I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t
201
9
,
p
p
.
2
9
3
2
~2
9
4
0
I
SS
N:
2
0
8
8
-
8
708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v9
i
4
.
p
p
2
9
3
2
-
2940
2932
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
co
r
e
.
co
m/
jo
u
r
n
a
ls
/in
d
ex
.
p
h
p
/
I
JE
C
E
Na
tural
l
a
ng
ua
g
e
d
escriptio
n o
f
i
ma
g
es
u
sing
h
y
brid
r
ecu
rren
t
n
eura
l
n
etw
o
rk
M
d.
Asi
f
uzza
m
a
n J
is
ha
n,
K
ha
n Ra
qib
M
a
h
m
u
d,
Abul
K
a
la
m
Al
Aza
d
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
,
Un
iv
e
rsit
y
o
f
Li
b
e
ra
l
A
rts
Ba
n
g
lad
e
sh
,
Dh
a
k
a
-
1
2
0
9
,
Ba
n
g
lad
e
sh
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
26
,
2
0
1
8
R
ev
i
s
ed
Mar
2
1
,
2
0
1
9
A
cc
ep
ted
A
p
r
4
,
2
0
1
9
W
e
p
re
se
n
ted
a
lea
rn
in
g
m
o
d
e
l
th
a
t
g
e
n
e
ra
ted
n
a
tu
ra
l
lan
g
u
a
g
e
d
e
sc
rip
ti
o
n
o
f
im
a
g
e
s.
T
h
e
m
o
d
e
l
u
ti
li
z
e
d
t
h
e
c
o
n
n
e
c
ti
o
n
s
b
e
tw
e
e
n
n
a
tu
ra
l
lan
g
u
a
g
e
a
n
d
vi
su
a
l
d
a
ta
b
y
p
ro
d
u
c
e
d
tex
t
l
in
e
b
a
s
e
d
c
o
n
ten
ts
f
ro
m
a
g
i
v
e
n
im
a
g
e
.
Ou
r
Hy
b
rid
Re
c
u
rre
n
t
Ne
u
ra
l
N
e
tw
o
rk
m
o
d
e
l
is
b
a
se
d
o
n
th
e
i
n
t
rica
c
ies
o
f
Co
n
v
o
l
u
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
(CNN
),
L
o
n
g
S
h
o
rt
-
T
e
rm
M
e
m
o
r
y
(
L
S
T
M
),
a
n
d
Bi
-
d
irec
ti
o
n
a
l
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
tw
o
rk
(BRNN
)
m
o
d
e
ls.
We
c
o
n
d
u
c
ted
e
x
p
e
rim
e
n
ts
o
n
th
re
e
b
e
n
c
h
m
a
rk
d
a
tas
e
ts,
e
.
g
.
,
F
li
c
k
r8
K,
F
li
c
k
r3
0
K,
a
n
d
M
S
COCO
.
Ou
r
h
y
b
rid
m
o
d
e
l
u
ti
li
z
e
d
L
S
T
M
m
o
d
e
l
to
e
n
c
o
d
e
tex
t
li
n
e
o
r
se
n
ten
c
e
s
in
d
e
p
e
n
d
e
n
t
o
f
th
e
o
b
jec
t
l
o
c
a
ti
o
n
a
n
d
BRNN
f
o
r
w
o
rd
re
p
re
se
n
tatio
n
,
t
h
is
re
d
u
c
e
d
th
e
c
o
m
p
u
tatio
n
a
l
c
o
m
p
lex
it
i
e
s
w
it
h
o
u
t
c
o
m
p
ro
m
isin
g
th
e
a
c
c
u
ra
c
y
o
f
t
h
e
d
e
sc
rip
to
r
.
T
h
e
m
o
d
e
l
p
ro
d
u
c
e
d
b
e
tt
e
r
a
c
c
u
ra
c
y
in
re
tri
e
v
in
g
n
a
tu
ra
l
lan
g
u
a
g
e
b
a
se
d
d
e
sc
rip
ti
o
n
o
n
th
e
d
a
tas
e
t.
K
ey
w
o
r
d
s
:
Bi
-
d
ir
ec
tio
n
al
r
ec
u
r
r
en
t
n
e
u
r
a
l
n
et
w
o
r
k
L
o
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
Natu
r
al
la
n
g
u
a
g
e
d
escr
ip
to
r
s
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
H
y
b
r
id
r
ec
u
r
r
en
t n
e
u
r
al
n
et
w
o
r
k
Co
p
y
rig
h
t
©
2
0
1
9
In
stit
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Md
.
A
s
i
f
u
zz
a
m
a
n
J
is
h
an
,
Dep
ar
t
m
en
t o
f
C
o
m
p
u
ter
Sc
ie
n
ce
an
d
E
n
g
i
n
ee
r
in
g
,
Un
i
v
er
s
it
y
o
f
L
ib
er
al
A
r
ts
B
an
g
lad
es
h
,
Dh
a
n
m
o
n
d
i,
Dh
a
k
a
-
1
2
0
9
,
B
an
g
lad
es
h
.
E
m
ail: j
is
h
a
n
9
0
0
@
g
m
ai
l.c
o
m
1.
I
NT
RO
D
UCT
I
O
N
A
p
r
i
m
ar
y
m
o
t
iv
at
io
n
of
co
m
p
u
tatio
n
a
l
v
is
u
al
tas
k
s
is
to
e
m
u
late
th
e
r
e
m
ar
k
ab
le
h
u
m
an
ca
p
ab
ilit
y
to
co
m
p
r
e
h
en
d
v
i
s
u
al
s
ce
n
e
s
a
n
d
in
ter
p
r
et
d
etailed
in
f
o
r
m
at
io
n
f
r
o
m
t
h
e
m
w
i
th
asto
n
is
h
i
n
g
a
cc
u
r
ac
y
[
1
]
.
Fo
r
an
ar
tif
icial
s
y
s
t
em
to
e
m
u
late
th
i
s
ca
p
ab
ilit
y
o
f
i
m
ag
e
d
escr
ip
ti
o
n
is
n
o
t
m
er
el
y
r
e
s
tr
icted
to
r
ec
o
g
n
izi
n
g
i
m
a
g
es,
r
ath
er
it
is
i
m
p
o
r
tan
t
to
u
n
d
er
s
tan
d
b
o
th
s
y
n
tact
ic
an
d
s
e
m
a
n
tic
m
ea
n
in
g
o
f
th
e
i
m
a
g
es,
th
a
t
is
to
s
a
y
,
t
h
e
tas
k
m
u
s
t
i
n
v
o
l
v
e
u
n
d
er
s
ta
n
d
i
n
g
t
h
e
co
n
te
n
ts
of
th
e
i
m
a
g
e
a
n
d
also
th
e
i
n
ter
ac
tio
n
s
a
m
o
n
g
th
e
co
n
ten
t
s
[2
-
6
]
.
I
m
ag
e
d
escr
ip
tio
n
t
y
p
icall
y
i
s
th
e
g
e
n
er
atio
n
of
n
at
u
r
al
la
n
g
u
ag
e
b
ased
tex
tu
al
d
escr
ip
tio
n
of
an
i
m
a
g
e
w
h
ic
h
h
as
b
ee
n
an
ac
ti
v
e
ar
ea
of
r
ese
ar
ch
[7
-
1
2
]
.
Fig
u
r
e
1
d
ep
icts
an
ex
a
m
p
le
w
h
er
e
th
e
i
m
a
g
e
h
a
s
b
ee
n
u
til
ized
to
ex
tr
ac
t
a
n
atu
r
al
la
n
g
u
ag
e
b
ased
s
in
g
le
s
e
n
te
n
ce
d
escr
ip
tio
n
f
r
o
m
t
h
e
ap
p
ar
en
t
v
i
s
u
al
in
f
o
r
m
atio
n
.
Her
e
t
h
e
s
i
m
p
le
d
esc
r
ip
t
io
n
d
e
m
o
n
s
tr
ate
s
th
e
q
u
ite
r
e
m
ar
k
ab
le
d
ep
th
in
p
er
ce
p
tio
n
o
f
th
e
i
m
a
g
e
i
n
b
o
th
s
y
n
tactica
l
an
d
s
e
m
a
n
t
ic
m
ea
n
i
n
g
w
h
er
e
ap
p
ar
en
tl
y
t
h
e
o
b
j
ec
t
an
d
s
p
atial
co
n
ten
t
s
in
t
h
e
i
m
ag
e
(
e.
g
.
,
p
eo
p
le
an
d
s
tr
ee
t)
ar
e
co
n
n
ec
ted
s
e
m
a
n
tical
l
y
w
it
h
th
e
ac
tio
n
w
a
l
k
in
g
.
T
h
e
c
o
n
ten
t
b
ased
i
m
ag
e
in
ter
p
r
etat
io
n
tas
k
o
f
t
h
is
k
i
n
d
i
s
cr
u
cia
l
in
v
ar
io
u
s
p
r
ac
tical
ap
p
licatio
n
s
s
u
c
h
as
a
u
to
m
at
i
c
i
m
ag
e
i
n
d
ex
i
n
g
,
i
m
a
g
e
-
b
ase
d
w
eb
-
s
ea
r
c
h
i
n
g
,
au
to
m
atic
i
m
ag
e
ca
p
tio
n
in
g
in
n
e
w
s
a
n
d
s
o
cial
m
ed
ia
s
p
h
er
e
an
d
m
o
r
e
i
m
p
o
r
tan
t
l
y
i
n
a
u
to
m
atic
d
ia
g
n
o
s
is
o
f
d
is
ea
s
e
s
f
o
llo
w
ed
b
y
p
o
ten
tial
au
to
m
ated
m
ed
ical
ad
v
ice
g
e
n
er
atio
n
f
r
o
m
b
io
m
ed
ical
i
m
a
g
es
an
d
so
o
n
.
To
f
u
r
th
er
el
u
cid
ate
th
e
p
o
ten
tial
ap
p
licatio
n
s
of
a
u
to
m
ated
i
m
a
g
e
d
escr
ip
tio
n
t
h
e
f
o
l
l
o
w
in
g
m
o
tiv
atio
n
al
ex
a
m
p
les
m
a
y
b
e
n
o
ted
:
in
a
n
i
m
a
g
e
o
f
r
o
ad
w
it
h
co
m
p
licated
tr
af
f
ic
co
n
g
e
s
tio
n
,
a
f
o
cu
s
ed
ex
tr
ac
tio
n
o
f
v
is
u
al
i
n
f
o
r
m
atio
n
m
i
g
h
t
h
elp
w
it
h
s
i
m
p
le
r
etr
ie
v
al
f
ea
tu
r
es
lik
e
n
u
m
b
er
o
f
v
eh
ic
les
o
r
t
y
p
e
o
f
v
e
h
icle
s
o
r
av
er
ag
e
s
ep
ar
ati
o
n
len
g
t
h
b
et
w
ee
n
v
eh
ic
les
i
n
t
h
e
i
m
a
g
e,
o
r
in
a
cr
o
w
d
ed
s
p
ac
e
if
an
y
p
er
s
o
n
p
o
s
es
a
th
r
ea
t
w
it
h
a
g
g
r
e
s
s
i
v
e
g
e
s
t
u
r
e
or
by
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
N
a
tu
r
a
l la
n
g
u
a
g
e
d
escr
ip
tio
n
o
f ima
g
es u
s
in
g
h
y
b
r
id
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Md
.
A
s
ifu
z
z
a
ma
n
Jish
a
n
)
2933
ex
p
o
s
in
g
w
ea
p
o
n
s
,
a
s
i
m
p
le
an
d
q
u
ick
co
n
te
x
t
r
etr
iev
al
of
t
h
e
i
m
a
g
e
i
m
p
l
y
i
n
g
i
m
m
ed
iate
p
o
ten
tial
th
r
ea
t
to
p
u
b
lic
w
o
u
ld
b
e
m
o
s
t
d
esire
d
.
Fig
u
r
e
1
.
E
x
tr
ac
tio
n
o
f
a
s
i
m
p
l
e
n
atu
r
al
la
n
g
u
a
g
e
d
escr
ip
ti
o
n
f
r
o
m
v
is
u
al
d
ata
Giv
e
n
t
h
e
s
cien
tific
a
n
d
p
r
ac
tical
i
m
p
o
r
ta
n
ce
o
f
th
e
n
a
tu
r
al
lan
g
u
a
g
e
b
ased
d
escr
ip
tio
n
o
f
i
m
a
g
es,
it
h
as
b
ee
n
a
v
er
y
d
y
n
a
m
ic
r
esear
ch
en
d
ea
v
o
u
r
w
it
h
to
o
ls
an
d
tech
n
iq
u
e
s
o
f
b
o
th
tr
ad
itio
n
a
l
m
ac
h
in
e
lear
n
in
g
an
d
d
ee
p
m
ac
h
in
e
lear
n
i
n
g
h
a
v
e
b
ee
n
b
r
o
u
g
h
t
to
b
ea
r
o
n
ac
h
iev
in
g
e
x
p
ec
ted
p
er
f
o
r
m
a
n
ce
[
1
3
-
1
5
]
.
Ho
w
e
v
er
,
r
estricte
d
s
co
p
e
o
f
t
h
e
v
o
ca
b
u
lar
ies
f
o
r
d
escr
ib
in
g
v
i
s
u
al
co
n
ten
t
s
li
m
it
s
t
h
e
v
ar
ietie
s
o
f
n
ar
r
ativ
e
s
ab
o
u
t
a
v
is
u
al
s
p
ac
e,
an
d
t
h
e
te
m
p
lat
e
b
ased
i
m
a
g
e
d
escr
ip
tio
n
r
es
tr
icts
co
m
p
le
x
a
n
d
v
ar
ied
s
e
m
atic
in
ter
p
r
etatio
n
,
th
o
u
g
h
d
escr
ip
to
r
m
o
d
els
ca
n
p
r
o
d
u
ce
g
r
a
m
m
atica
ll
y
co
r
r
ec
t
tex
ts
.
Mo
r
eo
v
er
,
th
e
g
r
o
w
i
n
g
s
u
r
g
e
of
i
m
ag
e
a
n
d
v
id
eo
d
atasets
[
1
6
-
18]
p
u
ts
u
p
ch
alle
n
g
in
g
b
ar
s
a
g
ai
n
s
t
th
e
co
m
p
u
tatio
n
a
l
m
o
d
eli
n
g
ef
f
o
r
t
s
to
g
e
n
er
ate
s
y
n
ta
ctica
l
l
y
a
n
d
s
e
m
a
n
ticall
y
v
iab
le
n
at
u
r
al
lan
g
u
a
g
e
b
ase
d
d
escr
ip
tio
n
b
ey
o
n
d
th
e
p
r
e
-
ass
u
m
ed
te
m
p
lates
an
d
clo
s
ed
v
o
ca
b
u
lar
ies.
T
o
s
ee
k
to
cir
cu
m
v
en
t
th
e
s
ai
d
li
m
itatio
n
s
i
n
d
ev
e
lo
p
in
g
a
w
o
r
k
i
n
g
ar
ti
f
icial
n
eu
r
al
s
y
s
te
m
to
o
l
to
g
en
er
ate
n
at
u
r
al
la
n
g
u
a
g
e
b
as
ed
d
escr
ip
tio
n
o
f
i
m
a
g
es,
a
r
ath
er
co
m
p
le
x
m
o
d
el
is
r
eq
u
i
r
ed
to
y
ield
n
o
v
el
tex
t
u
al
d
escr
ip
tio
n
f
r
o
m
v
is
u
al
s
ce
n
e
s
w
it
h
m
u
lti
m
o
d
al
co
m
p
le
x
itie
s
.
I
n
t
h
is
s
t
u
d
y
,
w
e
a
lig
n
w
it
h
t
h
i
s
ap
p
r
o
ac
h
b
y
i
m
p
le
m
e
n
ti
n
g
a
d
ee
p
lear
n
in
g
h
y
b
r
id
i
m
a
g
e
d
escr
ip
to
r
m
o
d
el
co
n
ca
t
e
n
at
in
g
C
o
n
v
o
l
u
tio
n
al
Neu
r
al
Net
w
o
r
k
s
(
C
NN)
[
1
9
,
2
0
]
,
L
o
n
g
S
h
o
r
t
T
e
r
m
M
e
m
o
r
y
(
L
ST
M)
[
2
1
]
an
d
B
i
-
d
ir
ec
tio
n
al
Ne
u
r
al
Net
w
o
r
k
s
(
B
R
NN)
[
2
2
]
m
o
d
els.
W
ith
th
i
s
h
y
b
r
id
m
o
d
el
o
u
r
ap
p
r
o
ac
h
is
to
em
p
lo
y
C
NN
t
o
lear
n
ca
teg
o
r
ical
f
ea
t
u
r
es
f
r
o
m
i
m
ag
e
s
b
y
u
s
i
n
g
s
o
f
t
m
a
x
cla
s
s
i
f
ier
f
o
llo
w
ed
b
y
t
h
e
la
n
g
u
a
g
e
m
o
d
el
L
ST
M
to
lear
n
lo
n
g
er
p
atter
n
s
t
y
p
ical
o
f
n
at
u
r
al
lan
g
u
a
g
e
b
ased
tex
t
s
,
w
h
ic
h
i
n
tu
r
n
f
o
llo
w
ed
b
y
a
B
R
NN
m
o
d
el
to
lear
n
w
o
r
d
r
ep
r
esen
tatio
n
.
T
h
is
co
n
ca
ten
atio
n
o
f
i
m
a
g
e
cla
s
s
i
f
ier
a
n
d
lan
g
u
a
g
e
m
o
d
el
s
en
s
u
r
e
s
lear
n
in
g
of
m
u
lti
m
o
d
al
asp
ec
ts
of
i
m
ag
e
co
n
te
n
ts
alo
n
g
w
i
th
th
e
r
elate
d
n
at
u
r
al
la
n
g
u
a
g
e
te
x
t.
T
h
u
s
,
by
bi
-
d
ir
ec
ti
o
n
al
s
eq
u
en
c
in
g
o
f
i
m
a
g
es
an
d
tex
t
s
,
th
e
d
ee
p
lea
n
in
g
m
o
d
el
alo
n
g
w
i
th
it
s
r
ec
u
r
r
en
t
n
eu
r
al
n
et
w
o
r
k
s
co
u
n
ter
p
ar
ts
lear
n
s
r
elatio
n
b
et
w
ee
n
f
i
n
er
p
o
r
ti
o
n
s
o
f
i
m
a
g
e
alo
n
g
w
it
h
th
e
r
ele
v
an
t
p
o
r
tio
n
o
f
th
e
s
e
n
ten
ce
s
.
Fu
r
t
h
er
,
f
o
r
th
e
lear
n
in
g
an
d
ex
ec
u
t
io
n
o
f
t
h
e
m
o
d
el
w
e
h
a
v
e
u
s
ed
t
h
r
ee
b
e
n
ch
m
ar
k
v
is
u
al
d
atasets
f
o
r
n
at
u
r
al
la
n
g
u
a
g
e
b
ased
d
e
s
c
r
i
p
t
i
o
n
,
e.
g
.
,
Fli
ck
r
8
K,
Fli
c
k
r
3
0
K,
an
d
MSC
OC
O
d
atasets
u
tili
z
in
g
th
e
B
L
E
U
a
n
d
ME
T
E
OR
m
e
tr
ic
[
2
3
]
.
W
e
r
e
p
o
r
t
a
c
h
i
e
v
i
n
g
s
i
g
n
i
f
ica
n
t
i
m
p
r
o
v
e
m
en
t
i
n
t
h
e
tex
t
u
al
r
etr
iev
al
f
r
o
m
t
h
e
d
atase
ts
in
th
e
lear
n
in
g
a
n
d
tes
tin
g
p
h
ase
s
b
y
f
i
n
e
-
tu
n
i
n
g
ar
c
h
itect
u
r
e
an
d
h
y
p
er
p
ar
am
e
ter
s
o
f
t
h
e
m
o
d
el.
2.
M
E
T
H
O
DO
L
O
G
Y
2
.
1
.
B
a
ck
end
co
m
p
uta
t
io
na
l
m
o
d
el
Neu
r
al
s
y
s
te
m
is
i
n
s
till
ed
in
co
m
p
u
tatio
n
al
f
r
a
m
e
w
o
r
k
to
e
m
u
late
th
e
co
g
n
i
tiv
e
f
u
n
ctio
n
s
of
h
u
m
a
n
ce
r
eb
r
u
m
i
n
r
ec
o
g
n
izin
g
a
n
d
p
r
o
ce
s
s
in
g
v
i
s
u
al
i
n
f
o
r
m
ati
o
n
.
I
t
is
i
n
cr
ea
s
i
n
g
l
y
a
p
o
p
u
lar
co
m
p
u
tatio
n
a
l
f
r
a
m
e
w
o
r
k
n
o
w
ad
a
y
s
to
e
x
tr
ac
t
n
atu
r
al
la
n
g
u
a
g
e
b
ase
d
d
escr
ip
tio
n
to
v
is
u
al
in
f
o
r
m
at
io
n
.
T
h
er
e
ar
e
ess
en
tiall
y
th
r
ee
i
m
p
er
ati
v
e
p
ar
ts
co
n
s
i
s
ti
n
g
a
Neu
r
al
S
y
s
te
m
:
A
N
N
(
A
r
ti
f
icial
Neu
r
al
Net
w
o
r
k
)
,
C
NN
(
C
o
n
v
o
lu
t
io
n
a
l
n
eu
r
al
s
y
s
te
m
)
,
a
n
d
R
NN
(
R
e
cu
r
r
en
t N
e
u
r
al
Net
w
o
r
k
)
.
C
o
n
vo
lu
tio
n
a
l
N
eu
r
a
l
N
et
w
o
r
k
(
C
N
N
)
i
s
co
n
t
ai
n
ed
at
least
o
n
e
co
n
v
o
l
u
tio
n
a
l
la
y
er
s
an
d
af
ter
th
a
t
to
o
k
a
f
t
e
r
b
y
at
least
o
n
e
co
m
p
letel
y
ass
o
ciate
d
la
y
er
s
as
in
a
s
ta
n
d
ar
d
m
u
ltil
a
y
er
n
e
u
r
al
s
y
s
te
m
.
C
N
N
b
asicall
y
u
s
e
f
o
r
im
a
g
e
r
ec
o
g
n
i
tio
n
,
v
id
eo
a
n
al
y
s
i
s
s
y
s
te
m
,
n
at
u
r
al
la
n
g
u
a
g
e
p
r
o
ce
s
s
i
n
g
,
an
d
m
a
n
y
m
o
r
e.
I
n
C
NN,
in
p
u
t
la
y
er
,
co
n
v
o
l
u
tio
n
al
la
y
er
,
p
o
llin
g
la
y
er
,
f
u
ll
y
co
n
n
ec
ted
la
y
er
,
an
d
o
u
t
p
u
t
la
y
er
ex
i
s
t
[
2
4
]
(
Fig
u
r
e
2
)
.
I
n
in
p
u
t la
y
er
t
h
er
e
ar
e
th
r
ee
m
ea
s
u
r
e
m
e
n
t
s
a
n
d
th
e
y
ar
e
w
id
t
h
,
h
ei
g
h
t a
n
d
d
ep
th
.
I
t is a
f
r
a
m
e
w
o
r
k
o
f
p
ix
el
estee
m
.
A
t
t
h
at
p
o
in
t
th
e
co
n
v
o
l
u
tio
n
al
la
y
er
ex
is
tin
g
.
A
p
iece
of
t
h
e
p
ictu
r
e
is
ass
o
ciate
d
w
it
h
th
e
f
o
llo
w
in
g
C
o
n
v
o
l
u
tio
n
al
l
a
y
er
in
li
g
h
t
of
th
e
f
ac
t
th
a
t
if
e
v
er
y
o
n
e
o
f
t
h
e
p
ix
el
s
o
f
t
h
e
in
f
o
is
as
s
o
ciate
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
3
2
-
2940
2934
w
it
h
th
e
C
o
n
v
o
l
u
tio
n
al
la
y
er
.
Fil
ter
,
Ker
n
el,
o
r
Featu
r
e
Dete
cto
r
is
a
li
ttle
m
atr
i
x
u
s
ed
f
o
r
h
ig
h
li
g
h
ts
lo
ca
tio
n
.
Af
ter
co
n
v
o
lu
t
io
n
al
la
y
er
,
at
th
at
p
o
in
t
t
h
e
p
o
o
lin
g
la
y
er
p
ar
t
ex
is
ts
.
P
o
o
l
L
a
y
er
p
la
y
s
o
u
t
a
ca
p
ac
it
y
to
d
ec
r
ea
s
e
th
e
s
p
atial
m
ea
s
u
r
e
m
en
ts
o
f
th
e
i
n
f
o
r
m
atio
n
,
an
d
th
e
co
m
p
u
tatio
n
a
l
u
n
p
r
ed
ictab
ilit
y
o
f
o
u
r
m
o
d
el.
W
h
at’
s
m
o
r
e,
it
ad
d
itio
n
al
l
y
c
o
n
tr
o
ls
o
v
er
f
itti
n
g
.
Af
ter
p
o
o
lin
g
la
y
er
,
f
u
ll
y
co
n
n
ec
ted
la
y
er
p
ar
t
ex
is
ti
n
g
a
n
d
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
in
ter
f
a
ce
each
n
eu
r
o
n
in
o
n
e
la
y
er
to
each
n
e
u
r
o
n
in
a
n
o
th
er
la
y
er
.
T
h
e
last
f
u
ll
y
co
n
n
ec
ted
la
y
er
u
ti
lizes
a
s
o
f
t
m
a
x
i
n
itia
tio
n
w
o
r
k
f
o
r
ch
ar
ac
ter
i
zin
g
t
h
e
p
r
o
d
u
ce
d
h
i
g
h
li
g
h
ts
o
f
th
e
in
f
o
r
m
atio
n
p
ictu
r
e
in
to
d
if
f
er
en
t
clas
s
es
i
n
lig
h
t
o
f
th
e
tr
ain
in
g
d
ataset
an
d
af
ter
co
m
p
leti
n
g
t
h
is
la
y
er
th
en
w
e
g
et
a
n
o
u
tp
u
t
[
2
5
]
.
Fig
u
r
e
2
.
C
o
n
v
o
l
u
tio
n
al
n
e
u
r
a
l n
et
w
o
r
k
(
C
N
N)
R
ec
u
r
r
en
t
N
eu
r
a
l
N
et
w
o
r
k
(
R
N
N
)
is
e
m
p
lo
y
ed
to
m
a
k
e
u
ti
lizatio
n
o
f
co
n
s
ec
u
t
iv
e
d
ata.
R
ec
u
r
r
en
t
N
e
u
r
a
l
Net
w
o
r
k
tak
e
s
t
h
e
p
r
ev
io
u
s
o
u
tp
u
t
o
r
h
id
d
en
s
tate
as
i
n
p
u
ts
.
R
NN
b
asicall
y
u
t
ilized
f
o
r
la
n
g
u
a
g
e
d
em
o
n
s
tr
ati
n
g
a
n
d
cr
ea
tin
g
c
o
n
ten
t,
m
ac
h
i
n
e
tr
an
s
latio
n
,
s
p
ee
ch
r
ec
o
g
n
i
tio
n
,
g
en
er
ati
n
g
i
m
a
g
e
d
escr
ip
tio
n
.
W
h
en
it
m
a
k
es
a
d
ec
is
io
n
,
it
th
i
n
k
s
ab
o
u
t
th
e
cu
r
r
en
t
in
p
u
t
s
tate
a
n
d
f
u
r
t
h
er
m
o
r
e
w
h
at
it
h
as
g
ai
n
ed
f
r
o
m
th
e
i
n
f
o
r
m
atio
n
s
o
u
r
ce
s
it
r
ec
eiv
ed
a
l
r
e
a
d
y
[
2
6
]
.
A
r
ec
u
r
r
e
n
t
n
eu
r
al
n
et
w
o
r
k
is
ab
le
to
r
e
m
e
m
b
er
p
r
o
ce
s
s
e
s
ex
ac
tl
y
w
h
ile
a
w
o
r
d
p
r
o
ce
s
s
r
u
n
n
i
n
g
b
ec
au
s
e
o
f
it
h
a
s
in
te
r
n
al
m
e
m
o
r
y
.
I
t
also
p
r
ed
ict
w
h
ic
h
ch
ar
ac
ter
w
il
l
b
e
co
m
e
i
n
n
ex
t
a
n
d
p
r
o
d
u
ce
s
o
u
tp
u
t,
co
p
ies
th
e
o
u
tp
u
t
a
n
d
lo
o
p
s
it
b
ac
k
i
n
to
th
e
n
et
wo
r
k
p
ar
t.
Mo
r
eo
v
er
,
in
a
R
NN
h
a
v
e
t
w
o
i
n
p
u
ts
,
p
r
esen
t a
n
d
th
e
r
ec
en
t p
as
t [
2
7
]
.
Lo
n
g
s
h
o
r
t
-
term
mem
o
r
y
(
LS
T
M)
is
a
s
p
ec
ial
k
i
n
d
o
f
R
N
N
e
n
ab
led
to
lear
n
lo
n
g
ter
m
d
ep
en
d
en
cie
s
.
I
t
is
w
id
el
y
u
s
ed
b
ec
a
u
s
e
of
it
s
f
ea
t
u
r
e
of
r
e
m
e
m
b
er
in
g
i
n
f
o
r
m
atio
n
f
o
r
lo
n
g
p
er
io
d
s
of
ti
m
e
[
2
8
]
.
T
h
is
is
d
o
n
e
by
cr
ea
ti
n
g
s
p
ec
ial
m
o
d
u
les
t
h
at
i
s
d
esi
g
n
ed
to
allo
w
i
n
f
o
r
m
atio
n
to
b
e
g
a
ted
-
i
n
a
n
d
g
a
t
ed
-
o
u
t
w
h
e
n
n
ee
d
ed
.
Un
li
k
e
tr
ad
itio
n
al
R
NN,
L
S
T
M
s
to
r
es
in
f
o
r
m
atio
n
u
s
i
n
g
a
m
e
m
o
r
y
ce
ll
w
it
h
li
n
ea
r
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
f
o
llo
w
i
n
g
s
y
s
te
m
of
E
q
u
at
io
n
(
1
)
g
o
v
er
n
s
th
e
ac
ti
v
itie
s
o
f
L
ST
M
w
h
ic
h
in
cl
u
d
es
t
h
e
d
esig
n
o
f
a
m
e
m
o
r
y
ce
ll u
s
i
n
g
lo
g
i
s
tic
an
d
li
n
ea
r
u
n
its
w
i
t
h
m
u
ltip
le
in
ter
ac
t
io
n
s
[
2
9
]
.
=
(
(
)
+
(
ℎ
)
ℎ
−
1
+
(
)
−
1
+
(
)
)
=
(
(
)
+
(
ℎ
)
ℎ
−
1
+
(
)
−
1
+
(
)
)
=
∗
−
1
+
∗
ℎ
(
(
)
+
(
ℎ
)
ℎ
−
1
+
(
)
)
(
1
)
=
(
(
)
+
(
ℎ
)
ℎ
−
1
+
(
)
−
1
+
(
)
)
ℎ
=
∗
ℎ
(
)
Her
e
I
,
f
,
o
an
d
c
ar
e
r
esp
ec
tiv
el
y
t
h
e
i
n
p
u
t,
f
o
r
g
e
t,
o
u
tp
u
t
g
ate
an
d
m
e
m
o
r
y
ce
ll
ac
ti
v
ati
o
n
v
ec
to
r
s
.
E
ac
h
m
e
m
o
r
y
ce
l
l
c
t
h
as
it
s
n
e
t
in
p
u
t
m
o
d
u
lated
by
t
h
e
ac
ti
v
it
y
of
an
i
n
p
u
t
g
ate,
a
n
d
h
as
its
o
u
tp
u
t
m
o
d
u
lated
by
t
h
e
ac
t
iv
i
t
y
o
f
a
n
o
u
tp
u
t
g
ate.
T
h
ese
in
p
u
t
a
n
d
o
u
tp
u
t
g
at
es
p
r
o
v
id
e
a
co
n
tex
t
-
s
e
n
s
itiv
e
w
a
y
to
u
p
d
ate
th
e
co
n
te
n
t
s
o
f
a
m
e
m
o
r
y
ce
l
l.
T
h
e
f
o
r
g
et
g
ate
m
o
d
u
late
s
a
m
o
u
n
t
o
f
ac
t
i
v
at
io
n
o
f
m
e
m
o
r
y
ce
ll
k
ep
t
f
r
o
m
th
e
p
r
ev
io
u
s
ti
m
e
s
tep
,
p
r
o
v
id
in
g
a
m
et
h
o
d
to
q
u
ic
k
l
y
er
ase
th
e
co
n
ten
ts
o
f
m
e
m
o
r
y
ce
ll
s
.
T
h
u
s
t
h
e
r
es
u
lt
in
g
in
ter
p
la
y
o
f
t
h
ese
g
ate
s
p
av
e
s
th
e
w
a
y
to
lear
n
i
n
g
o
f
p
atter
n
s
w
it
h
lo
n
g
d
ep
en
d
e
n
cies
e
s
p
ec
iall
y
f
ea
t
u
r
ed
in
th
e
n
at
u
r
al
la
n
g
u
a
g
es
[
3
0
]
.
2
.
2
.
Im
ple
m
ent
a
t
io
n
2
.
2
.
1
.
Repre
s
ent
a
t
io
n
R
ep
r
esen
ti
n
g
i
m
a
g
e
i
s
m
o
s
t
i
m
p
o
r
ta
n
t
p
ar
t
f
o
r
i
m
a
g
e
p
r
o
ce
s
s
i
n
g
an
d
w
e
g
et
a
lo
t
o
f
id
ea
s
to
r
e
v
ie
w
m
a
n
y
r
ec
en
t
w
o
r
k
s
[
9
]
.
W
e
w
atc
h
t
h
at
s
e
n
te
n
ce
d
escr
ip
t
io
n
m
ak
e
v
i
s
it
r
ef
er
en
ce
s
to
o
b
j
ec
ts
an
d
th
eir
attr
ib
u
tes [
3
1
]
.
T
h
e
C
NN
is
p
r
e
-
p
r
ep
ar
ed
o
n
I
m
a
g
eNe
t [
1
6
]
an
d
f
i
n
et
u
n
ed
o
n
th
e
2
0
0
class
es o
f
t
h
e
I
m
a
g
e
Ne
t
Dete
ctio
n
C
h
alle
n
g
e
[
3
2
]
.
W
e
m
ai
n
ta
i
n
t
h
e
tec
h
n
iq
u
e
f
o
r
G
ir
s
h
ic
k
et
al
.
[
3
3
]
to
d
etec
t
e
ac
h
o
b
j
ec
t
in
ea
c
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
N
a
tu
r
a
l la
n
g
u
a
g
e
d
escr
ip
tio
n
o
f ima
g
es u
s
in
g
h
y
b
r
id
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Md
.
A
s
ifu
z
z
a
ma
n
Jish
a
n
)
2935
i
m
a
g
e
w
it
h
a
R
eg
io
n
C
o
n
v
o
l
u
tio
n
al
Neu
r
al
Net
w
o
r
k
(
R
C
NN)
.
T
h
e
R
C
NN
m
o
d
el
h
a
s
t
w
o
p
ar
ts
,
a
r
eg
io
n
p
r
o
p
o
s
al
n
et
w
o
r
k
an
d
an
o
th
e
r
o
n
e
is
b
in
ar
y
m
a
s
k
cla
s
s
i
f
i
er
.
Fo
llo
w
in
g
Kar
p
ath
y
et
al
.
[
3
1
]
,
w
e
u
s
e
th
e
p
r
im
ar
y
1
9
id
en
tifie
d
ar
ea
d
esp
ite
th
e
w
h
o
le
p
ict
u
r
e.
T
h
en
co
m
p
u
te
t
h
e
r
ep
r
esen
tatio
n
i
n
l
ig
h
t
o
f
t
h
e
p
ix
els
I
b
in
s
id
e
ea
c
h
b
o
u
n
d
i
n
g
b
o
x
as t
ak
es
a
f
ter
:
=
[
(
)
]
+
(
2
)
T
h
e
CNN
(
I
b
)
ch
an
g
es
t
h
e
p
ix
els
i
n
s
id
e
th
e
b
o
u
n
d
in
g
box
(
I
b
)
to
4
0
9
6
-
d
im
e
n
s
io
n
al
e
n
ac
t
m
e
n
t
of
th
e
f
u
l
l
y
co
n
n
ec
ted
la
y
er
in
a
s
p
lit
s
ec
o
n
d
b
ef
o
r
e
t
h
e
clas
s
i
f
ier
.
T
h
e
C
NN
p
ar
a
m
eter
s
θ
c
co
n
tain
ar
o
u
n
d
60
m
illi
o
n
p
ar
a
m
eter
s
.
T
h
e
m
atr
ix
W
m
h
a
s
m
ea
s
u
r
e
m
en
t
s
h
×
4
0
9
6
,
w
h
er
e
h
is
t
h
e
e
x
te
n
t
o
f
th
e
m
u
lt
i
m
o
d
al
in
s
er
ti
n
g
s
p
ac
e.
E
ac
h
i
m
ag
e
r
e
p
r
esen
t a
s
h
-
d
i
m
en
s
io
n
a
l
v
ec
t
o
r
s
.
R
ep
r
esen
ti
n
g
s
e
n
te
n
ce
i
s
a
cr
u
cial
p
a
r
t
o
f
o
u
r
m
o
d
el.
W
e
u
tili
ze
d
a
B
id
ir
ec
tio
n
al
R
ec
u
r
r
en
t
Ne
u
r
a
l
N
e
tw
o
r
k
(
B
R
NN)
[
2
2
]
to
c
o
m
p
u
te
th
e
w
o
r
d
r
ep
r
esen
tatio
n
.
B
id
ir
ec
tio
n
al
R
ec
u
r
r
en
t
Neu
r
a
l
Net
w
o
r
k
(
B
R
NN)
is
a
p
ar
t
of
R
N
N
s
ec
tio
n
a
n
d
wh
ich
is
u
s
e
a
f
i
n
ite
s
eq
u
e
n
ce
to
p
r
ed
ictio
n
.
In
B
R
NN
m
o
d
el,
th
er
e
ar
e
lab
el
ea
c
h
ele
m
e
n
t
of
th
e
s
eq
u
e
n
ce
b
ase
d
o
n
th
e
p
ast
a
n
d
f
u
t
u
r
e
co
n
t
ex
t
ele
m
en
t.
B
R
NN
co
n
d
u
cts
th
is
s
eq
u
e
n
ci
n
g
b
y
clo
s
e
-
o
u
tp
u
t
o
f
t
w
o
R
NN
s
a
n
d
o
n
e
p
r
o
ce
s
s
in
g
of
th
e
s
eq
u
e
n
ce
is
f
r
o
m
le
f
t
to
r
ig
h
t,
th
e
an
o
th
er
s
eq
u
e
n
ce
f
r
o
m
r
ig
h
t
to
lef
t.
T
h
e
j
o
in
ed
o
u
tp
u
t
s
ar
e
t
h
e
p
r
ed
ictio
n
o
f
th
e
g
i
v
e
n
tar
g
et
s
i
g
n
al
s
.
Fo
r
o
u
r
m
o
d
el,
th
e
B
R
NN
ta
k
es
a
s
eq
u
en
ce
o
f
N
w
o
r
d
s
an
d
t
h
en
i
t
tr
an
s
f
o
r
m
s
ea
ch
to
h
-
d
i
m
e
n
s
io
n
al
v
ec
to
r
.
Utiliz
in
g
t
h
e
lis
t
t
=
1
.
.
.
N
to
in
d
icate
th
e
s
it
u
atio
n
o
f
a
w
o
r
d
in
a
s
en
te
n
ce
,
th
e
e
x
ac
t s
h
ap
e
o
f
th
e
B
R
NN
i
s
as p
er
th
e
f
o
llo
w
i
n
g
:
=
=
(
+
)
ℎ
=
(
+
ℎ
−
1
+
)
(
3
)
ℎ
=
(
+
ℎ
+
1
+
)
=
(
(
ℎ
+
ℎ
)
+
)
T
h
e
w
ei
g
h
ts
W
w
d
eter
m
i
n
e
a
w
o
r
d
in
s
e
r
ti
n
g
n
et
w
o
r
k
th
at
we
in
s
tate
w
it
h
300
-
d
i
m
en
s
io
n
al
w
o
r
d
2
v
ec
[
3
4
]
w
ei
g
h
ts
an
d
k
e
ep
f
ix
ed
b
ec
au
s
e
o
f
o
v
er
f
itti
n
g
co
n
ce
r
n
s
.
is
a
p
o
in
ter
co
lu
m
n
v
ec
to
r
th
at
h
a
s
a
s
in
g
le
o
n
e
at
t
h
e
r
ec
o
r
d
o
f
t
h
e
t
-
t
h
w
o
r
d
i
n
a
w
o
r
d
v
o
ca
b
u
lar
y
.
T
h
e
B
R
NN
co
m
p
r
i
s
es
o
f
t
w
o
i
n
d
ep
en
d
en
t
s
tr
ea
m
s
o
f
h
a
n
d
li
n
g
,
o
n
e
m
o
v
in
g
lef
t
to
r
i
g
h
t
(
ℎ
)
an
d
t
h
e
o
th
er
r
ig
h
t
to
lef
t
(
ℎ
)
.
W
e
s
et
th
e
ac
tiv
at
io
n
f
u
n
ctio
n
f
to
th
e
r
ec
ti
f
ier
lin
ea
r
u
n
it (
R
e
L
U)
.
2
.
2
.
2
.
Dec
o
din
g
Dec
o
d
in
g
co
n
s
id
er
s
a
p
ictu
r
e
f
r
o
m
t
h
e
tr
ain
in
g
s
et
a
n
d
its
co
m
p
ar
i
n
g
s
en
ten
ce
.
W
e
ar
e
u
lti
m
atel
y
in
ter
ested
i
n
p
r
o
d
u
cin
g
s
n
ip
p
ets
o
f
co
n
te
n
t
o
f
s
i
n
g
le
w
o
r
d
s
,
w
e
m
i
g
h
t
w
an
t
to
alig
n
ex
ten
d
ed
,
ad
j
ac
en
t
s
eq
u
en
ce
s
o
f
w
o
r
d
s
to
a
s
in
g
le
b
o
u
n
d
in
g
b
o
x
.
We
can
tr
a
n
s
late
th
e
a
m
o
u
n
t
v
T
s
t
as
t
h
e
u
n
n
o
r
m
al
ized
lo
g
lik
eli
h
o
o
d
of
th
e
t
-
th
w
o
r
d
d
ep
ictin
g
a
n
y
o
f
th
e
b
o
u
n
d
i
n
g
b
o
x
es
in
th
e
i
m
ag
e.
No
te
th
a
t t
h
e
n
ai
v
e
ar
r
an
g
e
m
en
t
th
at
as
s
i
g
n
s
ea
ch
w
o
r
d
f
r
ee
l
y
to
th
e
h
i
g
h
est
s
co
r
i
n
g
lo
ca
le
is
lack
i
n
g
in
li
g
h
t
of
t
h
e
f
ac
t
t
h
at
it
p
r
o
m
p
t
s
w
o
r
d
s
g
etti
n
g
s
ca
tter
ed
co
n
f
licti
n
g
l
y
to
v
ar
io
u
s
r
eg
io
n
s
.
W
e
r
eg
ar
d
th
e
g
e
n
u
i
n
e
ar
r
an
g
e
m
e
n
ts
as
i
n
ac
ti
v
e
f
ac
to
r
s
in
a
Ma
r
k
o
v
R
an
d
o
m
Field
(
M
R
F)
w
h
er
e
t
h
e
b
in
ar
y
co
llab
o
r
atio
n
s
b
et
w
ee
n
n
eig
h
b
o
r
in
g
w
o
r
d
s
u
r
g
e
an
ar
r
an
g
e
m
en
t
to
a
s
i
m
ilar
d
is
tr
i
ct.
So
lid
l
y
,
g
i
v
en
a
s
e
n
te
n
ce
with
N
w
o
r
d
s
a
n
d
a
p
ictu
r
e
w
it
h
M
j
u
m
p
i
n
g
b
o
x
es,
w
e
p
r
esen
t
th
e
i
n
ac
ti
v
e
ar
r
an
g
e
m
en
t
v
ar
iab
le
a
j
€
1
.
.
.
M
f
o
r
j
=
1
...N
.
Her
e,
d
ef
in
e
a
MR
F
i
n
a
ch
ai
n
s
tr
u
ctu
r
e
alo
n
g
t
h
e
s
e
n
te
n
ce
as tak
e
s
af
t
er
:
(
)
=
∑
(
)
+
∑
(
,
+
1
)
(
=
1
…
−
1
)
(
=
1
…
)
(
=
)
=
(
4
)
(
,
+
1
)
=
[
=
+
1
]
Her
e,
β
i
s
a
h
y
p
er
p
ar
a
m
eter
t
h
at
co
n
tr
o
ls
th
e
p
ar
tialit
y
to
w
ar
d
s
lo
n
g
er
w
o
r
d
p
h
r
ase
s
.
T
h
is
p
ar
am
eter
en
ab
les
u
s
to
i
n
tr
o
d
u
ce
b
et
wee
n
s
in
g
le
-
w
o
r
d
ar
r
an
g
e
m
e
n
t
s
(
β
=
0
)
a
n
d
ad
j
u
s
ti
n
g
t
h
e
w
h
o
le
s
en
te
n
ce
to
a
s
o
litar
y
,
m
a
x
i
m
all
y
s
co
r
i
n
g
a
r
ea
w
h
e
n
β
is
e
x
te
n
s
i
v
e.
T
h
e
y
ield
o
f
th
is
p
r
o
ce
d
u
r
e
is
a
s
et
o
f
i
m
ag
e
ar
ea
s
ex
p
lain
ed
w
it
h
f
r
ag
m
en
ts
o
f
c
o
n
ten
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
3
2
-
2940
2936
2
.
2
.
3
.
O
pti
m
iza
t
io
n
We
u
tili
ze
S
GD
w
it
h
m
i
n
i
b
atch
of
100
p
ictu
r
e
s
e
n
te
n
ce
s
ets
f
u
r
th
er
m
o
r
e,
s
p
ee
d
of
0
.
9
to
o
p
tim
izatio
n
to
th
e
ali
g
n
m
e
n
t
m
o
d
el.
W
e
cr
o
s
s
-
ap
p
r
o
v
e
th
e
lear
n
in
g
r
ate
a
n
d
th
e
w
ei
g
h
t
r
o
t.
W
e
lik
e
w
i
s
e
u
tili
ze
d
r
o
p
o
u
t
r
eg
u
lar
izatio
n
in
all
la
y
er
s
w
it
h
th
e
e
x
ce
p
tio
n
o
f
in
th
e
r
ec
u
r
r
e
n
t
la
y
er
s
[
3
5
]
an
d
clip
g
r
ad
ien
t
ele
m
e
n
t
w
i
s
e
at
5
(
ess
e
n
tia
l)
.
T
h
e
g
en
er
ati
v
e
R
NN
is
h
ar
d
e
r
to
o
p
ti
m
iza
tio
n
b
ec
au
s
e
o
f
t
h
e
w
o
r
d
f
r
eq
u
e
n
c
y
d
if
f
er
e
n
ce
b
et
w
ee
n
u
n
co
m
m
o
n
w
o
r
d
s
a
n
d
co
m
m
o
n
w
o
r
d
s
.
W
e
ac
co
m
p
li
s
h
t
h
e
b
est
o
u
tco
m
es
u
t
ilizi
n
g
R
MSp
r
o
p
[
3
6
]
.
T
h
at
m
et
h
o
d
is
a
v
er
s
at
ile
ad
v
a
n
ce
s
ize
s
tr
ateg
y
t
h
at
s
c
ale
s
t
h
e
r
ef
r
e
s
h
o
f
ea
c
h
w
ei
g
h
t
b
y
a
r
u
n
n
i
n
g
n
o
r
m
al
o
f
i
ts
g
r
a
d
ien
t
s
tan
d
ar
d
.
3.
SI
M
UL
AT
I
O
N
3
.
1
.
Da
t
a
s
et
W
e
u
tili
ze
t
h
e
Fl
ick
r
8
K
[
1
7
]
,
Fli
c
k
r
3
0
K
[
2
3
]
an
d
MSC
O
C
O
[
1
8
]
d
atasets
f
o
r
o
u
r
ex
p
er
im
e
n
t
.
Fli
ck
r
8
K
d
ataset
co
n
ta
in
8
,
0
0
0
,
Fli
ck
r
3
0
K
d
ataset
co
n
tai
n
3
1
,
0
0
0
an
d
MSC
OC
O
d
ata
s
et
co
n
tain
1
2
3
,
0
0
0
i
m
a
g
es.
Fo
r
F
lick
r
8
K
a
n
d
Fli
c
k
r
3
0
K
d
ataset,
w
e
u
til
ize
1
,
0
0
0
p
ictu
r
es f
o
r
v
al
id
atio
n
,
1
,
0
0
0
f
o
r
test
in
g
a
n
d
th
e
r
est
p
ictu
r
es
f
o
r
tr
ain
i
n
g
.
Fo
r
MS
C
OC
O,
we
u
til
ize
5
,
0
0
0
im
ag
e
s
f
o
r
v
alid
atio
n
an
d
test
in
g
b
o
th
p
ar
ts
.
We
u
s
e
NVI
DI
A
G1
G
A
MI
N
G
GP
U
f
o
r
t
r
a
i
n
th
e
d
ataset.
3
.
2
.
Da
t
a
p
re
pro
ce
s
s
ing
We
p
r
e
p
r
o
ce
s
s
o
u
r
d
ataset
b
ef
o
r
e
tr
ain
in
g
tas
k
.
We
co
n
v
er
t
all
s
en
te
n
ce
s
of
o
u
r
d
ataset
to
lo
w
er
ca
s
e,
d
is
ca
r
d
n
o
n
-
alp
h
a
n
u
m
er
ic
ch
a
r
ac
ter
s
.
W
e
f
ilter
w
o
r
d
s
w
h
ic
h
is
o
cc
u
r
5
ti
m
e
s
i
n
t
h
e
tr
ain
i
n
g
s
et,
w
h
ic
h
r
es
u
lt
in
2
5
3
8
w
o
r
d
s
f
o
r
Fli
c
k
r
8
K,
7
4
1
4
w
o
r
d
s
f
o
r
Fli
c
k
r
3
0
K,
an
d
8
7
9
1
w
o
r
d
s
f
o
r
MSC
O
C
O
d
ataset.
3
.
3
.
I
m
a
g
e
p
ro
ce
s
s
ing
W
e
r
esized
th
e
i
m
a
g
e
s
of
all
o
u
r
d
atasets
to
en
s
u
r
e
b
etter
g
en
er
alit
y
a
n
d
to
av
o
id
an
y
n
u
m
er
ica
l
in
co
n
s
is
te
n
c
y
d
u
r
in
g
tr
ai
n
i
n
g
an
d
test
i
n
g
p
h
a
s
es.
We
u
s
e
r
a
w
i
m
a
g
e
f
ile
s
of
each
d
at
as
et
al
o
n
g
s
id
e
J
SON
f
ile
an
d
VGG
C
N
N
f
ea
tu
r
e
s
f
o
r
o
u
r
th
r
ee
b
en
c
h
m
ar
k
d
ataset
Fl
ick
r
8
K,
Fli
ck
r
3
0
K,
an
d
MS
C
OC
O.
T
h
e
in
p
u
t
is
a
d
ataset
of
i
m
a
g
es
a
n
d
5
s
e
n
ten
ce
d
escr
ip
tio
n
s
w
h
ic
h
w
er
e
co
llected
w
i
th
Am
az
o
n
Me
c
h
a
n
ical
T
u
r
k
.
In
p
ar
ticu
lar
,
t
h
is
co
d
e
b
ase
is
s
et
up
f
o
r
Fli
ck
r
8
K,
Fli
c
k
r
3
0
K,
an
d
MS
C
O
C
O
d
ata
s
ets.
I
n
t
h
e
tr
ai
n
i
n
g
s
ec
tio
n
,
all
o
f
i
m
a
g
es
ar
e
f
ed
as
i
n
p
u
t
t
o
R
NN
a
n
d
R
NN
a
s
k
ed
to
p
r
e
d
ict
th
e
w
o
r
d
of
th
e
s
e
n
ten
ce
s
.
Fo
r
th
e
p
r
ed
ictio
n
p
ar
t,
im
a
g
es
ar
e
p
ass
ed
to
R
NN
an
d
R
NN
g
e
n
er
ates
t
h
e
s
en
ten
ce
w
o
r
d
at
a
ti
m
e
a
n
d
w
e
g
et
r
esu
lt
o
f
o
u
r
ev
alu
a
tio
n
w
it
h
B
L
E
U
an
d
M
E
T
E
OR
s
ca
le.
W
e
u
s
e
j
s
o
n
,
d
ateti
m
e,
p
ic
k
le,
m
at
h
,
ca
f
f
e,
n
u
m
p
y
,
s
c
i
p
y
,
te
n
s
o
r
f
lo
w
,
co
d
e,
s
o
c
k
et,
ar
g
p
ar
s
e,
o
s
,
an
d
ti
m
e
lib
r
ar
y
f
o
r
o
u
r
i
m
a
g
e
to
tex
t
g
e
n
er
atio
n
w
o
r
k
.
W
e
also
u
s
e
v
g
g
_
f
ea
t
s
.
m
at
w
h
ic
h
is
a
.
m
a
t
f
ile
a
n
d
th
at
s
to
r
es
th
e
C
N
N
f
ea
tu
r
e
s
.
W
e
u
s
e
5
1
2
h
id
d
en
la
y
er
s
an
d
f
r
o
m
i
m
a
g
er
n
n
.
d
ata_
p
r
o
v
id
er
u
s
e
g
etDa
taP
r
o
v
id
er
f
o
r
th
is
p
r
o
j
ec
t.
W
e
also
i
n
v
o
l
v
e
s
o
lv
er
,
d
ec
o
d
e
g
en
er
ato
r
,
ev
al_
s
p
lit
f
r
o
m
t
h
e
i
m
a
g
er
n
n
.
d
ata
_
p
r
o
v
id
er
.
We
also
u
s
e
i
m
r
ea
d
,
i
m
r
e
s
ize
f
o
r
im
a
g
e
r
esizin
g
or
r
esh
ap
in
g
.
Af
ter
co
m
p
letin
g
r
esize
of
i
m
ag
e
s
,
th
en
we
a
tte
m
p
t
to
tr
ain
th
e
w
h
o
le
d
ata
s
et.
As
r
eg
ar
d
s
to
t
h
e
co
m
p
u
tatio
n
al
d
u
r
atio
n
,
Fli
ck
r
8
K
tak
e
s
1
d
a
y
,
Fli
ck
r
3
0
K
tak
es
1
0
d
ay
s
,
an
d
MS
C
OC
O
tak
e
s
2
4
d
ay
s
to
co
m
p
lete
th
e
tr
ai
n
i
n
g
o
f
w
h
o
le
d
ataset.
4.
RE
SU
L
T
S
W
e
in
v
esti
g
ate
th
e
ab
ilit
y
o
f
t
h
e
w
o
r
k
in
g
h
y
b
r
id
d
ee
p
lear
n
in
g
m
o
d
el
b
y
ex
p
lo
r
in
g
h
o
w
w
ell
it
ca
n
g
en
er
ate
r
ea
li
s
tic
d
escr
ip
tio
n
of
th
e
tes
t
i
m
a
g
es.
We
tr
ain
e
d
o
u
r
m
o
d
el
to
lear
n
th
e
r
ela
tio
n
b
et
w
ee
n
f
i
n
er
p
o
r
tio
n
s
of
i
m
a
g
e
alo
n
g
w
it
h
t
h
e
r
elev
a
n
t
p
o
r
tio
n
o
f
t
h
e
s
en
t
en
ce
s
.
W
e
p
r
esen
t
th
e
B
L
E
U
a
n
d
ME
T
E
OR
s
co
r
e
to
ass
ess
th
e
p
er
f
o
r
m
an
ce
o
f
o
u
r
m
o
d
el.
T
h
ese
tech
n
iq
u
es
allo
w
u
s
to
co
m
p
u
te
a
s
co
r
e
th
e
m
ea
s
u
r
es
h
o
w
s
en
s
ib
le
is
th
e
i
m
a
g
e
d
escr
ip
tio
n
s
.
T
h
e
in
tu
i
tio
n
is
to
m
e
asu
r
e
h
o
w
c
lo
s
e
th
e
m
o
d
el
g
en
er
ated
s
e
n
ten
c
e
m
atc
h
es
w
it
h
a
n
y
o
f
th
e
f
iv
e
r
ef
er
en
ce
s
e
n
te
n
ce
s
p
r
o
v
id
ed
w
i
th
t
h
e
d
ataset.
W
e
r
ep
o
r
t
th
e
s
e
e
v
alu
a
tio
n
m
etr
ics o
f
o
u
r
m
o
d
el
an
d
p
r
esen
t a
co
m
p
ar
is
o
n
w
i
th
o
t
h
er
s
t
ate
-
of
-
t
h
e
-
ar
t
r
esu
lts
.
We
tr
ain
o
u
r
m
o
d
el
on
F
lick
r
8
K
an
d
Fli
c
k
r
3
0
K
d
atasets
a
n
d
o
b
s
er
v
e
th
e
ev
a
lu
at
io
n
of
f
u
ll
i
m
a
g
e
p
r
ed
ictio
n
s
o
n
1000
test
i
m
a
g
es.
T
h
e
B
L
E
U
-
1,
2,
3,
4
ev
alu
a
tio
n
s
co
r
es
an
d
ME
T
E
OR
m
e
tr
ic
s
co
r
e
ar
e
ass
es
s
ed
an
d
a
co
m
p
ar
is
o
n
o
f
t
h
e
r
es
u
lt
s
w
it
h
o
th
er
s
tate
-
of
-
t
h
e
-
ar
t
r
esu
lts
is
d
eli
n
ea
ted
in
T
a
b
l
e
1
an
d
T
a
b
l
e
2.
Fo
r
th
e
ex
p
er
i
m
e
n
t
s
,
1
,
0
0
0
im
ag
es
f
r
o
m
th
e
d
atase
ts
ar
e
u
s
e
d
f
o
r
test
i
n
g
an
d
v
alid
atio
n
p
u
r
p
o
s
e
an
d
th
e
r
ests
f
o
r
th
e
tr
ain
i
n
g
p
u
r
p
o
s
e.
Her
e
in
t
h
e
T
a
b
l
e
s
,
(
-
)
in
d
icate
s
an
u
n
k
n
o
w
n
m
etr
ic
o
f
t
h
i
s
d
ataset
r
esu
lt.
Fro
m
t
h
e
e
x
p
er
i
m
e
n
t
o
f
tr
ain
in
g
o
u
r
m
o
d
el
o
n
M
SC
O
C
O
d
ataset,
w
e
s
t
u
d
y
th
e
ev
a
lu
ati
o
n
o
f
f
u
l
l
i
m
a
g
e
p
r
ed
ictio
n
s
o
n
5
,
0
0
0
te
s
t
i
m
ag
e
s
.
T
h
e
B
L
E
U
-
1
,
2
,
3
,
4
ev
alu
at
io
n
s
co
r
es
a
n
d
ME
T
E
OR
m
e
tr
ic
s
co
r
e
ar
e
ass
ess
ed
a
n
d
a
co
m
p
ar
is
o
n
of
t
h
e
r
es
u
lts
w
it
h
o
th
er
s
tate
of
th
e
ar
t
r
esu
lts
is
d
eli
n
ea
ted
in
T
a
b
l
e
3.
Fo
r
th
i
s
ex
p
er
i
m
e
n
t,
5
,
0
0
0
im
a
g
es
f
r
o
m
th
e
d
ataset
s
ar
e
u
s
ed
f
o
r
test
in
g
an
d
v
alid
atio
n
p
u
r
p
o
s
e
an
d
th
e
r
ests
ar
e
u
s
ed
f
o
r
th
e
tr
ain
in
g
p
u
r
p
o
s
e.
Her
e
in
th
e
T
a
b
l
e
3
,
(
-
)
also
in
d
icate
s
an
u
n
k
n
o
w
n
m
etr
ic
o
f
t
h
is
d
ataset
r
es
u
lt.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
N
a
tu
r
a
l la
n
g
u
a
g
e
d
escr
ip
tio
n
o
f ima
g
es u
s
in
g
h
y
b
r
id
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Md
.
A
s
ifu
z
z
a
ma
n
Jish
a
n
)
2937
T
ab
le
1
.
B
L
E
U
s
co
r
es a
n
d
M
E
T
E
OR
s
co
r
e
f
o
r
Fli
ck
r
8
K
d
a
taset
D
a
t
a
se
t
M
o
d
e
l
B
L
EU
-
1
B
L
EU
-
2
B
L
EU
-
3
B
L
EU
-
4
M
ET
EO
R
M
a
o
e
t
a
l
.
[
3
7
]
58
28
23
-
-
G
o
o
g
l
e
N
I
C
[
2
]
63
41
27
-
-
F
l
i
c
k
r
8
K
L
R
C
N
[
3
8
]
-
-
-
-
-
M
S
R
e
se
a
r
c
h
[
3
9
]
-
-
-
-
-
C
h
e
n
a
n
d
Z
i
t
n
i
c
k
[
4
0
]
-
-
-
1
4
.
1
-
H
y
b
r
i
d
R
N
N
M
o
d
e
l
5
2
.
6
3
4
.
4
2
1
.
8
1
4
.
1
1
6
.
4
9
5
5
4
3
T
ab
le
2
.
B
L
E
U
s
co
r
es a
n
d
M
E
T
E
OR
s
co
r
e
f
o
r
Fli
ck
r
3
0
K
d
ataset
D
a
t
a
se
t
M
o
d
e
l
B
L
EU
-
1
B
L
EU
-
2
B
L
EU
-
3
B
L
EU
-
4
M
ET
EO
R
M
a
o
e
t
a
l
.
[
3
7
]
55
24
20
-
-
G
o
o
g
l
e
N
I
C
[
2
]
6
6
.
3
4
2
.
3
2
7
.
7
1
8
.
3
-
F
l
i
c
k
r
3
0
K
L
R
C
N
[
3
8
]
5
8
.
8
3
9
.
1
2
5
.
1
1
6
.
5
-
M
S
R
e
se
a
r
c
h
[
3
9
]
-
-
-
-
-
C
h
e
n
a
n
d
Z
i
t
n
i
c
k
[
4
0
]
-
-
-
1
2
.
6
-
H
y
b
r
i
d
R
N
N
M
o
d
e
l
5
6
.
8
3
7
.
3
2
4
.
1
1
5
.
6
1
9
.
4
4
1
4
5
2
T
ab
le
3
.
B
L
E
U
s
co
r
es a
n
d
M
E
T
E
OR
s
co
r
e
f
o
r
MSC
OC
O
d
ataset
D
a
t
a
se
t
M
o
d
e
l
B
L
EU
-
1
B
L
EU
-
2
B
L
EU
-
3
B
L
EU
-
4
M
ET
EO
R
M
a
o
e
t
a
l
.
[
3
7
]
-
-
-
-
-
G
o
o
g
l
e
N
I
C
[
2
]
6
6
.
6
4
6
.
1
3
2
.
9
2
4
.
6
-
M
S
C
O
C
O
L
R
C
N
[
3
8
]
6
2
.
8
4
4
.
2
3
0
.
4
-
-
M
S
R
e
se
a
r
c
h
[
3
9
]
-
-
-
2
1
.
1
2
0
.
7
C
h
e
n
a
n
d
Z
i
t
n
i
c
k
[
4
0
]
-
-
-
1
9
.
0
2
0
.
4
H
y
b
r
i
d
R
N
N
M
o
d
e
l
6
4
.
4
4
5
.
4
3
0
.
9
2
1
.
2
1
9
.
6
1
3
2
2
7
(
a)
(
b
)
Fig
u
r
e
3
.
Gr
ap
h
ical
r
ep
r
esen
ta
tio
n
o
f
tr
ain
in
g
ti
m
e
u
s
i
n
g
t
h
r
e
e
b
en
ch
m
ar
k
d
ataset
(
ep
o
ch
v
s
.
ac
cu
r
ac
y
a
n
d
ep
o
ch
v
s
.
lo
s
s
)
,
(
a)
E
p
o
ch
v
s
.
lo
s
s
,
a
n
d
E
p
o
ch
v
s
.
ac
cu
r
ac
y
f
o
r
Fli
c
k
r
8
k
,
(
b
)
E
p
o
ch
v
s
.
lo
s
s
,
a
n
d
E
p
o
ch
v
s
.
ac
cu
r
ac
y
f
o
r
Fli
c
k
r
3
0
k
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
3
2
-
2940
2938
(
c)
Fig
u
r
e
3
.
Gr
ap
h
ical
r
ep
r
es
en
ta
tio
n
o
f
tr
ain
in
g
ti
m
e
u
s
i
n
g
t
h
r
e
e
b
en
ch
m
ar
k
d
ataset
(
ep
o
ch
v
s
.
ac
cu
r
ac
y
a
n
d
ep
o
ch
v
s
.
lo
s
s
)
,
(
c)
E
p
o
ch
v
s
.
lo
s
s
,
a
n
d
E
p
o
ch
v
s
.
ac
cu
r
ac
y
f
o
r
MS
C
OC
O
We
tr
ain
o
u
r
h
y
b
r
id
d
ee
p
le
ar
n
in
g
m
o
d
el
th
at
e
n
s
u
r
es
le
ar
n
in
g
of
m
u
lt
i
m
o
d
al
asp
ec
ts
of
i
m
a
g
e
co
n
ten
t
s
alo
n
g
w
it
h
th
e
r
elate
d
n
atu
r
al
la
n
g
u
a
g
e
te
x
t.
W
e
r
e
p
r
esen
t
th
e
tr
ain
i
n
g
r
esu
lt
o
f
e
ac
h
d
ataset
u
s
i
n
g
i
n
d
if
f
er
e
n
t
g
r
ap
h
.
I
n
Fi
g
u
r
e
3
,
w
e
r
ep
r
ese
n
t
t
h
e
tr
ai
n
i
n
g
ac
c
u
r
ac
y
v
s
.
ep
o
ch
an
d
lo
s
s
v
s
.
ep
o
ch
in
d
if
f
er
en
t
g
r
ap
h
.
T
h
is
s
h
o
w
s
th
a
t o
u
r
m
o
d
el
d
o
es n
o
t e
n
co
u
n
ter
o
v
er
f
itt
in
g
p
r
o
b
le
m
.
4
.
1
.
Dis
cu
s
s
io
n
W
e
tr
ain
o
u
r
h
y
b
r
id
m
o
d
el
t
o
g
en
er
ate
s
e
n
te
n
ce
s
o
n
f
u
ll
i
m
a
g
es
w
h
er
e
w
e
u
s
e
C
N
N
to
lear
n
c
a
t
e
g
o
r
i
c
a
l
f
ea
tu
r
es
f
r
o
m
i
m
a
g
es
an
d
a
lan
g
u
ag
e
m
o
d
el
to
ass
is
t
t
h
e
m
ap
p
in
g
f
r
o
m
i
m
a
g
e
d
ata
to
s
eq
u
en
ce
s
o
f
w
o
r
d
s
,
f
o
llo
w
ed
b
y
a
B
R
NN
th
at
lear
n
s
th
e
w
o
r
d
r
ep
r
esen
t
atio
n
.
T
h
is
co
n
ca
ten
at
io
n
o
f
i
m
ag
e
clas
s
i
f
ier
a
n
d
lan
g
u
a
g
e
m
o
d
els
e
n
s
u
r
es
lear
n
in
g
o
f
m
u
lti
m
o
d
al
asp
ec
ts
o
f
i
m
ag
e
co
n
te
n
t
s
alo
n
g
w
it
h
th
e
r
elate
d
n
at
u
r
al
lan
g
u
a
g
e
te
x
t.
W
e
o
b
s
er
v
e
th
a
t
o
u
r
h
y
b
r
id
m
o
d
el
can
g
e
n
er
a
te
r
ea
s
o
n
ab
le
d
escr
ip
tio
n
s
of
i
m
ag
e
s
as
s
h
o
w
n
i
n
Fig
u
r
e
4
(
a)
ev
en
f
o
r
r
elativ
el
y
s
m
a
ll
or
r
ar
e
o
b
j
ec
ts
r
ef
er
Fi
g
u
r
e
4
(
b
)
w
h
ic
h
is
a
s
ig
n
i
f
ica
n
t
i
m
p
r
o
v
e
m
e
n
t
in
th
e
te
x
t
u
al
r
etr
ie
v
al
f
r
o
m
t
h
e
i
m
a
g
es.
Fo
r
th
e
lear
n
i
n
g
an
d
test
in
g
p
h
ase
of
o
u
r
m
o
d
el
we
h
a
v
e
u
s
ed
t
h
r
ee
b
en
ch
m
ar
k
v
is
u
al
d
atase
ts
f
o
r
n
atu
r
al
la
n
g
u
a
g
e
b
ased
d
escr
ip
tio
n
,
e.
g
.
,
Fli
c
k
r
8
K,
Fli
c
k
r
3
0
K
an
d
MS
C
OC
O
d
atasets
an
d
w
e
h
av
e
r
ep
o
r
ted
th
e
B
L
E
U
an
d
ME
T
E
OR
s
co
r
es
f
o
r
th
e
co
m
p
ar
is
o
n
.
C
o
m
p
ar
ed
to
th
e
o
t
h
er
s
tate
of
t
h
e
ar
t
m
o
d
el,
o
u
r
m
o
d
el
s
h
o
w
s
t
h
e
b
etter
p
er
f
o
r
m
an
c
e
or
co
m
p
ar
ab
le
to
t
h
e
m
,
as
our
m
o
d
el
f
in
e
-
t
u
n
es
th
e
ar
ch
itect
u
r
e
an
d
h
y
p
er
p
ar
am
eter
s
o
f
t
h
e
m
o
d
el,
r
es
u
lts
i
n
T
a
b
l
e
1
-
3.
(
a)
Fo
r
eac
h
test
p
ictu
r
e,
we
g
o
t
th
e
m
o
s
t
p
er
f
ec
t
test
s
e
n
ten
ce
(
b
)
W
e
g
o
t
th
e
ab
s
o
lu
te
b
est te
s
t se
n
ten
ce
f
o
r
test
i
m
ag
e
Fig
u
r
e
4.
E
x
a
m
p
le
of
s
e
n
te
n
ce
p
r
ed
icted
by
o
u
r
m
o
d
el.
Fo
r
ev
er
y
test
i
m
a
g
e,
we
g
o
t
th
e
m
o
s
t
co
m
p
atib
le
test
s
e
n
ten
c
e
We
ev
alu
ated
th
e
BL
EU
-
1,
2,
3,
4
s
co
r
es
an
d
ME
T
E
OR
s
co
r
es
an
d
co
m
p
ar
ed
o
u
r
r
esu
lt
s
w
it
h
th
e
b
en
ch
m
ar
k
r
es
u
lts
o
f
Ma
o
et
al
.
[
3
7
]
,
Go
o
g
le
NI
C
[
2
]
,
L
R
C
N
[
3
8
]
,
MS
R
e
s
ea
r
ch
[
3
9
]
,
an
d
C
h
en
an
d
Z
it
n
ic
k
[
4
0
]
m
o
d
el.
Fo
r
B
L
E
U
-
1
s
co
r
e,
it
is
o
b
s
er
v
ed
t
h
at
f
o
r
Fli
ck
r
3
0
k
o
u
r
m
o
d
el
g
i
v
es
b
etter
ac
cu
r
ac
y
th
a
n
th
e
Ma
o
et
al
.
m
o
d
el
[
3
7
]
an
d
f
o
r
M
SC
O
C
O
d
ataset
w
e
g
et
b
etter
r
es
u
lt
t
h
a
n
t
h
e
L
R
C
N
m
o
d
el
[
3
8
]
.
Seco
n
d
l
y
,
in
B
L
E
U
-
2
e
v
al
u
ati
o
n
,
o
u
r
m
o
d
el
g
iv
e
s
b
etter
r
es
u
lt
f
o
r
all
t
h
r
ee
b
en
c
h
m
ar
k
d
at
asets
.
Fo
r
B
L
E
U
-
3
ev
alu
a
tio
n
,
w
e
g
et
b
etter
p
er
f
o
r
m
a
n
ce
f
o
r
Fli
ck
r
3
0
k
a
n
d
MSC
O
C
O
w
h
ic
h
i
s
b
etter
th
an
th
e
Ma
o
el
al.
m
o
d
el
[
3
7
]
an
d
L
R
C
N
[
3
8
]
m
o
d
el
r
esp
ec
ti
v
el
y
.
Fo
r
B
L
E
U
-
4
s
co
r
e,
Fli
ck
r
3
0
k
a
n
d
MSC
O
C
O
g
iv
e
b
etter
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
N
a
tu
r
a
l la
n
g
u
a
g
e
d
escr
ip
tio
n
o
f ima
g
es u
s
in
g
h
y
b
r
id
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Md
.
A
s
ifu
z
z
a
ma
n
Jish
a
n
)
2939
p
er
f
o
r
m
a
n
ce
co
m
p
ar
ed
w
it
h
MS
R
esear
c
h
[
3
9
]
,
an
d
C
h
en
a
n
d
Z
it
n
ic
k
[
4
0
]
m
o
d
el.
Fin
al
l
y
,
w
e
u
s
e
ME
T
E
OR
ev
alu
atio
n
an
d
g
e
t
1
6
.
4
9
5
5
4
3
,
1
9
.
4
4
1
4
5
2
an
d
1
9
.
6
1
3
2
2
7
f
o
r
th
e
b
en
ch
m
ar
k
d
atasets
r
esp
ec
ti
v
el
y
an
d
o
b
s
er
v
e
i
m
p
r
o
v
e
m
e
n
ts
i
n
o
u
r
r
esu
lts
.
On
e
li
m
itatio
n
s
o
f
o
th
er
m
o
d
el
is
th
at
t
h
e
y
ar
e
u
n
ab
le
to
g
en
er
at
e
d
if
f
er
e
n
t
p
atter
n
o
f
s
e
n
te
n
ce
r
ea
lizatio
n
s
a
s
t
h
e
d
ataset
s
co
n
s
is
t
s
o
f
h
an
d
m
ad
e
an
n
o
tatio
n
s
,
b
u
t
o
u
r
m
o
d
el
ca
n
g
en
er
ate
d
y
n
a
m
ic
o
u
tp
u
t a
s
o
u
r
m
o
d
el
lear
n
s
to
m
o
d
u
late
t
h
e
m
ag
n
it
u
d
e
o
f
t
h
e
r
eg
io
n
a
n
d
w
o
r
d
e
m
b
ed
d
in
g
.
I
n
s
p
ite
o
f
th
e
f
ac
t
t
h
at
o
u
r
o
u
tco
m
e
s
ar
e
en
co
u
r
ag
i
n
g
,
th
e
m
o
d
el
o
f
Mu
l
ti
m
o
d
al
R
N
N
(
R
e
cu
r
r
en
t
Neu
r
al
Net
w
o
r
k
)
h
a
s
d
i
f
f
er
e
n
t
t
y
p
e
o
f
li
m
itat
io
n
s
.
First
o
f
a
l
l,
th
i
s
M
u
lti
m
o
d
al
R
NN
m
o
d
el
ca
n
o
n
l
y
g
en
er
ate
a
d
escr
ip
tio
n
o
r
s
e
n
te
n
ce
o
f
o
n
l
y
o
n
e
i
n
p
u
t
ar
r
a
y
an
d
th
a
t
ar
r
a
y
o
f
p
ix
e
ls
a
t
a
f
i
x
ed
r
eso
lu
tio
n
.
An
o
th
er
s
en
s
ib
le
ap
p
r
o
ac
h
is
to
u
s
e
m
u
lti
p
le
s
ac
ca
d
es
id
e
n
ti
f
y
t
h
e
a
ll
o
f
e
n
ti
ties
ar
o
u
n
d
t
h
e
i
m
a
g
e
an
d
t
h
eir
co
m
m
o
n
co
llab
o
r
atio
n
s
an
d
m
o
r
e
e
x
ten
s
iv
e
s
ett
in
g
b
e
f
o
r
e
p
r
o
d
u
cin
g
a
d
escr
ip
tio
n
.
A
l
s
o
,
th
e
R
NN
(
R
ec
u
r
r
en
t
Ne
u
r
al
Net
w
o
r
k
)
ca
n
r
ec
eiv
e
t
h
e
i
n
f
o
r
m
at
io
n
of
al
l
i
m
ag
e
s
o
n
l
y
th
r
o
u
g
h
ad
d
iti
v
e
b
ias
i
n
ter
ac
t
io
n
s
w
h
ic
h
ar
e
less
ex
p
r
ess
i
v
e
th
a
n
m
o
r
e
co
m
p
lic
ated
m
u
lt
ip
licati
v
e
in
ter
ac
tio
n
s
.
5.
CO
NCLU
SI
O
N
W
e
s
tu
d
y
i
n
t
h
is
p
ap
er
a
co
m
p
lex
h
y
b
r
id
n
e
u
r
al
n
et
w
o
r
k
m
o
d
el
w
h
ich
s
h
o
w
s
r
e
m
ar
k
ab
le
ab
ilit
y
to
g
en
er
ate
n
a
tu
r
al
lan
g
u
ag
e
b
as
ed
s
in
g
le
s
en
te
n
ce
d
es
cr
ip
tio
n
f
r
o
m
a
g
i
v
e
n
tes
t
i
m
ag
e.
T
h
e
m
o
d
el
id
en
ti
f
ie
s
t
h
e
i
m
a
g
e
r
eg
io
n
an
d
g
e
n
er
ates
n
atu
r
al
la
n
g
u
a
g
e
d
escr
ip
tio
n
of
i
m
ag
e
s
.
O
u
r
ap
p
r
o
ac
h
in
clu
d
es
a
lo
w
er
in
g
of
r
eso
lu
tio
n
i
m
ag
e
s
t
h
at
ad
j
u
s
t
ed
p
ar
ts
o
f
v
i
s
u
al
a
n
d
lan
g
u
ag
e
m
o
d
alitie
s
t
h
r
o
u
g
h
th
e
in
ter
p
la
y
o
f
d
ee
p
co
n
v
o
lu
tio
n
lear
n
in
g
m
o
d
el
w
it
h
it
s
e
f
f
i
c
i
e
n
t
L
ST
M
an
d
B
R
NN
co
u
n
ter
p
ar
ts
.
Mo
r
eo
v
er
,
w
e
o
b
tain
b
etter
p
er
f
o
r
m
a
n
ce
co
m
p
ar
ed
to
b
en
ch
m
ar
k
r
es
u
lt
s
b
y
ea
r
lier
atte
m
p
t
s
.
W
e
r
ep
o
r
t
p
er
f
o
r
m
an
ce
r
es
u
lts
w
it
h
ap
p
r
o
p
r
iate
r
e
p
r
esen
tatio
n
alo
n
g
w
it
h
co
m
p
le
m
e
n
ta
r
y
ill
u
s
t
r
atio
n
s
f
o
r
b
etter
u
n
d
er
s
ta
n
d
in
g
.
O
u
r
ex
p
lo
r
atio
n
o
f
th
e
m
o
d
el
i
n
f
er
s
th
at
b
ette
r
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
w
id
en
i
n
g
r
an
g
e
o
f
d
atasets
m
a
y
b
e
a
ch
iev
ed
v
ia
m
o
d
el
f
i
n
e
-
t
u
n
i
n
g
a
n
d
ar
ch
itect
u
r
al
a
u
g
m
e
n
tatio
n
.
RE
F
E
R
E
NC
E
S
[1
]
L
.
F
e
i
-
F
e
i,
e
t
a
l
.
,
“
W
h
a
t
d
o
w
e
p
e
r
c
e
iv
e
in
a
g
lan
c
e
o
f
a
re
a
l
-
w
o
rld
sc
e
n
e
?
’
J
o
u
rn
a
l
o
f
v
isi
o
n
,
v
o
l/
issu
e
:
7
(1
)
,
p
p
.
10
,
2
0
0
7
.
[2
]
O
.
V
i
n
y
a
ls,
e
t
a
l
.,
“
S
h
o
w
a
n
d
tell
:
A
n
e
u
ra
l
im
a
g
e
c
a
p
ti
o
n
g
e
n
e
ra
to
r
,
”
a
rXiv:
1
4
1
1
.
4
5
5
5
v
2
,
2
0
1
5
.
[3
]
O
.
V
in
y
a
ls,
e
t
a
l
.,
“
S
h
o
w
a
n
d
te
ll
:
L
e
ss
o
n
s
lea
rn
e
d
f
ro
m
th
e
2
0
1
5
m
s
c
o
c
o
ima
g
e
c
a
p
ti
o
n
in
g
c
h
a
ll
e
n
g
e
,
”
a
rXiv:
1
6
0
9
.
0
6
6
4
7
v
1
,
2
0
1
6
.
[4
]
S
.
V
e
n
u
g
o
p
a
lan
,
e
t
a
l
.,
“
Ca
p
ti
o
n
i
n
g
im
a
g
e
s
w
it
h
d
iv
e
rse
o
b
jec
ts,
”
a
rXiv:
1
6
0
6
.
0
7
7
7
0
v
3
,
2
0
1
7
.
[5
]
L.
J.
Li
a
n
d
L.
F
e
i
-
F
e
i
,
“
W
h
a
t,
w
h
e
re
a
n
d
w
h
o
?
Clas
sify
in
g
e
v
e
n
ts
by
s
c
e
n
e
a
n
d
o
b
jec
t
re
c
o
g
n
it
i
o
n
,
”
I
CCV
,
2
0
0
7
.
[6
]
L.
J.
L
i,
e
t
a
l
.,
“
T
o
w
a
r
d
s
to
tal
sc
e
n
e
u
n
-
d
e
rsta
n
d
i
n
g
:
Clas
sif
ica
ti
o
n
,
a
n
n
o
tati
o
n
a
n
d
se
g
-
m
e
n
tatio
n
in
a
n
a
u
t
o
m
a
ti
c
f
ra
m
e
w
o
rk
,”
Co
mp
u
ter
Vi
sio
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
I
E
E
E
C
o
n
f
e
re
n
c
e
o
n
,
p
p.
2
0
3
6
-
2
0
4
3
,
2
0
0
9
.
[7
]
S
.
F
i
d
ler,
e
t
a
l
.,
“
A
se
n
ten
c
e
is
w
o
rth
a
t
h
o
u
sa
n
d
p
ix
e
ls,
”
CVP
R
,
2
0
1
3
.
[8
]
A
.
G
u
p
ta
a
n
d
P
.
M
a
n
n
e
m
,
“
F
ro
m
im
a
g
e
a
n
n
o
tatio
n
t
o
im
a
g
e
d
e
sc
ri
p
ti
o
n
,
”
Ne
u
ra
l
i
n
f
o
rm
a
ti
o
n
p
r
o
c
e
ss
in
g
,
S
p
ri
n
g
e
r,
2
0
1
2
.
[9
]
G
.
Ku
lk
a
rn
i
,
e
t
a
l
.
,
“
Ba
b
y
talk
:
U
n
d
e
rsta
n
d
in
g
a
n
d
g
e
n
e
ra
ti
n
g
sim
p
le i
m
a
g
e
d
e
sc
rip
ti
o
n
s,
”
CVP
R
,
2
0
1
1
.
[1
0
]
P.
Ku
z
n
e
tso
v
a
,
e
t
a
l
.
,
“
Co
l
lec
ti
v
e
g
e
n
e
ra
ti
o
n
of
n
a
tu
ra
l
im
a
g
e
d
e
sc
rip
ti
o
n
s
,
”
ACL
,
2
0
1
2
.
[1
1
]
P.
Ku
z
n
e
tso
v
a
,
e
t
a
l
.
,
“
T
re
e
talk
:
Co
m
p
o
siti
o
n
a
n
d
c
o
m
p
re
ss
io
n
of
tree
s
f
o
r
i
m
a
g
e
d
e
sc
rip
ti
o
n
s,
”
T
ra
n
sa
c
ti
o
n
s
of
t
h
e
Asso
c
ia
ti
o
n
f
o
r
Co
mp
u
t
a
ti
o
n
a
l
L
i
n
g
u
isti
c
s,
v
o
l
/
i
s
s
u
e
:
2
(1
0
)
,
p
p
.
3
5
1
-
3
6
2
,
2
0
1
4
.
[1
2
]
A.
F
a
rh
a
d
i,
e
t
a
l
.,
“
Ev
e
r
y
p
ictu
re
tells a sto
ry
:
G
e
n
e
ra
ti
n
g
se
n
ten
c
e
s f
ro
m
i
m
a
g
e
s
,
”
ECCV
,
2
0
1
0
.
[1
3
]
S
.
Ba
i
a
n
d
S
.
A
n
,
“
A
S
u
rv
e
y
on
Au
to
m
a
ti
c
I
m
a
g
e
Ca
p
ti
o
n
G
e
n
e
ra
ti
o
n
,
”
Ne
u
ro
c
o
mp
u
ti
n
g
,
2
0
1
8
.
[1
4
]
R
.
Be
rn
a
rd
i,
e
t
a
l
.,
“
A
u
to
m
a
ti
c
D
e
sc
rip
ti
o
n
G
e
n
e
ra
ti
o
n
f
ro
m
I
m
a
g
e
s:
A
S
u
rv
e
y
o
f
M
o
d
e
ls,
Da
tas
e
ts,
a
n
d
Ev
a
lu
a
ti
o
n
M
e
a
su
re
s
,
”
J
o
u
rn
a
l
o
f
Arti
fi
c
i
a
l
I
n
telli
g
e
n
c
e
Res
e
a
rc
h
(
J
AIR
)
,
v
o
l.
55,
p
p
.
4
0
9
-
4
4
2
,
2
0
1
6
.
[1
5
]
A
.
Ku
m
a
r
a
n
d
S
.
G
o
e
l,
“
A
su
rv
e
y
o
f
e
v
o
lu
ti
o
n
o
f
i
m
a
g
e
c
a
p
ti
o
n
i
n
g
tec
h
n
iq
u
e
s
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Hy
b
rid
In
telli
g
e
n
t
S
y
ste
ms
Pre
p
rin
t
,
p
p
.
1
-
19
,
2
0
1
7
.
[1
6
]
J.
De
n
g
,
e
t
a
l
.
,
“
Im
a
g
e
n
e
t:
A
larg
e
-
sc
a
le h
iera
rc
h
ica
l
i
m
a
g
e
d
a
tab
a
se
,
”
CVP
R
,
2
0
0
9
.
[1
7
]
M
.
Ho
d
o
sh
,
e
t
a
l
.
,
“
F
ra
m
in
g
ima
g
e
d
e
sc
rip
ti
o
n
a
s
a
ra
n
k
in
g
tas
k
:
d
a
ta,
m
o
d
e
ls
a
n
d
e
v
a
lu
a
ti
o
n
m
e
tri
c
s
,
”
J
o
u
r
n
a
l
o
f
Arti
fi
c
ia
l
I
n
telli
g
e
n
c
e
Res
e
a
rc
h
,
2
0
1
3
.
[1
8
]
T.
Y
.
L
in
,
e
t
a
l
.,
“
M
icro
so
f
t
c
o
c
o
:
Co
m
-
m
o
n
o
b
jec
ts i
n
c
o
n
tex
t
,
”
a
r
Xi
v
p
re
p
rin
t
a
rXiv:
1
4
0
5
.
0
3
1
2
,
2
0
1
4
.
[1
9
]
Y.
L
e
Cu
n
,
e
t
a
l
.,
“
G
ra
d
ien
t
-
b
a
se
d
lea
rn
in
g
a
p
p
li
e
d
to
d
o
c
u
m
e
n
t
re
c
o
g
n
it
io
n
,
”
Pr
o
c
e
e
d
-
in
g
s
o
f
t
h
e
IEE
E,
v
o
l/
issu
e
:
8
6
(1
1
)
,
p
p
.
2
2
7
8
-
2
3
2
4
,
1
9
9
8
.
[2
0
]
A.
Kriz
h
e
v
sk
y
,
e
t
a
l
.
,
“
I
m
a
g
e
n
e
t
c
las
sif
i
c
a
ti
o
n
w
it
h
d
e
e
p
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
,
”
NIPS
,
2
0
1
2
.
[2
1
]
S.
Ho
c
h
re
it
e
r
a
n
d
J.
S
c
h
m
id
h
u
b
e
r,
“
L
o
n
g
sh
o
rt
-
term
m
e
m
o
r
y
,
”
Ne
u
ra
l
c
o
mp
u
ta
t
io
n
,
v
o
l
/
i
s
s
u
e
:
9
(8
)
,
p
p
.
1
7
3
5
-
1
7
8
0
,
1
9
9
7
.
[2
2
]
M.
S
c
h
u
ste
r
a
n
d
K.
K.
P
a
li
w
a
l,
“
Bid
irec
ti
o
n
a
l
re
c
u
rre
n
t
n
e
u
ra
l
n
e
tw
o
rk
s,
”
S
ig
n
a
l
Pro
c
e
ss
in
g
,
IE
EE
T
r
a
n
s
a
c
t
i
o
n
s
,
1
9
9
7
.
[2
3
]
P.
Y
o
u
n
g
,
e
t
a
l
.,
“
F
ro
m
i
m
a
g
e
d
e
sc
rip
ti
o
n
s
to
v
isu
a
l
d
e
n
o
tati
o
n
s:
Ne
w
s
im
i
-
larity
m
e
tri
c
s
f
o
r
se
m
a
n
ti
c
in
f
e
re
n
c
e
o
v
e
r
e
v
e
n
t
d
e
sc
rip
t
io
n
s,
”
T
A
C
L
,
2
0
1
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
3
2
-
2940
2940
[2
4
]
Y
.
Ki
m
,
“
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
N
e
tw
o
rk
s
f
o
r
S
e
n
ten
c
e
Clas
sif
ic
a
ti
o
n
,
”
a
rXiv:
1
4
0
8
.
5
8
8
2
v
2
,
2
0
1
4
.
[2
5
]
D
.
Cires
Ã
.,
e
t
a
l
.,
“
M
u
lt
i
-
c
o
lu
m
n
De
e
p
Ne
u
ra
l
Ne
tw
o
rk
s
f
o
r
Im
a
g
e
Clas
sif
ic
a
ti
o
n
,
”
a
rX
iv
:
1
2
0
2
.
2
7
4
5
v
1
,
2
0
1
2
.
[2
6
]
J
.
Ch
u
n
,
e
t
a
l
.,
“
Em
p
iri
c
a
l
Ev
a
lu
a
ti
o
n
of
G
a
ted
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
tw
o
rk
s
o
n
S
e
q
u
e
n
c
e
M
o
d
e
li
n
g
,
”
a
rXiv:
1
4
1
2
.
3
5
5
5
v
1
,
2
0
1
4
.
[2
7
]
Y
.
F
a
n
,
e
t
a
l
.,
“
TTS
S
y
n
th
e
sis
with
Bid
i
re
c
ti
o
n
a
l
L
S
T
M
b
a
se
d
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
tw
o
rk
s
,
”
Co
n
fer
e
n
c
e
o
f
th
e
In
ter
n
a
t
io
n
a
l
S
p
e
e
c
h
C
o
mm
u
n
ica
ti
o
n
Asso
c
ia
t
io
n
,
2
0
1
4
.
[2
8
]
J
.
S
o
n
g
,
e
t
a
l
.,
“
L
S
T
M
-
in
-
L
S
T
M
f
o
r
g
e
n
e
ra
ti
n
g
l
o
n
g
d
e
s
c
rip
ti
o
n
s
o
f
i
m
a
g
e
s,
”
Co
mp
u
ta
ti
o
n
a
l
V
isu
a
l
M
e
d
ia
,
2
0
1
6
.
[2
9
]
Z
.
C.
L
ip
to
n
,
e
t
a
l
.,
“
A
Crit
ica
l
Re
v
i
e
w
of
R
e
c
u
rre
n
t
Ne
u
ra
l
Ne
t
w
o
rk
s
f
o
r
S
e
q
u
e
n
c
e
L
e
a
rn
in
g
,
”
a
rXiv:
1
5
0
6
.
0
0
0
1
9
v
4
,
2
0
1
5
.
[3
0
]
J
.
Oh
,
e
t
a
l
.
,
“
A
c
ti
o
n
-
Co
n
d
it
i
o
n
a
l
V
i
d
e
o
P
r
e
d
icti
o
n
u
sin
g
De
e
p
N
e
tw
o
rk
s
in
A
tari
G
a
m
e
s
,
”
a
rXiv:
1
5
0
7
.
0
8
7
5
0
v
2
,
2
0
1
5
.
[3
1
]
A.
K
a
rp
a
th
y
,
e
t
a
l
.,
“
De
e
p
f
ra
g
m
e
n
t
e
m
b
e
d
d
in
g
s
f
o
r
b
id
irec
ti
o
n
a
l
im
a
g
e
se
n
ten
c
e
m
a
p
p
in
g
,
”
a
rXiv
p
re
p
rin
t
a
rX
iv:
1
4
0
6
.
5
6
7
9
,
2
0
1
4
.
[3
2
]
O.
Ru
ss
a
k
o
v
sk
y
,
e
t
a
l
.,
“
Im
a
g
e
n
e
t
larg
e
sc
a
le
v
isu
a
l
re
c
o
g
n
it
io
n
c
h
a
ll
e
n
g
e
,
”
a
rXiv:
1
4
0
9
.
0
5
7
5
v
3
,
2
0
1
5
.
[3
3
]
R.
G
irsh
ick
,
e
t
a
l
.,
“
Rich
f
e
a
tu
re
h
iera
rc
h
ies
f
o
r
a
c
c
u
ra
te o
b
jec
t
d
e
t
e
c
ti
o
n
a
n
d
se
m
a
n
ti
c
se
g
m
e
n
tatio
n
,
”
CVP
R
,
2
0
1
4
.
[3
4
]
T.
M
i
k
o
l
o
v
,
e
t
a
l
.,
“
Distri
b
u
ted
re
p
re
se
n
tatio
n
s
of
w
o
rd
s
a
n
d
p
h
ra
se
s
a
n
d
t
h
e
ir
c
o
m
p
o
siti
o
n
a
li
ty
,
”
NIPS
,
2
0
1
3
.
[3
5
]
W
.
Zare
m
b
a
,
e
t
a
l
.,
“
Re
c
u
rre
n
t
n
e
u
ra
l
n
e
tw
o
rk
re
g
u
lariz
a
ti
o
n
,
”
a
rXi
v
p
re
p
rin
t
a
rXiv:
1
4
0
9
.
2
3
2
9
,
2
0
1
4
.
[3
6
]
T.
T
iele
m
a
n
a
n
d
G.
E.
Hin
to
n
,
“
Lec
tu
re
6
.
5
-
rm
sp
ro
p
:
Div
id
e
t
h
e
g
ra
d
ien
t
by
a
ru
n
n
in
g
a
v
e
ra
g
e
of
its
re
c
e
n
t
m
a
g
n
it
u
d
e
,
”
2
0
1
2
.
[3
7
]
J.
M
a
o
,
e
t
a
l
.,
“
Ex
p
lain
ima
g
e
s
w
it
h
m
u
lt
im
o
d
a
l
re
c
u
rre
n
t
n
e
u
ra
l
n
e
t
w
o
rk
s
,
”
a
rXiv p
re
p
rin
t
a
rX
iv:
1
4
1
0
.
1
0
9
0
,
2
0
1
4
.
[3
8
]
J.
Do
n
a
h
u
e
,
e
t
a
l
.,
“
L
o
n
g
-
term
re
c
u
rre
n
t
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
tw
o
rk
s
f
o
r
v
is
u
a
l
r
e
c
o
g
n
it
io
n
a
n
d
d
e
sc
rip
ti
o
n
,
”
a
rXiv
p
re
p
rin
t
a
rXiv:
1
4
1
1
.
4
3
8
9
,
2
0
1
4
.
[3
9
]
H.
F
a
n
g
,
e
t
a
l
.
,
“
F
ro
m
c
a
p
ti
o
n
s t
o
v
isu
a
l
c
o
n
c
e
p
ts an
d
b
a
c
k
,
”
a
rXiv
p
re
p
rin
t
a
rXiv:
1
4
1
1
.
4
9
5
2
,
2
0
1
4
.
[4
0
]
X
.
C
h
e
n
a
n
d
C.
L
.
Zi
tn
ick
,
“
L
e
a
rn
in
g
a
re
c
u
rre
n
t
v
isu
a
l
re
p
re
s
e
n
tati
o
n
f
o
r
im
a
g
e
c
a
p
ti
o
n
g
e
n
e
ra
ti
o
n
,
”
C
o
RR
,
a
b
s/1
4
1
1
.
5
6
5
4
,
2
0
1
4
.
B
I
O
G
RAP
H
I
E
S
O
F
AUTH
O
RS
M
d
.
As
ifu
z
z
a
m
a
n
J
ish
a
n
is
stu
d
y
in
g
as
a
stu
d
e
n
t
of
Ba
c
h
e
lo
r
of
S
c
ien
c
e
in
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
w
it
h
in
t
h
e
De
p
a
rtme
n
t
of
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
at
th
e
Un
iv
e
rsity
of
L
ib
e
ra
l
A
rts
Ba
n
g
l
a
d
e
sh
(ULAB).
He
h
a
s
e
x
p
e
rti
se
in
C,
Ja
v
a
,
P
y
th
o
n
,
M
A
T
L
A
B
a
n
d
C+
+
p
ro
g
ra
m
m
in
g
lan
g
u
a
g
e
.
H
e
h
a
s
a
lso
w
o
r
k
in
g
k
n
o
w
led
g
e
in
d
iff
e
r
e
n
t
w
e
b
p
ro
g
ra
m
m
in
g
lan
g
u
a
g
e
:
HT
M
L
,
CS
S
,
Ja
v
a
S
c
rip
t
(JS),
L
a
ra
v
e
l
f
ra
m
e
w
o
r
k
,
a
n
d
d
a
tab
a
se
sy
ste
m
.
He
h
a
s
b
e
e
n
a
c
ti
v
e
in
t
h
e
re
se
a
rc
h
w
it
h
re
se
a
rc
h
in
tere
st
i
n
t
h
e
a
re
a
o
f
im
a
g
e
p
ro
c
e
ss
in
g
,
a
rti
f
icia
l
in
telli
g
e
n
c
e
,
m
a
c
h
in
e
lea
rn
in
g
a
n
d
n
e
u
r
a
l
s
y
s
t
e
m
.
K
h
a
n
R
a
q
i
b
M
a
h
m
u
d
c
u
rre
n
tl
y
w
o
rk
in
g
as
a
lec
tu
re
r
w
it
h
in
th
e
d
e
p
a
rtme
n
t
of
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
at
th
e
Un
iv
e
rsity
of
L
ib
e
ra
l
A
rts
B
a
n
g
lad
e
sh
(U
LA
B).
He
h
a
s
c
o
m
p
lete
d
Ba
c
h
e
lo
r
o
f
S
c
ien
c
e
(Ho
n
o
rs)
a
n
d
M
a
ste
r
o
f
S
c
ien
c
e
in
M
a
th
e
m
a
ti
c
s
f
ro
m
S
h
a
h
Ja
lal
Un
iv
e
rsit
y
o
f
S
c
ien
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
Ba
n
g
lad
e
sh
.
He
re
c
e
i
v
e
d
a
n
Eras
m
u
s
M
u
n
d
u
s
S
c
h
o
lars
h
i
p
f
ro
m
th
e
Ed
u
c
a
ti
o
n
,
A
u
d
io
v
isu
a
l
a
n
d
Cu
lt
u
re
Ex
e
c
u
ti
v
e
Ag
e
n
c
y
of
th
e
Eu
ro
p
e
a
n
Co
m
m
issio
n
,
to
p
u
rsu
e
a
d
o
u
b
le
M
a
ste
rs
in
S
c
ien
c
e
d
e
g
re
e
i
n
C
o
m
p
u
ter
S
im
u
latio
n
f
o
r
S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
a
n
d
Co
m
p
u
tatio
n
a
l
E
n
g
in
e
e
rin
g
,
f
ro
m
Ge
r
m
a
n
y
a
n
d
S
w
e
d
e
n
.
He
w
a
s
a
n
M
S
c
th
e
sis
stu
d
e
n
t
w
it
h
in
t
h
e
Co
m
p
u
tatio
n
a
l
T
e
c
h
n
o
lo
g
y
L
a
b
o
ra
to
ry
o
f
th
e
De
p
a
rt
m
e
n
t
o
f
Hi
g
h
P
e
rf
o
rm
a
n
c
e
Co
m
p
u
ti
n
g
a
n
d
V
isu
a
li
z
a
ti
o
n
a
t
KT
H
Ro
y
a
l
In
stit
u
te
o
f
T
e
c
h
n
o
l
o
g
y
,
S
w
e
d
e
n
.
His
re
se
a
rc
h
w
o
rk
c
o
n
c
e
n
trate
d
o
n
th
e
stu
d
y
o
f
th
e
se
n
siti
v
it
y
a
n
a
l
y
sis
o
f
Ne
a
r
W
a
l
l
T
u
rb
u
len
c
e
M
o
d
e
li
n
g
o
f
In
c
o
m
p
re
ss
ib
le
F
lo
w
s.
His
c
u
rre
n
t
re
s
e
a
rc
h
in
tere
st
in
c
l
u
d
e
s
m
a
c
h
in
e
lea
rn
in
g
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
im
a
g
e
p
ro
c
e
ss
in
g
a
n
d
c
o
m
p
u
ter v
isio
n
a
n
d
a
d
a
p
t
iv
e
d
y
n
a
m
ic
s
y
ste
m
.
Abu
l
K
a
la
m
al
A
z
a
d
re
c
e
i
v
e
d
h
is
P
h
D
in
A
p
p
li
e
d
M
a
t
h
e
m
a
ti
c
s
f
r
o
m
Un
iv
e
rsit
y
of
Ex
e
ter,
Un
it
e
d
Kin
g
d
o
m
,
M
a
ste
rs
of
S
c
ien
c
e
in
T
h
e
o
re
ti
c
a
l
P
h
y
sic
s
a
n
d
Ba
c
h
e
lo
r
of
S
c
ien
c
e
in
P
h
y
sic
s
f
ro
m
U
n
i
v
e
rsity
o
f
Dh
a
k
a
.
He
is
c
u
rre
n
tl
y
a
n
A
ss
o
c
iat
e
P
r
o
f
e
ss
o
r
a
t
th
e
De
p
a
rtm
e
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
,
Un
i
v
e
rsit
y
o
f
L
ib
e
ra
l
A
rts
Ba
n
g
l
a
d
e
sh
(ULA
B).
P
re
v
io
u
sly
,
h
e
u
n
d
e
rt
o
o
k
p
o
st
-
d
o
c
to
ra
l
re
se
a
rc
h
a
t
De
p
a
rt
m
e
n
t
o
f
Co
m
p
u
ti
n
g
a
n
d
M
a
t
h
e
m
a
ti
c
s,
Un
iv
e
rsit
y
o
f
P
ly
m
o
u
th
,
Un
it
e
d
Kin
g
d
o
m
,
a
n
d
S
c
h
o
o
l
of
Bio
lo
g
ica
l
S
c
ien
c
e
s,
Un
iv
e
rsity
of
Bristo
l,
Un
it
e
d
Kin
g
d
o
m
,
on
a
BBS
RC
f
e
ll
o
ws
h
ip
.
His
re
se
a
rc
h
in
tere
st
in
c
l
u
d
e
s
a
re
a
s
of
th
e
o
re
ti
c
a
l
a
n
d
c
o
m
p
u
tatio
n
a
l
n
e
u
ro
sc
ien
c
e
,
c
o
n
-
n
e
c
to
m
ics
,
m
u
lt
i
-
ti
m
e
s
c
a
le
d
y
n
a
m
ics
,
se
l
f
-
or
g
a
n
ize
d
c
rit
ica
li
t
y
(S
OC)
a
n
d
a
rti
f
icia
l
in
telli
g
e
n
c
e
.
He
h
a
s
p
u
b
li
sh
e
d
a
n
u
m
b
e
r
of
p
a
p
e
rs
in
p
e
e
r
-
re
v
ie
w
e
d
in
tern
a
ti
o
n
a
l
j
o
u
r
n
a
ls
a
n
d
p
re
se
n
ted
o
rig
in
a
l
re
se
a
rc
h
a
rti
c
les
in
n
u
m
e
ro
u
s
in
tern
a
ti
o
n
a
l
c
o
n
f
e
re
n
c
e
s.
He
re
c
e
i
v
e
d
v
a
rio
u
s
sc
h
o
lars
h
ip
s,
re
se
a
rc
h
a
n
d
trav
e
l
g
ra
n
ts
a
s
re
c
o
g
n
it
io
n
o
f
h
is
re
a
c
h
w
o
rk
.
He
w
a
s a
m
e
m
b
e
r
o
f
M
NN
,
OCN
S
a
n
d
S
IA
M
.
Evaluation Warning : The document was created with Spire.PDF for Python.