tp
u
t
la
y
er
also
n
ee
d
s
to
b
e
m
o
d
if
ie
d
to
m
atch
th
e
n
u
m
b
er
o
f
class
es
u
s
ed
f
o
r
class
if
icatio
n
.
T
h
e
n
u
m
b
er
o
f
f
ilter
s
in
ea
ch
co
n
v
o
lu
tio
n
al
lay
er
is
r
ed
u
ce
d
a
n
d
ea
c
h
lay
er
is
in
cr
ea
s
ed
with
d
r
o
p
o
u
t
to
p
r
ev
en
t
o
v
er
f
itti
n
g
.
T
h
ese
m
o
d
if
icatio
n
s
,
b
y
m
ain
tain
in
g
th
e
b
asic
s
tr
u
ctu
r
e
o
f
th
e
C
NN
-
2
D
ar
ch
itectu
r
e
w
h
ile
a
d
ap
tin
g
it
to
wo
r
k
with
1
D
d
ata
s
u
ch
as
th
e
s
ch
izo
p
h
r
en
ia
E
E
G
d
ataset.
T
o
o
p
tim
ize
th
e
L
eNe
t
-
5
ar
ch
itectu
r
e
to
w
o
r
k
with
1
D
d
ata
,
s
u
ch
as
E
E
G
s
ig
n
al
d
ata,
th
e
o
r
ig
in
al
2
D
co
n
v
o
l
u
tio
n
al
lay
er
(
C
o
n
v
2
D)
is
r
ep
lace
d
with
a
1
D
co
n
v
o
l
u
tio
n
al
lay
er
(
C
o
n
v
1
D
)
.
T
h
e
in
p
u
t
_
s
h
ap
e
p
ar
am
eter
is
s
et
to
(
n
_
tim
estep
s
,
n
_
f
ea
tu
r
es),
wh
er
e
n
_
tim
e
s
tep
s
is
th
e
len
g
th
o
f
th
e
tim
e
s
er
ies
(
in
th
is
ca
s
e,
th
e
n
u
m
b
er
o
f
E
E
G
s
am
p
les)
an
d
n
_
f
ea
tu
r
es
is
th
e
n
u
m
b
er
o
f
ch
a
n
n
els
(
in
th
is
ca
s
e,
th
e
n
u
m
b
er
o
f
elec
tr
o
d
es).
T
h
e
p
o
o
lin
g
lay
er
(
Ma
x
Po
o
lin
g
1
D)
is
also
m
o
d
if
ied
t
o
wo
r
k
with
1
D
d
ata,
an
d
th
e
f
u
lly
co
n
n
ec
ted
la
y
er
(
d
en
s
e)
r
e
m
ain
s
th
e
s
am
e
as
in
th
e
o
r
ig
in
al
L
eNe
t
-
5
ar
ch
itectu
r
e
as
th
e
n
u
m
b
er
o
f
f
ilter
s
i
n
th
e
co
n
v
o
lu
tio
n
al
la
y
er
a
n
d
th
e
k
er
n
el
s
ize
n
ee
d
to
b
e
ad
j
u
s
ted
to
ca
p
t
u
r
e
r
ele
v
an
t
f
ea
tu
r
es
in
th
e
E
E
G
s
ig
n
al
as
s
h
o
wn
in
Fig
u
r
e
5
.
On
ce
t
h
e
ar
ch
itectu
r
e
an
d
h
y
p
er
p
a
r
a
m
eter
s
ar
e
o
p
tim
ize
d
(
T
a
b
le
1
)
,
th
e
m
o
d
el
ca
n
b
e
tr
ain
ed
u
s
in
g
s
tan
d
ar
d
o
p
ti
m
izatio
n
tech
n
iq
u
es,
s
u
ch
a
s
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t
o
r
Ad
am
.
Mo
d
el
p
er
f
o
r
m
an
ce
ca
n
t
h
en
b
e
e
v
al
u
ated
u
s
in
g
ap
p
r
o
p
r
iate
m
etr
ic
s
,
s
u
ch
as
ac
cu
r
ac
y
o
r
F1
s
co
r
e,
o
n
th
e
v
alid
atio
n
s
et
o
r
th
r
o
u
g
h
cr
o
s
s
-
v
alid
atio
n
.
Fig
u
r
e
4
.
E
x
am
p
le
d
iag
r
am
o
f
h
o
w
th
e
GR
U
wo
r
k
s
[
3
2
]
Fig
u
r
e
5
.
T
h
e
L
eNe
t
-
5
a
r
ch
ite
ctu
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
cc
u
r
a
cy
o
f n
e
u
r
a
l n
etw
o
r
ks in
b
r
a
in
w
a
ve
d
ia
g
n
o
s
is
o
f sch
i
z
o
p
h
r
en
ia
(
S
u
ke
mi)
1317
T
ab
le
1
.
T
h
e
h
y
p
er
p
a
r
am
eter
o
f
1
D
L
eNe
t
-
5
ar
ch
itectu
r
e
La
y
e
r
s
La
y
e
r
s
N
a
m
e
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
s
A
c
t
i
v
a
t
i
o
n
1
C
o
n
v
1
D
_
1
6
5
1
R
e
LU
2
M
a
x
P
o
o
l
i
n
g
1
D
_
1
-
2
2
-
3
C
o
n
v
1
D
_
2
16
5
1
R
e
LU
4
M
a
x
P
o
o
l
i
n
g
1
D
_
2
-
2
2
-
5
C
o
n
v
1
D
_
3
1
2
0
5
1
R
e
LU
6
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
1
D
-
-
-
-
7
D
e
n
se
_
1
84
-
-
R
e
LU
8
D
e
n
se
_
2
1
-
-
S
i
g
m
o
i
d
T
o
ad
a
p
t
Alex
Net
to
w
o
r
k
o
n
1
D
d
ata,
it
is
n
ec
ess
ar
y
to
m
o
d
if
y
t
h
e
in
p
u
t
lay
er
to
ac
ce
p
t
1
D
s
ig
n
als
r
ath
er
th
an
3
-
c
h
an
n
el
im
ag
es.
I
t
n
ee
d
s
to
ad
ju
s
t
th
e
n
u
m
b
er
o
f
f
ilter
s
an
d
k
er
n
el
s
ize
in
th
e
co
n
v
o
lu
tio
n
al
lay
er
to
b
etter
ca
p
tu
r
e
r
elev
a
n
t
f
ea
tu
r
es
in
th
e
E
E
G
s
ig
n
al.
On
e
ap
p
r
o
ac
h
t
o
o
p
tim
ize
Alex
Net
f
o
r
1
D
E
E
G
d
ata
is
to
u
s
e
C
o
n
v
1
D
in
s
tead
o
f
C
o
n
v
2
D.
I
t
also
r
ep
lace
s
t
h
e
Flatten
lay
er
with
g
lo
b
al
av
e
r
ag
e
p
o
o
lin
g
1
D
to
r
ed
u
ce
o
v
er
f
itti
n
g
.
T
h
e
n
u
m
b
er
o
f
f
ilter
s
in
t
h
e
co
n
v
o
lu
ti
o
n
al
lay
er
ca
n
b
e
i
n
cr
ea
s
ed
t
o
ca
p
tu
r
e
co
m
p
le
x
p
atter
n
s
in
th
e
E
E
G
s
ig
n
al.
T
h
e
last
lay
er
o
f
Alex
Net
u
s
es so
f
tm
ax
lay
er
f
o
r
class
if
icatio
n
(
T
ab
le
2
)
.
T
ab
le
2
.
T
h
e
h
y
p
er
p
a
r
am
eter
o
f
1
D
Alex
Net
ar
ch
itectu
r
e
La
y
e
r
s
La
y
e
r
s
N
a
m
e
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
s
A
c
t
i
v
a
t
i
o
n
1
C
o
n
v
1
D
_
1
96
11
4
R
e
LU
2
M
a
x
P
o
o
l
i
n
g
1
D
_
1
-
3
2
-
3
C
o
n
v
1
D
_
2
2
5
6
5
1
R
e
LU
4
M
a
x
P
o
o
l
i
n
g
1
D
_
2
-
3
2
-
5
C
o
n
v
1
D
_
3
3
8
4
3
1
R
e
LU
6
C
o
n
v
1
D
_
4
3
8
4
3
1
R
e
LU
7
C
o
n
v
1
D
_
5
2
5
6
3
1
R
e
LU
8
M
a
x
P
o
o
l
i
n
g
1
D
_
3
-
3
2
-
9
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
1
D
-
-
-
-
10
D
e
n
se
_
1
4
,
0
9
6
-
-
R
e
LU
11
D
r
o
p
o
u
t
_
1
-
-
-
-
12
D
e
n
se
_
2
4
,
0
9
6
-
-
R
e
LU
13
D
r
o
p
o
u
t
_
2
-
-
-
-
14
D
e
n
se
_
3
1
-
-
S
i
g
m
o
i
d
Fo
r
E
E
G
d
ata,
we
u
s
e
a
d
if
f
er
en
t
o
u
tp
u
t
lay
er
,
s
u
ch
as
a
s
ig
m
o
id
lay
er
f
o
r
b
in
a
r
y
cla
s
s
if
icatio
n
.
I
n
th
e
o
r
ig
in
al
VGG
-
1
6
ar
c
h
itectu
r
e,
th
e
in
p
u
t
s
h
ap
e
is
(
2
2
4
,
2
2
4
,
3
)
,
wh
ich
co
r
r
esp
o
n
d
s
to
a
2
D
im
ag
e
with
224
×
2
2
4
p
i
x
els
an
d
3
c
o
lo
r
ch
an
n
els.
T
o
m
a
k
e
it
wo
r
k
with
1
D
d
ata,
we
ca
n
c
h
an
g
e
th
e
in
p
u
t
s
h
a
p
e
to
(
n
,
1
,
1
)
,
wh
er
e
n
is
th
e
len
g
t
h
o
f
th
e
1
D
s
ig
n
al.
T
h
is
will
allo
w
th
e
a
r
ch
itectu
r
e
to
ac
c
e
p
t
1
D
in
p
u
t
s
ig
n
als.
T
h
e
VGG
-
1
6
ar
ch
itectu
r
e
h
as
s
ev
er
al
co
n
v
o
lu
tio
n
al
lay
er
s
with
lar
g
e
f
ilter
s
izes,
wh
ich
m
ay
n
o
t
b
e
s
u
itab
le
f
o
r
1
D
d
ata
(
T
ab
le
3
)
.
So
m
e
o
f
th
ese
lay
er
s
ar
e
r
em
o
v
e
d
an
d
r
ep
lace
d
with
s
m
aller
f
ilter
s
izes.
T
h
e
3
×
3
f
ilter
s
in
th
e
two
co
n
v
o
lu
tio
n
al
lay
er
s
ar
e
r
ep
lace
d
with
1
×
3
f
ilter
s
to
ca
p
tu
r
e
tem
p
o
r
al
f
ea
tu
r
es a
l
o
n
g
th
e
s
ig
n
al.
Mo
r
e
co
n
v
o
l
u
tio
n
al
la
y
er
s
ar
e
ad
d
e
d
to
ca
p
t
u
r
e
m
o
r
e
co
m
p
lex
f
ea
tu
r
es
alo
n
g
th
e
s
ig
n
al.
T
h
es
e
ad
d
itio
n
al
la
y
er
s
s
h
o
u
ld
h
av
e
s
m
aller
f
ilter
s
izes
to
av
o
id
o
v
e
r
f
itti
n
g
a
n
d
im
p
r
o
v
e
th
e
m
o
d
el'
s
ab
ilit
y
to
g
en
er
alize
.
T
h
e
VGG
-
1
6
ar
ch
itectu
r
e
u
s
es
m
ax
lay
er
p
o
o
lin
g
with
a
2
×
2
p
o
o
l
s
ize
to
s
h
r
in
k
th
e
f
ea
tu
r
e
m
ap
.
Fo
r
1
D
d
ata,
we
ca
n
u
s
e
a
v
er
ag
e
p
o
o
lin
g
in
s
tead
o
f
m
a
x
p
o
o
lin
g
to
ca
p
tu
r
e
th
e
av
e
r
ag
e
v
alu
e
o
f
th
e
f
ea
tu
r
e
m
ap
th
r
o
u
g
h
o
u
t
th
e
s
ig
n
al.
T
h
e
VGG
-
1
6
ar
ch
itectu
r
e
h
a
s
th
r
ee
f
u
lly
co
n
n
ec
ted
lay
er
s
at
th
e
en
d
o
f
th
e
ar
ch
itectu
r
e,
wh
ich
m
ay
n
o
t
b
e
s
u
itab
le
f
o
r
1
D
d
ata
an
d
ca
n
ad
d
m
o
r
e
f
u
lly
co
n
n
e
cted
lay
er
s
to
th
e
ar
ch
itectu
r
e
to
ca
p
t
u
r
e
m
o
r
e
c
o
m
p
lex
r
elatio
n
s
h
ip
s
b
etwe
en
f
ea
tu
r
es a
lo
n
g
th
e
s
ig
n
al.
T
h
e
n
u
m
b
er
o
f
f
ilter
s
in
ea
ch
co
n
v
o
lu
tio
n
lay
er
was r
e
d
u
ce
d
to
h
alf
to
p
r
ev
e
n
t o
v
er
f
i
ttin
g
an
d
to
allo
w
th
e
p
r
o
ce
s
s
to
r
u
n
o
n
Mic
r
o
s
o
f
t
Vis
u
al
Stu
d
io
C
o
d
e
(
3
2
-
b
it).
Sin
ce
th
e
ar
ch
itectu
r
e
is
m
o
d
i
f
ied
f
o
r
1
D
d
ata,
th
e
o
p
tim
al
l
ea
r
n
in
g
r
ate
m
ay
b
e
s
m
aller
th
an
th
e
o
r
ig
in
al
VGG
-
1
6
ar
ch
itectu
r
e.
W
e
ca
n
tr
y
d
if
f
er
en
t
lear
n
in
g
r
ates
an
d
ch
o
o
s
e
th
e
o
n
e
th
at
g
iv
es th
e
b
est p
er
f
o
r
m
an
ce
o
n
th
e
v
alid
atio
n
s
et.
I
n
T
ab
le
4
,
t
h
e
m
o
d
if
ied
R
esNet
-
1
8
ar
ch
itectu
r
e
f
o
r
1
D
d
a
ta,
s
u
ch
as
E
E
G
s
ig
n
als,
r
etain
s
th
e
b
asic
p
r
in
cip
les
o
f
th
e
o
r
ig
in
al
R
esNet
-
1
8
wh
ile
ad
ap
tin
g
it
to
h
an
d
le
1
D
in
p
u
ts
.
T
h
e
ar
ch
itectu
r
e
co
n
s
is
ts
o
f
s
ev
er
al
k
ey
c
o
m
p
o
n
en
ts
.
First,
th
e
in
p
u
t
lay
er
tak
es
th
e
f
o
r
m
o
f
a
1
D
E
E
G
s
ig
n
al.
T
h
e
s
u
b
s
eq
u
en
t
s
tag
es
f
o
llo
w
th
e
R
esNet
-
1
8
s
tr
u
ctu
r
e,
with
ea
ch
s
tag
e
co
n
s
is
tin
g
o
f
a
s
er
ies
o
f
r
esid
u
e
b
lo
ck
s
.
Ho
wev
er
,
i
n
th
is
m
o
d
if
ied
v
er
s
io
n
,
th
e
o
r
ig
i
n
al
2
D
co
n
v
o
l
u
tio
n
al
lay
er
is
r
ep
lace
d
with
a
1
D
c
o
n
v
o
lu
ti
o
n
al
lay
er
.
T
h
is
ad
ju
s
tm
en
t
allo
ws
th
e
m
o
d
el
to
ca
p
tu
r
e
tem
p
o
r
al
d
ep
en
d
en
cies
in
E
E
G
d
ata
al
o
n
g
o
n
e
d
im
en
s
io
n
.
R
esid
u
e
b
lo
ck
s
,
wh
ich
ar
e
th
e
b
u
ild
in
g
b
lo
ck
s
o
f
R
esNet,
h
elp
to
o
v
e
r
co
m
e
th
e
v
an
is
h
in
g
g
r
ad
ie
n
t
p
r
o
b
lem
an
d
en
a
b
le
d
ee
p
n
etwo
r
k
t
r
ain
in
g
.
E
ac
h
b
lo
c
k
co
n
s
is
ts
o
f
c
o
n
v
o
lu
ti
o
n
al
lay
er
s
with
s
k
ip
co
n
n
ec
tio
n
s
,
wh
ic
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
2
,
Ap
r
il 2
0
2
5
:
1
3
1
1
-
1
3
2
5
1318
allo
w
g
r
ad
ie
n
ts
to
f
lo
w
d
ir
ec
tly
th
r
o
u
g
h
t
h
e
n
etwo
r
k
.
T
h
e
n
u
m
b
er
o
f
f
ilter
s
in
th
e
co
n
v
o
l
u
tio
n
al
lay
e
r
v
ar
ies
ac
r
o
s
s
s
tag
es
to
ca
p
tu
r
e
d
if
f
er
en
t
lev
els
o
f
f
ea
tu
r
es.
T
o
ad
ap
t
th
e
o
u
tp
u
t
o
f
th
e
R
esNet
-
1
8
m
o
d
el
to
th
e
class
if
icatio
n
ta
s
k
,
a
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
lay
er
is
ap
p
lied
to
r
ed
u
ce
th
e
s
p
atial
d
im
en
s
io
n
o
f
th
e
f
ea
tu
r
e
m
ap
to
o
n
e
v
alu
e
p
e
r
ch
an
n
el.
T
h
is
p
o
o
lin
g
o
p
er
atio
n
a
v
er
ag
es
th
e
v
alu
es
o
f
ea
ch
c
h
an
n
el,
ef
f
ec
tiv
ely
s
u
m
m
ar
izin
g
th
e
lear
n
ed
f
ea
t
u
r
es.
Dr
o
p
o
u
t
r
e
g
u
lar
izatio
n
i
s
th
en
ap
p
lied
to
p
r
ev
en
t
o
v
er
f
itti
n
g
b
y
r
an
d
o
m
ly
d
is
ab
lin
g
a
s
m
all
p
o
r
tio
n
o
f
n
eu
r
o
n
s
.
Fin
ally
,
th
e
o
u
t
p
u
t
l
ay
er
co
n
s
is
ts
o
f
a
d
en
s
e
la
y
e
r
with
th
e
d
esire
d
n
u
m
b
er
o
f
class
es,
wh
ich
y
ield
s
th
e
p
r
ed
icted
class
p
r
o
b
ab
ilit
ies.
Ov
er
all,
th
is
m
o
d
if
ied
R
esNet
-
1
8
ar
ch
itectu
r
e
f
o
r
1
D
d
ata
p
r
o
v
id
es
a
r
o
b
u
s
t
f
r
am
ewo
r
k
f
o
r
ca
p
tu
r
in
g
tem
p
o
r
al
p
atter
n
s
an
d
f
ea
tu
r
es
in
E
E
G
s
ig
n
als,
m
ak
in
g
it
s
u
itab
le
f
o
r
task
s
s
u
ch
as
E
E
G
-
b
ased
class
if
icat
io
n
an
d
an
aly
s
is
in
th
e
co
n
tex
t
o
f
n
eu
r
o
lo
g
ical
d
is
o
r
d
e
r
s
an
d
b
r
a
in
-
co
m
p
u
ter
in
ter
f
ac
es.
T
ab
le
3
.
T
h
e
h
y
p
er
p
a
r
am
eter
o
f
1
D
VGG
-
1
6
ar
ch
itectu
r
e
La
y
e
r
s
La
y
e
r
s
N
a
m
e
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
s
A
c
t
i
v
a
t
i
o
n
1
C
o
n
v
1
D
_
1
32
3
1
R
e
LU
2
C
o
n
v
1
D
_
2
32
3
1
R
e
LU
3
M
a
x
P
o
o
l
i
n
g
1
D
_
1
-
2
2
-
4
C
o
n
v
1
D
_
3
64
3
1
R
e
LU
5
C
o
n
v
1
D
_
4
64
3
1
R
e
LU
6
M
a
x
P
o
o
l
i
n
g
1
D
_
2
-
2
2
-
7
C
o
n
v
1
D
_
5
1
2
8
3
1
R
e
LU
8
C
o
n
v
1
D
_
6
1
2
8
3
1
R
e
LU
9
C
o
n
v
1
D
_
7
1
2
8
3
1
R
e
LU
10
M
a
x
P
o
o
l
i
n
g
1
D
_
3
-
2
2
-
11
C
o
n
v
1
D
_
8
2
5
6
3
1
R
e
LU
12
C
o
n
v
1
D
_
9
2
5
6
3
1
R
e
LU
13
C
o
n
v
1
D
_
1
0
2
5
6
3
1
R
e
LU
14
M
a
x
P
o
o
l
i
n
g
1
D
_
4
-
2
2
-
15
C
o
n
v
1
D
_
1
1
2
5
6
3
1
R
e
LU
16
C
o
n
v
1
D
_
1
2
2
5
6
3
1
R
e
LU
17
C
o
n
v
1
D
_
1
3
2
5
6
3
1
R
e
LU
18
M
a
x
P
o
o
l
i
n
g
1
D
_
5
-
2
2
-
19
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
1
D
-
-
-
-
20
D
e
n
se
_
1
2
,
0
4
8
-
-
R
e
LU
21
D
r
o
p
o
u
t
_
1
0
.
5
-
-
-
22
D
e
n
se
_
2
2
,
0
4
8
-
-
R
e
LU
23
D
r
o
p
o
u
t
_
2
0
.
5
-
-
-
24
D
e
n
se
_
3
1
-
-
S
i
g
m
o
i
d
T
ab
le
4
.
T
h
e
h
y
p
er
p
a
r
am
eter
o
f
1
D
R
esNet
-
1
8
ar
ch
itectu
r
e
La
y
e
r
s
La
y
e
r
s
N
a
m
e
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
s
A
c
t
i
v
a
t
i
o
n
1
C
o
n
v
1
D
_
1
64
7
2
R
e
LU
2
M
a
x
P
o
o
l
i
n
g
1
D
_
1
-
3
2
-
3
R
e
si
d
u
a
l
B
l
o
c
k
_
1
-
-
-
-
4
R
e
si
d
u
a
l
B
l
o
c
k
_
2
-
-
-
-
5
R
e
si
d
u
a
l
B
l
o
c
k
_
3
-
-
-
-
6
R
e
si
d
u
a
l
B
l
o
c
k
_
4
-
-
-
-
7
C
o
n
v
1
D
_
2
1
2
8
3
2
R
e
LU
8
R
e
si
d
u
a
l
B
l
o
c
k
_
5
-
-
-
-
9
R
e
si
d
u
a
l
B
l
o
c
k
_
6
-
-
-
-
10
R
e
si
d
u
a
l
B
l
o
c
k
_
7
-
-
-
-
11
R
e
si
d
u
a
l
B
l
o
c
k
_
8
-
-
-
-
12
C
o
n
v
1
D
_
3
2
5
6
3
2
R
e
LU
13
R
e
si
d
u
a
l
B
l
o
c
k
_
9
-
-
-
-
14
R
e
si
d
u
a
l
B
l
o
c
k
_
1
0
-
-
-
-
15
R
e
si
d
u
a
l
B
l
o
c
k
_
1
1
-
-
-
-
16
R
e
si
d
u
a
l
B
l
o
c
k
_
1
2
-
-
-
-
17
C
o
n
v
1
D
_
4
5
1
2
3
2
R
e
LU
18
R
e
si
d
u
a
l
B
l
o
c
k
_
1
3
-
-
-
-
19
R
e
si
d
u
a
l
B
l
o
c
k
_
1
4
-
-
-
-
20
R
e
si
d
u
a
l
B
l
o
c
k
_
1
5
-
-
-
-
21
R
e
si
d
u
a
l
B
l
o
c
k
_
1
6
-
-
-
-
22
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
1
D
-
-
-
-
23
D
e
n
se
_
1
1
-
-
S
i
g
m
o
i
d
I
n
th
e
p
r
o
p
o
s
ed
C
NN
-
L
STM
ar
ch
itectu
r
e,
it
co
n
s
is
ts
o
f
a
to
tal
o
f
1
3
lay
er
s
,
co
m
b
in
in
g
C
NN
an
d
L
STM
lay
er
s
.
Similar
to
th
e
m
o
d
el
in
[
1
8
]
,
t
h
is
ar
ch
itectu
r
e
u
s
es
th
e
g
lo
b
al
av
er
a
g
e
p
o
o
l
in
g
1
D
lay
er
i
n
s
tead
o
f
f
latten
to
r
ed
u
ce
th
e
s
p
atia
l
d
im
en
s
io
n
o
f
th
e
f
ea
tu
r
e
m
a
p
.
T
h
e
f
ir
s
t
1
0
lay
er
s
o
f
th
is
C
NN
-
L
STM
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
cc
u
r
a
cy
o
f n
e
u
r
a
l n
etw
o
r
ks in
b
r
a
in
w
a
ve
d
ia
g
n
o
s
is
o
f sch
i
z
o
p
h
r
en
ia
(
S
u
ke
mi)
1319
r
em
ain
th
e
s
am
e
as th
e
p
r
ev
io
u
s
ar
ch
itectu
r
e,
with
th
e
C
NN
lay
er
r
esp
o
n
s
ib
le
f
o
r
ex
tr
ac
tin
g
lo
ca
l f
ea
tu
r
es a
n
d
p
atter
n
s
f
r
o
m
th
e
in
p
u
t
d
ata.
T
h
e
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
1
D
lay
er
f
o
llo
ws,
wh
ich
p
er
f
o
r
m
s
a
p
o
o
lin
g
o
p
er
atio
n
to
s
u
m
m
ar
ize
th
e
ex
tr
ac
ted
f
ea
tu
r
es
i
n
ea
ch
ch
an
n
el.
B
y
u
s
in
g
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
1
D,
aim
f
o
r
a
f
ix
ed
-
len
g
th
r
ep
r
esen
tatio
n
r
e
g
ar
d
less
o
f
th
e
len
g
th
o
f
th
e
in
p
u
t seq
u
en
ce
.
I
n
th
e
n
e
x
t
lay
e
r
,
a
d
e
n
s
e
lay
e
r
with
5
0
n
eu
r
o
n
s
is
u
s
ed
a
n
d
a
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
ac
tiv
atio
n
f
u
n
ctio
n
is
u
s
ed
t
o
in
tr
o
d
u
ce
n
o
n
-
lin
ea
r
ity
to
th
e
m
o
d
el.
T
h
is
lay
er
m
ak
es
it
p
o
s
s
ib
le
to
l
ea
r
n
m
o
r
e
c
o
m
p
lex
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
.
T
o
p
r
ev
en
t
o
v
er
f
itti
n
g
,
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
2
5
was
ad
d
ed
t
o
th
e
1
2
th
lay
er
.
Fin
ally
,
at
th
e
1
3
th
lay
er
,
a
d
en
s
e
lay
er
with
a
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
is
u
s
ed
to
p
er
f
o
r
m
th
e
class
if
icatio
n
ta
s
k
.
B
y
co
m
b
in
in
g
th
e
C
NN
an
d
L
STM
co
m
p
o
n
en
ts
,
alo
n
g
with
th
e
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
1
D
lay
er
,
th
e
p
r
o
p
o
s
ed
C
NN
-
L
STM
m
o
d
el
is
ab
le
to
ca
p
tu
r
e
b
o
th
s
p
ati
al
an
d
tem
p
o
r
al
in
f
o
r
m
atio
n
f
r
o
m
E
E
G
s
ig
n
als.
T
h
is
ar
ch
itectu
r
e
allo
ws
th
e
n
etwo
r
k
to
lear
n
m
ea
n
in
g
f
u
l
r
e
p
r
esen
tatio
n
s
an
d
m
ak
e
ac
cu
r
ate
p
r
ed
ictio
n
s
f
o
r
E
E
G
s
ig
n
al
class
if
icatio
n
in
th
is
s
tu
d
y
(
T
ab
le
5
)
.
T
ab
le
5
.
T
h
e
h
y
p
er
p
a
r
am
eter
o
f
1
D
C
NN
-
L
STM
ar
ch
itectu
r
e
La
y
e
r
s
La
y
e
r
s
N
a
m
e
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
s
A
c
t
i
v
a
t
i
o
n
1
C
o
n
v
1
D
64
3
1
R
e
LU
2
C
o
n
v
1
D
64
3
1
R
e
LU
3
D
r
o
p
o
u
t
0
.
5
-
-
-
4
M
a
x
P
o
o
l
i
n
g
1
D
-
2
1
-
6
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
1
D
-
-
-
-
7
LSTM
1
0
0
-
-
-
8
D
r
o
p
o
u
t
0
.
5
-
-
-
9
D
e
n
se
1
0
0
-
-
R
e
LU
10
D
r
o
p
o
u
t
0
.
2
5
-
-
-
11
D
e
n
se
50
-
-
R
e
LU
12
D
r
o
p
o
u
t
0
.
2
5
-
-
13
D
e
n
se
1
-
-
S
i
g
m
o
i
d
T
h
e
p
r
o
p
o
s
ed
C
NN
-
GR
U
m
o
d
el
is
u
s
ed
f
o
r
co
m
p
ar
is
o
n
in
th
is
s
tu
d
y
.
T
h
is
ar
ch
itectu
r
e
c
o
m
b
in
es
a
C
NN
lay
er
with
a
GR
U
lay
e
r
,
p
r
o
v
id
in
g
a
d
if
f
e
r
en
t
ap
p
r
o
ac
h
to
ca
p
t
u
r
e
th
e
tem
p
o
r
al
d
y
n
am
ics
o
f
E
E
G
s
ig
n
als.
A
f
ix
ed
g
lo
b
al
a
v
er
a
g
e
p
o
o
lin
g
1
D
lay
er
is
u
s
ed
to
r
ep
lace
f
latten
in
s
u
m
m
ar
i
zin
g
th
e
e
x
tr
ac
ted
f
ea
tu
r
es
in
ea
ch
ch
an
n
el.
Sim
ilar
to
th
e
p
r
ev
io
u
s
m
o
d
el,
th
e
f
ir
s
t
1
0
lay
e
r
s
o
f
t
h
is
C
NN
-
GR
U
ar
ch
itectu
r
e
r
em
ain
u
n
ch
a
n
g
ed
,
co
n
s
is
tin
g
o
f
a
C
NN
lay
er
th
at
e
x
tr
ac
ts
lo
ca
l
f
ea
tu
r
es
f
r
o
m
t
h
e
in
p
u
t
E
E
G
d
ata.
T
h
e
s
u
b
s
eq
u
en
t
g
lo
b
al
av
e
r
a
g
e
p
o
o
lin
g
1
D
lay
er
is
u
s
ed
to
r
ed
u
ce
th
e
s
p
atial
d
im
en
s
io
n
an
d
o
b
tain
a
f
ix
ed
-
len
g
th
r
ep
r
esen
tatio
n
,
r
eg
ar
d
less
o
f
th
e
len
g
th
o
f
th
e
in
p
u
t
s
eq
u
en
ce
Fo
llo
win
g
th
e
p
o
o
lin
g
lay
e
r
,
th
e
GR
U
lay
er
is
u
s
ed
in
s
tead
o
f
th
e
L
STM
.
GR
U
i
s
a
ty
p
e
o
f
R
NN
th
at
o
f
f
er
s
a
m
o
r
e
ef
f
icien
t
ar
ch
itectu
r
e
b
y
u
tili
zin
g
a
g
atin
g
m
ec
h
a
n
is
m
to
s
elec
tiv
ely
s
to
r
e
im
p
o
r
tan
t
in
f
o
r
m
atio
n
an
d
d
is
ca
r
d
ir
r
elev
an
t
in
f
o
r
m
atio
n
f
r
o
m
th
e
s
eq
u
en
ce
.
T
h
is
h
elp
s
th
e
m
o
d
el
ef
f
ec
tiv
ely
ca
p
tu
r
e
th
e
tem
p
o
r
al
d
ep
e
n
d
en
ci
es
p
r
esen
t
in
E
E
G
s
ig
n
als.
T
o
f
u
r
th
er
en
h
a
n
ce
t
h
e
r
ep
r
esen
tatio
n
ca
p
ac
ity
o
f
th
e
m
o
d
el,
a
d
en
s
e
lay
e
r
with
5
0
n
e
u
r
o
n
s
a
n
d
a
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
wa
s
also
ap
p
lied
to
t
h
is
m
o
d
el.
T
h
is
lay
er
allo
ws
t
h
e
n
etwo
r
k
to
lear
n
co
m
p
lex
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
.
T
o
r
e
d
u
ce
o
v
e
r
f
itti
n
g
,
a
d
r
o
p
o
u
t
lay
er
with
a
r
ate
o
f
0
.
2
5
is
in
clu
d
ed
af
ter
th
e
d
en
s
e
lay
er
.
I
n
th
e
last
lay
er
,
a
d
e
n
s
e
lay
er
with
a
s
ig
m
o
i
d
ac
tiv
atio
n
f
u
n
ctio
n
is
u
s
ed
f
o
r
class
if
icatio
n
,
allo
win
g
th
e
m
o
d
el
to
o
u
tp
u
t a
p
r
o
b
ab
ilit
y
v
alu
e
in
d
icatin
g
th
e
lik
elih
o
o
d
o
f
th
e
in
p
u
t E
E
G
s
ig
n
al
b
elo
n
g
in
g
to
a
p
ar
ticu
lar
class
.
B
y
co
m
b
in
in
g
th
e
C
NN
lay
er
,
GR
U
lay
er
,
an
d
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
1
D
l
ay
er
,
th
e
p
r
o
p
o
s
ed
C
NN
-
GR
U
m
o
d
el
ca
p
tu
r
es
b
o
th
s
p
atial
a
n
d
tem
p
o
r
al
in
f
o
r
m
atio
n
f
r
o
m
th
e
E
E
G
s
ig
n
al
.
T
h
is
a
r
ch
itectu
r
e
en
ab
les
th
e
n
etwo
r
k
t
o
ef
f
ec
tiv
ely
lear
n
r
elev
a
n
t
p
atter
n
s
a
n
d
m
ak
e
ac
cu
r
ate
p
r
e
d
ictio
n
s
f
o
r
th
e
class
if
icatio
n
task
in
th
is
s
tu
d
y
.
3.
RE
SU
L
T
S
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
i
f
ied
m
o
d
els,
in
clu
d
in
g
Alex
Net,
VGG
-
1
6
,
L
eNe
t
-
5
,
an
d
R
esNet
-
1
8
,
was
ev
alu
ated
u
s
in
g
two
m
ain
m
etr
ics:
te
s
t
ac
cu
r
ac
y
an
d
F1
s
co
r
e
.
T
h
e
r
esu
lts
d
em
o
n
s
tr
ate
th
e
ef
f
ec
tiv
en
ess
o
f
th
ese
m
o
d
els
in
class
if
y
in
g
E
E
G
s
ig
n
als
f
o
r
s
ch
izo
p
h
r
e
n
ia
d
iag
n
o
s
is
.
B
o
th
Alex
Net
an
d
VGG
-
1
6
ac
h
ie
v
ed
o
u
ts
tan
d
in
g
p
e
r
f
o
r
m
an
ce
,
wit
h
test
ac
cu
r
ac
y
an
d
F1
s
co
r
e
s
o
f
0
.
9
9
,
i
n
d
icatin
g
h
i
g
h
ly
a
cc
u
r
ate
p
r
ed
ictio
n
s
.
T
h
ese
m
o
d
els
d
em
o
n
s
tr
ated
o
u
ts
tan
d
in
g
d
is
cr
im
in
ativ
e
ab
ilit
y
,
ac
cu
r
ately
d
is
tin
g
u
is
h
in
g
b
etwe
en
h
ea
lth
y
in
d
iv
id
u
als
an
d
th
o
s
e
with
p
ar
an
o
id
s
ch
izo
p
h
r
e
n
ia.
I
n
a
d
d
itio
n
,
L
eNe
t
-
5
an
d
R
esNe
t
-
1
8
ac
h
iev
ed
test
ac
cu
r
ac
y
a
n
d
F1
s
co
r
es
o
f
0
.
9
8
,
d
em
o
n
s
tr
atin
g
s
tr
o
n
g
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
i
f
ied
m
o
d
el
was
ev
alu
ated
u
s
in
g
a
co
m
p
r
e
h
en
s
iv
e
E
E
G
s
ig
n
al
d
ataset
co
llected
f
r
o
m
p
atien
ts
d
iag
n
o
s
ed
with
s
ch
izo
p
h
r
en
ia
an
d
a
g
r
o
u
p
o
f
h
ea
lth
y
in
d
i
v
i
d
u
als.
Fo
u
r
wid
ely
r
ec
o
g
n
ized
C
NN
ar
ch
itectu
r
e
s
,
n
am
ely
L
eNe
t
-
5
,
Alex
Net,
VGG
-
1
6
,
an
d
R
esNet
-
1
8
,
wer
e
ad
ap
ted
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
2
,
Ap
r
il 2
0
2
5
:
1
3
1
1
-
1
3
2
5
1320
m
o
d
if
ied
to
e
f
f
ec
tiv
ely
h
a
n
d
l
e
th
e
u
n
iq
u
e
c
h
ar
ac
ter
is
tics
o
f
1
D
E
E
G
d
ata.
T
h
e
m
o
d
el
u
n
d
er
g
o
es
r
ig
o
r
o
u
s
tr
ain
in
g
an
d
test
in
g
p
r
o
ce
d
u
r
e
s
u
s
in
g
ap
p
r
o
p
r
iate
d
ata
s
p
lits
to
en
s
u
r
e
r
eliab
le
ev
alu
atio
n
.
E
v
alu
atio
n
m
etr
ics
u
s
ed
to
ass
ess
m
o
d
el
p
er
f
o
r
m
an
ce
in
clu
d
e
test
ac
cu
r
ac
y
a
n
d
F1
s
co
r
es,
wh
ich
ar
e
c
o
m
m
o
n
m
ea
s
u
r
es
u
s
ed
in
class
if
icatio
n
task
s
.
T
h
e
Fig
u
r
es
6
to
1
0
a
r
e
th
e
o
u
tp
u
t
r
esu
lts
f
r
o
m
ea
ch
m
o
d
el
u
s
ed
.
Am
o
n
g
th
e
m
o
d
if
ied
m
o
d
els,
th
e
m
o
d
if
ied
Alex
Net
an
d
VGG
-
1
6
m
o
d
els
s
to
o
d
o
u
t
with
o
u
ts
tan
d
in
g
p
er
f
o
r
m
an
ce
,
ac
h
iev
in
g
im
p
r
ess
iv
e
test
in
g
ac
cu
r
ac
y
a
n
d
an
F1
s
co
r
e
o
f
0
.
9
9
.
T
h
e
s
e
m
o
d
els
s
h
o
wed
r
em
ar
k
ab
le
ab
ilit
y
in
ac
cu
r
at
ely
class
if
y
in
g
E
E
G
s
ig
n
als
f
r
o
m
in
d
iv
id
u
als
with
s
ch
izo
p
h
r
en
ia
an
d
h
ea
lth
y
co
n
tr
o
ls
.
T
h
e
r
o
b
u
s
tn
ess
an
d
ef
f
ec
tiv
en
ess
o
f
t
h
is
m
o
d
el
ca
n
b
e
attr
ib
u
ted
to
th
e
in
co
r
p
o
r
atio
n
o
f
s
p
ec
ialized
co
n
v
o
l
u
tio
n
al
lay
er
s
,
p
o
o
lin
g
o
p
er
atio
n
s
,
a
n
d
n
o
n
-
li
n
ea
r
ac
t
iv
atio
n
,
wh
ich
en
ab
le
e
f
f
ec
tiv
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
an
d
d
is
cr
im
in
atio
n
o
f
r
elev
an
t
p
atter
n
s
in
E
E
G
d
ata.
I
n
ad
d
itio
n
,
th
e
m
o
d
i
f
ied
L
eNe
t
-
5
an
d
R
esNet
-
1
8
m
o
d
els
also
s
h
o
wed
s
tr
o
n
g
p
er
f
o
r
m
an
ce
,
ac
h
iev
in
g
co
m
m
en
d
a
b
le
test
a
cc
u
r
ac
y
an
d
a
n
F1
s
co
r
e
o
f
0
.
9
8
.
T
h
ese
m
o
d
els,
alth
o
u
g
h
s
lig
h
tly
b
eh
in
d
th
eir
Alex
Net
an
d
VGG
-
1
6
co
u
n
t
er
p
ar
ts
,
s
till
d
em
o
n
s
tr
ated
th
eir
s
u
itab
ilit
y
f
o
r
E
E
G
class
i
f
icatio
n
task
s
.
T
h
e
m
o
d
if
ied
L
eNe
t
-
5
m
o
d
el,
with
its
r
elativ
ely
s
im
p
ler
ar
c
h
itectu
r
e,
ex
h
ib
its
th
e
e
f
f
ica
cy
o
f
u
tili
zin
g
1
D
co
n
v
o
l
u
tio
n
al
lay
er
s
in
ca
p
tu
r
i
n
g
r
elev
an
t
tem
p
o
r
al
d
ep
en
d
e
n
cies
an
d
ex
tr
ac
tin
g
d
is
cr
im
in
ativ
e
f
ea
tu
r
es
f
r
o
m
E
E
G
s
ig
n
als.
Similar
ly
,
th
e
m
o
d
if
ied
R
esNet
-
1
8
m
o
d
el,
wit
h
its
d
ee
p
r
esid
u
al
ar
ch
itectu
r
e,
d
em
o
n
s
tr
ated
its
ab
ilit
y
to
ef
f
ec
tiv
ely
lea
r
n
co
m
p
lex
r
ep
r
esen
tatio
n
s
an
d
h
an
d
le
p
o
ten
tial c
h
allen
g
es in
m
o
d
elin
g
E
E
G
d
ata.
Ov
er
all,
th
e
r
esu
lts
o
b
tain
ed
f
r
o
m
th
e
ev
alu
atio
n
o
f
m
o
d
if
ied
C
NN
ar
ch
itectu
r
es
u
n
d
er
lin
e
th
eir
s
ig
n
if
ican
t
p
o
ten
tial
f
o
r
t
h
e
cl
ass
if
icatio
n
an
d
d
iag
n
o
s
is
o
f
s
ch
izo
p
h
r
e
n
ia
b
ased
o
n
E
E
G
d
ata.
T
h
ese
f
in
d
i
n
g
s
p
r
o
v
id
e
v
al
u
ab
le
in
s
ig
h
ts
f
o
r
th
e
f
ield
o
f
p
s
y
ch
iatr
y
r
esear
ch
,
h
ig
h
lig
h
tin
g
th
e
ef
f
icac
y
o
f
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
in
an
aly
zin
g
an
d
i
n
ter
p
r
etin
g
E
E
G
s
ig
n
als.
T
h
e
h
ig
h
test
ac
cu
r
ac
y
an
d
F1
s
co
r
es
ac
h
iev
ed
b
y
th
e
m
o
d
if
ied
m
o
d
el
r
ef
lect
its
r
o
b
u
s
tn
ess
an
d
r
eliab
ilit
y
,
d
em
o
n
s
tr
atin
g
its
p
o
ten
tial
f
o
r
p
r
ac
tical
ap
p
licatio
n
in
clin
ical
s
ettin
g
s
f
o
r
ea
r
ly
d
etec
tio
n
an
d
m
o
n
ito
r
i
n
g
o
f
s
ch
izo
p
h
r
en
ia.
A
c
c
u
r
a
c
y
Lo
ss
C
o
n
f
u
s
i
o
n
M
a
t
r
i
x
R
O
C
C
u
r
v
e
Fig
u
r
e
6
.
Acc
u
r
ac
y
c
u
r
v
e
a
n
d
co
n
f
u
s
io
n
m
atr
ix
d
ia
g
r
am
o
f
t
h
e
1
D
L
eNe
t
-
5
m
o
d
e
l
Evaluation Warning : The document was created with Spire.PDF for Python.