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.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
,
p
p
.
4
5
6
5
~
4
5
7
8
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
6
.
p
p
4
5
6
5
-
4
5
7
8
4565
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Em
o
V
i
be:
A
I
-
dr
i
v
e
n m
ul
ti
m
o
da
l
em
o
ti
o
n an
a
l
y
s
i
s
f
o
r m
ent
a
l
he
a
l
th
v
i
a
s
o
c
i
a
l
m
ed
i
a
d
a
s
hbo
a
rds
Dee
pa
li Vo
ra
,
Ary
a
n Sha
rma
,
M
ud
it
G
a
rg
,
Ste
v
e
F
r
a
ncis
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
a
n
d
En
g
i
n
e
e
r
i
n
g
,
S
y
m
b
i
o
si
s
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
P
u
n
e
,
S
y
mb
i
o
si
s I
n
t
e
r
n
a
t
i
o
n
a
l
(
D
e
e
me
d
U
n
i
v
e
r
si
t
y
)
,
P
u
n
e
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
r
1
2
,
2
0
2
5
R
ev
is
ed
Sep
8
,
2
0
2
5
Acc
ep
ted
Oct
1
8
,
2
0
2
5
M
o
n
i
to
ri
n
g
m
e
n
tal
h
e
a
lt
h
v
ia
so
c
ial
m
e
d
ia
o
ften
u
t
il
ize
s
u
n
imo
d
a
l
a
p
p
ro
a
c
h
e
s,
su
c
h
a
s
se
n
ti
m
e
n
t
a
n
a
ly
sis
o
n
tex
t
o
r
sin
g
le
-
sta
g
e
d
ima
g
e
c
a
teg
o
riza
ti
o
n
,
o
r
e
x
e
c
u
tes
e
a
rl
y
fe
a
tu
re
f
u
sio
n
.
Ho
we
v
e
r,
in
re
a
l
-
wo
rld
c
o
n
tex
ts
w
h
e
re
e
m
o
ti
o
n
s
a
re
c
o
n
v
e
y
e
d
v
ia
te
x
t,
e
m
o
ji
s,
a
n
d
ima
g
e
s,
u
n
imo
d
a
l
a
p
p
r
o
a
c
h
lea
d
s
to
o
b
sc
u
re
d
d
e
c
isio
n
-
m
a
k
in
g
p
a
t
h
wa
y
s
a
n
d
o
v
e
ra
ll
d
imin
ish
e
d
p
e
rfo
rm
a
n
c
e
.
T
o
o
v
e
rc
o
m
e
th
e
se
li
m
it
a
ti
o
n
s,
we
p
r
o
p
o
se
Emo
Vib
e
,
a
h
y
b
r
id
m
u
lt
imo
d
a
l
AI
fra
m
e
wo
rk
fo
r
e
m
o
ti
v
e
a
n
a
ly
sis.
Emo
Vib
e
u
se
s
a
tt
e
n
ti
o
n
-
b
a
se
d
la
te
fu
sio
n
stra
teg
y
,
w
h
e
re
tex
t
e
m
b
e
d
d
i
n
g
s
a
re
g
e
n
e
ra
ted
fro
m
b
i
d
irec
ti
o
n
a
l
e
n
c
o
d
e
r
re
p
re
se
n
tati
o
n
s
fro
m
tra
n
sfo
rm
e
rs
(
BERT
)
a
n
d
v
is
u
a
l
fe
a
tu
re
s
a
re
e
x
trac
ted
b
y
v
isio
n
tr
a
n
sfo
rm
e
r
.
S
u
b
se
q
u
e
n
tl
y
,
e
m
o
ti
c
o
n
v
e
c
to
rs
li
n
k
e
d
to
a
v
a
tars
a
re
p
ro
c
e
ss
e
d
in
d
e
p
e
n
d
e
n
t
ly
.
Late
r,
t
h
e
se
in
d
e
p
e
n
d
e
n
t
d
a
ta
fe
a
tu
re
s
a
re
i
n
teg
ra
te
d
a
t
h
ig
h
e
r
lev
e
ls,
e
n
h
a
n
c
in
g
i
n
terp
re
tab
il
it
y
a
n
d
p
e
rfo
rm
a
n
c
e
.
I
n
c
o
n
tras
t
to
e
a
rly
fu
sio
n
m
e
th
o
d
s
a
n
d
in
te
g
ra
ted
m
u
lt
imo
d
a
l
larg
e
lan
g
u
a
g
e
m
o
d
e
l
s
(
LL
M
s
)
li
k
e
CLIP
,
F
lam
in
g
o
,
G
P
T
-
4
V,
M
e
n
taLLaM
A,
a
n
d
d
o
m
a
in
-
a
d
a
p
ted
m
o
d
e
ls
li
k
e
Emo
BERT
a
,
Emo
Vi
b
e
p
re
se
rv
e
s
m
o
d
a
li
ty
-
s
p
e
c
ifi
c
c
o
n
tex
t
s
with
o
u
t
p
re
m
a
tu
re
fu
si
o
n
.
T
h
is
a
rc
h
i
tec
tu
re
sa
v
e
s
p
ro
c
e
ss
in
g
c
o
st
,
a
ll
o
win
g
f
o
r
c
lea
re
r,
u
n
a
m
b
ig
u
o
u
s
ra
ti
o
n
a
li
z
a
ti
o
n
a
n
d
e
x
p
la
n
a
ti
o
n
s.
Emo
Vib
e
o
u
t
p
e
rfo
rm
s
u
n
imo
d
a
l
b
a
se
li
n
e
s
a
n
d
e
a
rly
fu
sio
n
m
o
d
e
ls,
o
b
tain
i
n
g
8
9
.
7
%
a
c
c
u
ra
c
y
o
n
G
o
Emo
ti
o
n
s,
F
ER
,
a
n
d
Affe
c
tNe
t,
c
o
m
p
a
re
d
t
o
BER
T'
s
8
7
.
4
%
a
n
d
Re
sN
e
t
-
5
0
'
s
8
4
.
2
%
,
re
sp
e
c
ti
v
e
ly
.
F
u
rt
h
e
rm
o
re
,
a
c
u
sto
m
iza
b
le
,
re
a
l
-
ti
m
e
,
p
riv
a
c
y
-
a
wa
re
d
a
sh
b
o
a
r
d
i
s
c
re
a
ted
,
su
p
p
o
rt
in
g
p
h
y
sic
ian
s
a
n
d
e
n
d
u
se
rs.
Th
is
tec
h
n
o
lo
g
y
e
n
a
b
les
sc
a
lab
le
a
n
d
p
ro
a
c
ti
v
e
in
terv
e
n
ti
o
n
o
p
ti
o
n
s
a
n
d
fo
ste
rs
u
se
r
se
lf
-
a
wa
re
n
e
ss
o
f
m
e
n
tal
h
e
a
l
th
.
K
ey
w
o
r
d
s
:
Af
f
ec
tiv
e
s
tate
esti
m
atio
n
E
m
o
tio
n
d
ash
b
o
ar
d
Hy
b
r
id
tr
an
s
f
e
r
lear
n
in
g
L
ate
-
f
u
s
io
n
M
u
l
t
i
m
o
d
a
l
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
R
ea
l
-
tim
e
m
o
n
ito
r
in
g
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
:
Dee
p
ali
Vo
r
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
Sy
m
b
io
s
is
I
n
s
titu
te
o
f
T
ec
h
n
o
l
o
g
y
Pu
n
e
Sy
m
b
io
s
is
I
n
ter
n
atio
n
al
(
Dee
m
ed
Un
iv
er
s
ity
)
L
av
ale,
Mu
ls
h
i,
Pu
n
e,
Ma
h
ar
a
s
h
tr
a
4
1
2
1
1
5
,
I
n
d
ia
E
m
ail: d
ee
p
ali.
v
o
r
a@
s
itp
u
n
e.
ed
u
.
in
1.
I
NT
RO
D
UCT
I
O
N
Me
n
tal
h
ea
lth
is
a
s
ig
n
if
ican
t
co
m
p
o
n
en
t
o
f
f
ee
lin
g
g
o
o
d
h
ea
lth
-
wis
e.
I
t
in
f
lu
e
n
ce
s
em
o
tio
n
al,
p
s
y
ch
o
lo
g
ical,
an
d
s
o
cial
f
ac
to
r
s
th
at
d
ictate
h
o
w
in
d
iv
i
d
u
als
th
in
k
,
f
e
el,
an
d
in
ter
ac
t.
W
ith
to
d
ay
'
s
v
er
y
co
n
n
ec
ted
d
ig
ital
er
a,
s
o
cial
m
ed
ia
s
ites
ar
e
p
lace
s
wh
er
e
p
e
o
p
le
ca
n
ex
p
r
ess
th
em
s
elv
es
an
d
lear
n
[
1
]
.
W
h
ile
th
ese
s
ite
s
in
cite
d
i
s
cu
s
s
io
n
,
th
ey
also
cr
ea
te
a
d
if
f
er
en
t
k
in
d
o
f
s
o
cial
b
o
n
d
in
g
th
at
f
o
r
ce
s
u
s
to
co
m
p
ar
e
o
u
r
s
elv
es
to
o
th
er
s
.
C
y
b
er
b
u
lly
in
g
an
d
em
o
tio
n
al
d
is
tr
ess
ex
ac
er
b
ate
m
en
tal
h
ea
lth
is
s
u
es
s
u
ch
as
d
ep
r
ess
io
n
,
lo
n
elin
ess
,
an
d
a
n
x
iety
.
T
h
e
W
o
r
ld
Hea
lth
Or
g
a
n
izatio
n
es
tim
ates
th
at
alm
o
s
t
9
7
0
m
illi
o
n
p
eo
p
le
g
l
o
b
ally
a
r
e
im
p
ac
ted
b
y
m
e
n
tal
h
ea
lth
d
is
o
r
d
er
s
,
an
d
I
n
d
ia
alo
n
e
ac
c
o
u
n
ts
f
o
r
o
v
er
1
9
7
m
illi
o
n
c
ases
,
in
d
icatin
g
an
u
r
g
en
t
n
ee
d
f
o
r
s
ca
lab
le
an
d
a
cc
ess
ib
le
m
en
tal
h
ea
lth
m
o
n
it
o
r
in
g
s
o
lu
tio
n
s
[
2
]
,
[
3
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
5
6
5
-
4
5
7
8
4566
So
cial
m
ed
ia
s
ites
ar
e
ex
p
ec
ted
to
r
ea
ch
a
g
lo
b
al
u
s
er
b
ase
o
f
6
.
0
b
illi
o
n
b
y
2
0
2
8
.
B
ein
g
a
d
o
u
b
le
-
ed
g
ed
s
wo
r
d
,
th
ey
en
a
b
le
u
s
er
s
to
ex
p
r
ess
em
o
tio
n
s
an
d
g
ain
ap
p
r
o
v
al
to
en
h
an
ce
s
o
cial
co
n
n
ec
ted
n
ess
.
On
th
e
o
th
er
h
an
d
,
th
e
y
m
ag
n
if
y
th
e
ad
v
er
s
e
p
h
en
o
m
en
o
n
o
f
s
o
cial
co
m
p
ar
is
o
n
,
cy
b
er
b
u
lly
in
g
,
an
d
r
elian
ce
o
n
ex
ter
n
al
ap
p
r
o
v
al
b
y
lik
es
an
d
co
m
m
e
n
ts
.
T
h
ese
r
elatio
n
s
wo
r
s
en
p
s
y
ch
o
l
o
g
ical
d
is
o
r
d
e
r
s
,
s
u
ch
as
an
x
iety
,
d
ep
r
ess
io
n
,
a
n
d
lo
w
s
elf
-
estee
m
.
Acc
o
r
d
in
g
to
o
n
e
s
tu
d
y
,
m
an
y
in
d
iv
id
u
als
ex
p
er
ie
n
ce
m
en
tal
in
s
tab
ilit
y
an
d
s
elf
-
d
o
u
b
t b
ec
au
s
e
o
f
th
e
c
u
lt
u
r
e
o
f
s
o
cial
co
m
p
ar
is
o
n
wh
ic
h
is
en
co
u
r
a
g
ed
o
n
th
e
in
ter
n
et
[
4
]
.
T
h
ese
b
a
r
r
ier
s
ar
e
b
ec
o
m
in
g
m
o
r
e
an
d
m
o
r
e
well
-
k
n
o
wn
,
b
u
t
th
e
s
co
p
e
o
f
cu
r
r
e
n
tly
av
ailab
le
m
eth
o
d
s
o
f
m
en
tal
h
ea
lth
ev
alu
atio
n
r
e
m
ain
s
lo
w.
Un
i
m
o
d
al
in
f
o
r
m
atio
n
is
u
s
ed
in
m
o
s
t
s
tu
d
ies
an
d
th
is
ap
p
r
o
ac
h
f
ails
to
ap
p
r
ec
iate
th
e
in
tr
icac
y
o
f
ex
p
r
ess
io
n
o
f
em
o
tio
n
[
5
]
.
Sin
g
le
-
im
a
g
e
-
b
ased
em
o
tio
n
d
etec
tio
n
a
n
d
s
en
tim
en
t
an
aly
s
is
b
ased
o
n
tex
t
alo
n
e
ar
e
ju
s
t
two
in
s
tan
ce
s
o
f
th
is
.
Ad
d
itio
n
ally
,
s
in
ce
th
ese
m
o
d
els
ty
p
ically
u
tili
ze
r
etr
o
s
p
ec
tiv
e
ass
ess
m
en
ts
in
s
tead
o
f
r
e
al
-
tim
e
f
u
n
ctio
n
ality
,
th
ey
m
is
s
s
ig
n
if
ican
t
o
p
p
o
r
tu
n
ities
f
o
r
tim
ely
in
ter
v
en
tio
n
[
6
]
.
Hu
m
an
f
e
elin
g
s
ar
e
n
o
t
co
n
v
e
y
ed
in
a
s
in
g
le
s
en
ten
ce
d
u
e
to
th
eir
in
ter
n
al
co
m
p
lex
ity
.
C
u
r
r
en
t
m
o
d
els
ca
n
n
o
t
ef
f
icie
n
tly
lev
er
ag
e
t
h
e
elab
o
r
ate
m
u
ltimo
d
al
en
v
ir
o
n
m
e
n
t
g
en
e
r
a
ted
wh
en
a
p
ictu
r
e
o
r
em
o
ji d
escr
ib
es a
u
s
er
'
s
em
o
tio
n
al
r
esp
o
n
s
e
[
7
]
.
T
h
is
wo
r
k
s
u
g
g
ests
an
AI
m
u
ltimo
d
al
ap
p
r
o
ac
h
to
id
e
n
tif
y
in
g
em
o
tio
n
al
s
tates
f
r
o
m
s
o
cial
m
ed
ia
em
o
tico
n
s
,
im
ag
es,
an
d
tex
t.
I
t
ad
o
p
ts
h
y
b
r
id
tr
an
s
f
er
lea
r
n
in
g
to
m
atch
in
ter
-
m
o
d
al
f
ea
tu
r
es
an
d
u
tili
ze
s
s
tr
o
n
g
er
m
o
d
els
s
u
c
h
as
v
is
io
n
tr
a
n
s
f
o
r
m
e
r
s
f
o
r
im
a
g
es
an
d
b
id
ir
ec
tio
n
al
en
co
d
er
r
e
p
r
esen
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
er
s
(
B
E
R
T
)
f
o
r
tex
t.
T
h
is
ap
p
r
o
ac
h
ca
n
id
en
tify
c
o
m
p
lex
em
o
tio
n
al
s
tates,
in
cl
u
d
in
g
co
n
tr
ad
icto
r
y
em
o
tio
n
s
o
r
am
b
ig
u
o
u
s
ex
p
r
ess
io
n
s
,
f
r
eq
u
en
tly
o
v
er
lo
o
k
e
d
b
y
cu
r
r
en
t
m
eth
o
d
s
.
T
h
e
r
ea
l
-
tim
e
ca
p
ab
ilit
y
en
ab
les
p
r
o
ac
tiv
e
m
en
tal
h
ea
lt
h
s
cr
ee
n
in
g
an
d
tr
ea
tm
en
t,
with
a
p
er
s
o
n
alize
d
b
e
h
av
io
r
al
d
a
s
h
b
o
ar
d
m
a
k
in
g
it
u
s
ef
u
l
f
o
r
g
en
e
r
al
u
s
er
s
an
d
h
ea
lth
ca
r
e
p
r
o
v
id
er
s
.
T
h
e
m
o
d
el
f
o
llo
ws
eth
ical
s
tan
d
ar
d
s
a
n
d
in
clu
d
es
p
r
iv
ac
y
-
en
h
an
cin
g
ca
p
ab
ilit
ies f
o
r
m
a
n
ag
in
g
s
en
s
itiv
e
u
s
er
d
ata
s
ec
u
r
ely
.
T
h
e
lis
t o
f
k
ey
c
o
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
in
clu
d
es:
‒
E
m
o
Vib
e
is
a
m
u
ltimo
d
al
AI
f
r
am
ewo
r
k
co
m
b
in
in
g
tex
t,
i
m
ag
es,
an
d
em
o
tico
n
s
th
r
o
u
g
h
late
f
u
s
io
n
f
o
r
r
ea
l
-
tim
e
an
d
ac
c
u
r
ate
em
o
tio
n
r
ec
o
g
n
itio
n
.
‒
Hy
b
r
id
tr
an
s
f
er
lear
n
in
g
d
esig
n
,
in
teg
r
atin
g
p
r
e
-
tr
ain
ed
m
o
d
els
(
B
E
R
T
,
R
e
s
Net
-
5
0
)
w
ith
b
e
s
p
o
k
e
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
an
d
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
d
esig
n
s
to
m
ax
i
m
is
e
d
o
m
ain
-
s
en
s
itiv
e
em
o
tio
n
d
etec
tio
n
.
‒
An
in
n
o
v
ative
r
eal
-
ti
m
e
s
o
cial
m
ed
ia
em
o
ti
o
n
al
d
as
h
b
o
a
r
d
(
S
MED)
f
o
r
in
te
r
activ
e
m
o
n
it
o
r
in
g
o
f
em
o
tio
n
al
tr
en
d
s
an
d
tim
ely
i
n
ter
v
en
tio
n
.
‒
C
o
m
b
in
in
g
p
r
i
v
ac
y
-
p
r
eser
v
in
g
m
eth
o
d
s
an
d
eth
ical
d
ata
p
r
ac
tices
to
s
ec
u
r
ely
p
r
o
ce
s
s
s
e
n
s
itiv
e
m
en
tal
h
ea
lth
d
ata.
‒
L
ar
g
e
-
s
ca
le
b
e
n
ch
m
ar
k
in
g
o
n
Go
E
m
o
tio
n
s
,
FER,
a
n
d
Af
f
ec
tNet
d
atasets
with
b
etter
p
er
f
o
r
m
a
n
c
e
(
ac
cu
r
ac
y
o
f
u
p
to
8
9
.
7
%)
t
h
a
n
u
n
im
o
d
al
an
d
ea
r
ly
f
u
s
io
n
b
aselin
es.
T
h
e
s
u
g
g
ested
f
r
am
ewo
r
k
s
ig
n
if
ican
tly
ad
v
a
n
ce
s
co
m
p
r
e
h
e
n
d
in
g
an
d
en
h
a
n
cin
g
p
s
y
c
h
o
l
o
g
ical
well
-
b
ein
g
in
th
e
d
ig
ital
ag
e,
u
tili
zin
g
c
u
ttin
g
-
ed
g
e
m
ac
h
i
n
e
lear
n
i
n
g
m
o
d
els
an
d
m
u
ltimo
d
al
i
n
teg
r
a
tio
n
.
T
h
e
f
o
llo
win
g
wer
e
f
o
u
n
d
to
b
e
th
e
p
r
i
n
cip
al
r
esear
ch
q
u
esti
o
n
s
to
g
u
id
e
th
e
s
tu
d
y
:
i)
R
Q1
:
d
esig
n
a
m
u
lt
im
o
d
al
p
latf
o
r
m
to
in
teg
r
ate
an
d
ass
ess
em
o
tico
n
s
,
tex
t,
an
d
im
ag
es
o
n
s
o
cial
m
ed
ia
f
o
r
d
etec
tin
g
b
e
h
av
io
r
al
p
atter
n
s
s
u
ch
a
s
an
x
iety
,
an
g
er
,
a
n
d
d
ep
r
ess
io
n
;
ii)
R
Q2
:
h
y
b
r
id
tr
an
s
f
er
lear
n
in
g
is
a
p
p
lied
to
e
n
h
a
n
ce
ac
cu
r
ac
y
an
d
r
eliab
ilit
y
in
em
o
tio
n
d
etec
tio
n
in
n
u
m
er
o
u
s
m
o
d
alities
;
an
d
iii)
R
Q3
:
d
ev
elo
p
an
em
o
ti
o
n
al
d
ash
b
o
a
r
d
in
r
ea
l
-
tim
e
f
o
r
c
o
m
p
r
e
h
en
s
iv
e
v
is
u
aliza
tio
n
th
at
en
ab
les tim
ely
an
d
c
u
s
to
m
ized
m
en
tal
h
ea
lth
ca
r
e.
T
h
is
p
ap
er
is
d
iv
id
ed
in
to
s
e
v
en
s
ec
tio
n
s
.
Sectio
n
2
is
th
e
liter
atu
r
e
r
ev
iew
.
T
h
e
m
eth
o
d
o
lo
g
y
is
d
escr
ib
ed
in
s
ec
tio
n
3
.
T
h
e
im
p
lem
en
tatio
n
is
d
escr
ib
ed
i
n
s
ec
tio
n
4
.
S
ec
tio
n
5
p
r
esen
ts
th
e
f
in
d
in
g
s
an
d
co
m
m
en
ts
.
Sectio
n
6
p
r
esen
ts
th
e
m
ain
c
o
n
clu
s
io
n
s
an
d
im
p
licatio
n
s
.
Fin
ally
,
s
ec
tio
n
7
o
u
tlin
es
u
p
co
m
in
g
wo
r
k
,
an
d
f
o
llo
we
d
b
y
a
lis
t o
f
r
ef
er
e
n
ce
s
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
T
h
e
in
ter
s
ec
tio
n
o
f
AI
an
d
m
o
n
ito
r
in
g
o
f
m
en
tal
h
ea
lth
h
as
r
esu
lted
in
m
ass
iv
e
r
esear
ch
in
em
o
tio
n
r
ec
o
g
n
itio
n
u
s
in
g
d
ig
ital
d
ata
o
n
a
lar
g
e
s
ca
le.
E
ar
lier
r
esear
ch
u
s
ed
s
in
g
le
-
m
o
d
al
d
ata
,
s
u
ch
as
tex
t
an
d
im
ag
es,
to
id
en
tify
b
eh
a
v
io
r
al
tr
aits
,
b
u
t
it
f
ailed
to
ca
p
tu
r
e
e
m
o
tio
n
al
ex
p
r
ess
io
n
s
o
n
s
o
cia
l
m
ed
ia.
Ad
v
a
n
ce
d
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P
)
an
d
co
m
p
u
ter
v
is
io
n
m
o
d
els
allo
wed
th
e
d
e
v
elo
p
m
en
t
o
f
m
u
ltimo
d
al
s
y
s
tem
s
,
r
esu
ltin
g
in
m
o
r
e
c
o
n
tex
t
-
b
ased
m
en
tal
h
ea
lth
m
o
n
i
to
r
in
g
.
Hy
b
r
id
tr
an
s
f
er
lear
n
i
n
g
an
d
d
ata
f
u
s
io
n
m
eth
o
d
s
ad
d
r
ess
m
o
d
ality
-
s
p
ec
if
ic
s
h
o
r
tco
m
in
g
s
.
L
ate
f
u
s
io
n
ar
c
h
itectu
r
es,
f
o
r
i
n
s
tan
ce
,
en
h
an
ce
class
if
icatio
n
ac
cu
r
ac
y
b
y
alig
n
in
g
in
f
o
r
m
ati
o
n
in
tex
t,
im
ag
es,
an
d
em
o
tico
n
s
.
I
s
s
u
es
p
er
s
is
t,
h
o
wev
er
,
s
u
ch
as
co
m
p
u
tin
g
r
eq
u
ir
em
e
n
ts
,
p
r
iv
ac
y
,
a
n
d
r
ea
l
-
tim
e
an
aly
s
is
ca
p
ab
ilit
y
.
T
ab
le
1
ca
p
tu
r
es
in
f
lu
en
tial
f
in
d
in
g
s
,
m
eth
o
d
o
lo
g
ies,
an
d
co
n
tr
ib
u
tio
n
s
in
p
r
io
r
wo
r
k
,
estab
lis
h
in
g
th
e
g
o
als an
d
s
co
p
e
o
f
th
is
r
esear
ch
.
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
E
mo
V
ib
e:
A
I
-
d
r
iven
mu
ltimo
d
a
l e
mo
tio
n
a
n
a
lysi
s
fo
r
men
ta
l h
ea
lth
…
(
Dee
p
a
li V
o
r
a
)
4567
T
ab
le
1
.
Su
m
m
a
r
y
o
f
k
e
y
s
tu
d
ies o
n
m
u
ltimo
d
al
em
o
tio
n
d
et
ec
tio
n
Ref
M
e
t
h
o
d
o
l
o
g
y
D
a
t
a
s
e
t
u
s
e
d
P
e
r
f
o
r
ma
n
c
e
Li
mi
t
a
t
i
o
n
s
M
o
d
a
l
i
t
y
u
se
d
N
o
so
l
o
g
y
f
o
c
u
se
d
[
8
]
B
ER
T
-
b
a
se
d
t
e
x
t
su
mm
a
r
i
z
a
t
i
o
n
w
i
t
h
d
e
p
r
e
ss
i
o
n
d
e
t
e
c
t
i
o
n
D
A
I
C
-
W
O
Z
F1
-
sc
o
r
e
:
0
.
8
1
(
v
a
l
i
d
a
t
i
o
n
set
)
To
k
e
n
l
e
n
g
t
h
l
i
m
i
t
a
t
i
o
n
o
f
B
ER
T
m
o
d
e
l
s
Te
x
t
D
e
p
r
e
ss
i
o
n
[
9
]
H
y
b
r
i
d
d
e
e
p
l
e
a
r
n
i
n
g
mo
d
e
l
c
o
m
b
i
n
i
n
g
F
a
st
T
e
x
t
,
C
N
N
,
a
n
d
LSTM
f
o
r
d
e
p
r
e
ss
i
o
n
S
o
c
i
a
l
me
d
i
a
(
Tw
i
t
t
e
r
,
R
e
d
d
i
t
)
I
mp
r
o
v
e
d
a
c
c
u
r
a
c
y
o
v
e
r
st
a
t
e
-
of
-
t
h
e
-
a
r
t
mo
d
e
l
s
Li
mi
t
e
d
t
o
t
e
x
t
-
b
a
s
e
d
d
a
t
a
;
r
e
q
u
i
r
e
s
f
e
a
t
u
r
e
e
n
g
i
n
e
e
r
i
n
g
Te
x
t
D
e
p
r
e
ss
i
o
n
[
1
0
]
S
i
g
n
a
l
-
i
ma
g
e
e
n
c
o
d
i
n
g
w
i
t
h
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
e
mo
t
i
o
n
a
l
st
a
t
e
r
e
c
o
g
n
i
t
i
o
n
R
e
a
l
-
w
o
r
l
d
sen
s
o
r
d
a
t
a
9
8
.
5
%
a
c
c
u
r
a
c
y
f
o
r
e
m
o
t
i
o
n
c
l
a
ss
i
f
i
c
a
t
i
o
n
S
mal
l
t
r
a
i
n
i
n
g
d
a
t
a
se
t
,
l
i
mi
t
e
d
t
o
se
n
so
r
d
a
t
a
P
h
y
s
i
o
l
o
g
i
c
a
l
sen
s
o
r
s,
i
ma
g
e
M
e
n
t
a
l
w
e
l
l
-
b
e
i
n
g
,
e
m
o
t
i
o
n
a
l
s
t
a
t
e
s
[
1
1
]
D
e
e
p
l
e
a
r
n
i
n
g
w
i
t
h
c
o
n
t
e
x
t
u
a
l
e
m
o
t
i
o
n
d
e
t
e
c
t
i
o
n
u
si
n
g
i
m
a
g
e
d
a
t
a
EM
O
TI
C
,
M
S
C
O
C
O
,
A
D
E2
0
K
mA
P
:
7
9
.
6
%
O
n
l
y
c
o
n
si
d
e
r
s
i
ma
g
e
d
a
t
a
,
r
e
q
u
i
r
e
s
c
o
n
t
e
x
t
u
a
l
u
n
d
e
r
s
t
a
n
d
i
n
g
I
mag
e
(
b
o
d
y
l
a
n
g
u
a
g
e
,
c
o
n
t
e
x
t
)
Emo
t
i
o
n
d
e
t
e
c
t
i
o
n
(
c
o
n
t
e
x
t
u
a
l
)
[
1
2
]
F
u
n
c
t
i
o
n
a
l
n
e
t
w
o
r
k
c
o
n
n
e
c
t
i
v
i
t
y
w
i
t
h
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
me
n
t
a
l
h
e
a
l
t
h
rs
-
f
M
R
I
d
a
t
a
(
b
r
a
i
n
i
ma
g
i
n
g
)
I
mp
r
o
v
e
d
a
c
c
u
r
a
c
y
La
c
k
o
f
i
n
t
e
r
p
r
e
t
a
b
i
l
i
t
y
i
n
d
e
e
p
l
e
a
r
n
i
n
g
mo
d
e
l
s
N
e
u
r
o
i
ma
g
i
n
g
(rs
-
f
M
R
I
)
M
e
n
t
a
l
h
e
a
l
t
h
(
d
e
p
r
e
ssi
o
n
,
st
r
e
ss,
a
n
x
i
e
t
y
)
[
1
3
]
Ti
me
-
e
n
r
i
c
h
e
d
m
u
l
t
i
m
o
d
a
l
t
r
a
n
sf
o
r
mer f
o
r
d
e
p
r
e
ssi
o
n
d
e
t
e
c
t
i
o
n
Tw
i
t
t
e
r
,
R
e
d
d
i
t
,
mu
l
t
i
m
o
d
a
l
d
a
t
a
se
t
s
F1
-
sc
o
r
e
:
0
.
9
3
1
(
Tw
i
t
t
e
r
)
,
0
.
9
0
2
(
R
e
d
d
i
t
)
R
e
q
u
i
r
e
s
p
r
e
c
i
s
e
t
i
m
e
i
n
f
o
r
m
a
t
i
o
n
b
e
t
w
e
e
n
p
o
st
s f
o
r
o
p
t
i
m
a
l
r
e
s
u
l
t
s
Te
x
t
,
i
m
a
g
e
(
Emo
B
ER
T
a
,
C
LI
P
e
mb
e
d
d
i
n
g
s)
D
e
p
r
e
ss
i
o
n
[
1
4
]
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
f
o
r
mu
l
t
i
m
o
d
a
l
me
n
t
a
l
h
e
a
l
t
h
d
e
t
e
c
t
i
o
n
(
p
a
ssi
v
e
se
n
si
n
g
)
S
o
c
i
a
l
me
d
i
a
,
smar
t
p
h
o
n
e
s,
w
e
a
r
a
b
l
e
d
e
v
i
c
e
s,
a
u
d
i
o
,
a
n
d
v
i
d
e
o
V
a
r
i
e
s
b
y
a
p
p
r
o
a
c
h
,
g
e
n
e
r
a
l
l
y
i
m
p
r
o
v
e
s
w
i
t
h
f
u
s
i
o
n
R
e
q
u
i
r
e
s
c
a
r
e
f
u
l
f
u
si
o
n
o
f
f
e
a
t
u
r
e
s
f
r
o
m h
e
t
e
r
o
g
e
n
e
o
u
s
d
a
t
a
Te
x
t
,
a
u
d
i
o
,
v
i
d
e
o
,
w
e
a
r
a
b
l
e
s
M
u
l
t
i
p
l
e
men
t
a
l
h
e
a
l
t
h
d
i
s
o
r
d
e
r
s
(
d
e
p
r
e
ssi
o
n
,
a
n
d
a
n
x
i
e
t
y
)
[
1
5
]
M
o
b
i
l
e
-
b
a
s
e
d
a
p
p
l
i
c
a
t
i
o
n
f
o
r
p
r
e
v
e
n
t
i
n
g
a
n
d
t
r
e
a
t
i
n
g
men
t
a
l
h
e
a
l
t
h
i
ss
u
e
s
M
o
b
i
l
e
a
p
p
l
i
c
a
t
i
o
n
d
a
t
a
se
t
N
o
s
p
e
c
i
f
i
c
p
e
r
f
o
r
m
a
n
c
e
met
r
i
c
s wer
e
p
r
o
v
i
d
e
d
Li
mi
t
e
d
t
o
m
o
b
i
l
e
a
p
p
f
u
n
c
t
i
o
n
a
l
i
t
y
a
n
d
a
c
c
e
ss
t
o
h
e
a
l
t
h
d
e
p
a
r
t
me
n
t
s
M
o
b
i
l
e
a
p
p
,
b
e
h
a
v
i
o
r
a
l
t
r
a
c
k
i
n
g
M
e
n
t
a
l
h
e
a
l
t
h
d
i
s
o
r
d
e
r
s i
n
a
d
o
l
e
sce
n
t
s
[
1
6
]
B
ER
T
-
b
a
se
d
c
l
a
ss
i
f
i
c
a
t
i
o
n
f
o
r
p
h
o
b
i
a
su
b
t
y
p
e
s
i
n
t
w
e
e
t
s
N
o
v
e
l
t
w
e
e
t
d
a
t
a
se
t
(
8
1
1
,
5
6
9
t
w
e
e
t
s)
F1
-
sc
o
r
e
:
7
8
.
4
4
%
(
b
i
n
a
r
y
)
,
2
4
.
0
1
%
(
mu
l
t
i
-
c
l
a
ss)
Li
mi
t
e
d
t
o
t
e
x
t
d
a
t
a
;
n
o
t
a
p
p
l
i
c
a
b
l
e
t
o
a
l
l
p
h
o
b
i
a
s
u
b
t
y
p
e
s
Te
x
t
P
h
o
b
i
a
,
a
n
x
i
e
t
y
[
1
7
]
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
f
o
r
men
t
a
l
h
e
a
l
t
h
d
e
t
e
c
t
i
o
n
u
si
n
g
p
a
ss
i
v
e
se
n
s
i
n
g
M
u
l
t
i
p
l
e
p
a
ss
i
v
e
s
e
n
s
i
n
g
d
a
t
a
se
t
s
V
a
r
i
e
s w
i
t
h
t
h
e
met
h
o
d
a
p
p
l
i
e
d
P
r
i
v
a
c
y
c
o
n
c
e
r
n
s:
r
e
q
u
i
r
e
s
f
u
s
i
o
n
o
f
mu
l
t
i
-
s
o
u
r
c
e
d
a
t
a
Te
x
t
,
i
m
a
g
e
,
a
u
d
i
o
,
w
e
a
r
a
b
l
e
s,
v
i
d
e
o
V
a
r
i
o
u
s m
e
n
t
a
l
h
e
a
l
t
h
d
i
so
r
d
e
r
s
(
g
e
n
e
r
a
l
)
[
1
8
]
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
f
o
r
e
mo
t
i
o
n
d
e
t
e
c
t
i
o
n
a
n
d
sen
t
i
m
e
n
t
a
n
a
l
y
si
s
1
2
3
p
a
p
e
r
s
r
e
v
i
e
w
e
d
V
a
r
i
e
s
b
y
m
e
t
h
o
d
Li
mi
t
e
d
t
o
sen
t
i
m
e
n
t
a
n
a
l
y
si
s
;
d
a
t
a
a
n
d
a
p
p
l
i
c
a
t
i
o
n
d
o
m
a
i
n
-
f
o
c
u
s
e
d
Te
x
t
Emo
t
i
o
n
d
e
t
e
c
t
i
o
n
,
sen
t
i
m
e
n
t
a
n
a
l
y
si
s
[
1
9
]
M
e
n
t
a
l
h
e
a
l
t
h
a
n
a
l
y
si
s
i
n
so
c
i
a
l
me
d
i
a
p
o
st
s (s
u
r
v
e
y
)
S
o
c
i
a
l
me
d
i
a
(
Tw
i
t
t
e
r
,
R
e
d
d
i
t
,
S
i
n
a
W
e
i
b
o
)
V
a
r
i
e
s
b
y
m
e
t
h
o
d
a
n
d
d
a
t
a
se
t
La
c
k
s s
t
a
n
d
a
r
d
i
z
e
d
met
r
i
c
s
a
c
r
o
ss
st
u
d
i
e
s;
s
i
g
n
i
f
i
c
a
n
t
v
a
r
i
a
t
i
o
n
s
i
n
a
p
p
r
o
a
c
h
e
s.
Te
x
t
D
e
p
r
e
ss
i
o
n
,
st
r
e
ss,
su
i
c
i
d
e
r
i
sk
[
2
0
]
I
n
t
e
r
p
r
e
t
a
b
l
e
me
n
t
a
l
h
e
a
l
t
h
a
n
a
l
y
si
s
u
s
i
n
g
l
a
r
g
e
l
a
n
g
u
a
g
e
mo
d
e
l
s (L
LM
s)
S
o
c
i
a
l
me
d
i
a
(
Tw
i
t
t
e
r
,
R
e
d
d
i
t
)
S
t
a
t
e
-
of
-
t
h
e
-
a
r
t
i
n
t
e
r
p
r
e
t
a
b
i
l
i
t
y
a
n
d
a
c
c
u
r
a
c
y
R
e
q
u
i
r
e
s
d
o
ma
i
n
-
sp
e
c
i
f
i
c
f
i
n
e
-
t
u
n
i
n
g
;
l
i
m
i
t
e
d
o
p
e
n
-
so
u
r
c
e
d
a
t
a
se
t
s
Te
x
t
,
i
m
a
g
e
(
v
i
a
LL
M
)
M
e
n
t
a
l
h
e
a
l
t
h
(
g
e
n
e
r
a
l
)
[
2
1
]
M
u
l
t
i
m
o
d
a
l
l
e
a
r
n
i
n
g
w
i
t
h
t
r
a
n
sf
o
r
mers
f
o
r
m
e
n
t
a
l
h
e
a
l
t
h
a
n
a
l
y
s
i
s
S
o
c
i
a
l
me
d
i
a
a
n
d
m
u
l
t
i
m
o
d
a
l
d
a
t
a
se
t
s
V
a
r
i
e
s
b
y
a
p
p
l
i
c
a
t
i
o
n
N
e
e
d
s
e
x
t
e
n
s
i
v
e
d
a
t
a
a
n
d
m
u
l
t
i
p
l
e
mo
d
a
l
i
t
i
e
s
;
c
h
a
l
l
e
n
g
e
s wi
t
h
i
n
t
e
r
-
mo
d
a
l
i
t
y
Te
x
t
,
a
u
d
i
o
,
i
ma
g
e
G
e
n
e
r
a
l
men
t
a
l
h
e
a
l
t
h
a
n
a
l
y
si
s
[
2
2
]
M
u
l
t
i
m
o
d
a
l
a
n
a
l
y
s
i
s fo
r
d
e
p
r
e
ss
i
o
n
d
e
t
e
c
t
i
o
n
i
n
so
c
i
a
l
me
d
i
a
Tw
i
t
t
e
r
(
8
,
7
7
0
a
n
n
o
t
a
t
e
d
u
sers)
I
mp
r
o
v
e
d
F
1
-
sc
o
r
e
o
v
e
r
u
n
i
mo
d
a
l
a
p
p
r
o
a
c
h
e
s
C
h
a
l
l
e
n
g
e
s wi
t
h
r
e
a
l
-
t
i
me
d
e
p
l
o
y
me
n
t
;
n
e
e
d
f
o
r
e
t
h
i
c
a
l
d
a
t
a
u
sa
g
e
Te
x
t
,
i
m
a
g
e
,
v
i
d
eo
D
e
p
r
e
ss
i
o
n
,
su
i
c
i
d
a
l
t
e
n
d
e
n
c
i
e
s
R
ec
en
t
m
u
ltimo
d
al
L
L
M
f
r
a
m
ewo
r
k
s
h
av
e
g
r
ea
tly
im
p
r
o
v
ed
v
is
io
n
-
lan
g
u
ag
e
co
m
p
r
eh
e
n
s
io
n
.
C
L
I
P
[
2
3
]
co
n
s
tr
u
cts
jo
in
t
em
b
ed
d
i
n
g
s
v
ia
co
n
tr
asti
v
e
p
r
etr
ain
in
g
;
Flam
in
g
o
[
2
4
]
in
tr
o
d
u
ce
s
v
is
u
al
co
n
tex
ts
in
to
L
L
Ms
th
r
o
u
g
h
g
ate
d
cr
o
s
s
-
at
ten
tio
n
;
GPT
-
4V
[
2
5
]
p
r
o
v
id
es
a
u
n
if
ied
ap
p
r
o
ac
h
f
o
r
m
u
ltimo
d
al
r
ea
s
o
n
in
g
;
an
d
Me
n
taL
L
aM
A
[
1
9
]
f
o
cu
s
es
o
n
in
ter
p
r
etab
ilit
y
f
o
r
m
en
t
al
h
ea
lth
ev
alu
atio
n
.
W
h
ile
th
ese
ap
p
r
o
ac
h
es
ar
e
v
alu
ab
le,
m
o
s
t
m
o
d
els
u
tili
ze
ea
r
ly
,
o
r
to
k
en
-
le
v
el
f
u
s
io
n
,
wh
ich
ten
d
s
to
d
im
in
is
h
m
o
d
ality
-
s
p
ec
if
ic
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
5
6
5
-
4
5
7
8
4568
attr
ib
u
tes an
d
s
u
f
f
er
f
r
o
m
h
ig
h
co
m
p
u
tatio
n
al
co
s
ts
.
I
n
co
n
tr
ast,
E
m
o
Vib
e
u
tili
ze
s
atten
tio
n
-
g
u
id
e
d
late
f
u
s
io
n
to
m
ain
tain
in
d
ep
en
d
en
t
h
ig
h
-
lev
el
em
b
ed
d
in
g
s
,
alig
n
th
em
,
an
d
n
eu
tr
alize
co
n
f
licts
b
ef
o
r
e
m
er
g
i
n
g
.
T
h
is
also
ap
p
lies
to
em
o
tico
n
r
ep
r
esen
tatio
n
s
,
wh
ich
en
a
b
les
b
e
tter
in
ter
p
r
etab
ilit
y
,
r
ed
u
ce
d
r
eso
u
r
ce
u
s
ag
e,
an
d
r
ea
l
-
tim
e
r
esp
o
n
s
iv
en
ess
f
o
r
clin
ician
-
f
ac
in
g
d
ash
b
o
a
r
d
s
.
C
u
r
r
en
t
r
esear
ch
h
ig
h
lig
h
ts
s
ev
er
al
ess
en
tial
ch
allen
g
es
in
m
u
ltimo
d
al
AI
s
y
s
tem
s
f
o
r
m
en
tal
h
ea
lth
.
Un
im
o
d
al
ap
p
r
o
ac
h
es f
ail
to
c
o
n
s
id
er
th
e
co
m
p
lex
ity
o
f
em
o
tio
n
al
ex
p
r
ess
io
n
s
co
n
v
e
y
ed
t
h
r
o
u
g
h
s
ev
er
al
d
ata
ty
p
es
lead
in
g
to
im
p
r
ec
is
e
o
r
lack
in
g
ju
d
g
m
e
n
ts
.
Fu
r
th
er
m
o
r
e,
p
r
iv
ac
y
an
d
eth
ical
co
n
ce
r
n
s
m
ak
e
d
ata
g
ath
er
in
g
a
n
d
an
al
y
s
is
m
u
ch
m
o
r
e
co
m
p
lex
,
an
d
th
e
lack
o
f
r
ea
l
-
tim
e
m
o
n
ito
r
in
g
h
as
m
ad
e
it
m
o
r
e
d
if
f
icu
lt
to
ac
t
q
u
ick
ly
.
Fu
r
th
er
m
o
r
e,
a
lth
o
u
g
h
ex
is
tin
g
m
u
ltimo
d
al
m
o
d
els
in
[
1
4
]
,
[
1
5
]
,
[
1
7
]
,
[
1
9
]
ex
h
ib
it
p
o
ten
tial,
th
ey
h
a
v
e
d
if
f
icu
lty
s
tr
ik
in
g
a
c
o
m
p
r
o
m
is
e
b
etwe
en
co
m
p
u
tin
g
ec
o
n
o
m
y
an
d
ac
c
u
r
ac
y
.
Sin
ce
m
o
d
els
f
r
eq
u
e
n
tly
en
co
u
n
ter
co
n
s
tr
ai
n
ts
b
ec
au
s
e
o
f
th
e
s
ca
r
city
o
f
h
ig
h
-
q
u
ality
,
o
p
e
n
-
s
o
u
r
ce
tr
ai
n
in
g
d
ata
as
in
[
1
6
]
,
[
1
8
]
,
[
2
0
]
t
h
er
e
is
a
n
o
tab
l
e
v
o
id
in
in
teg
r
atin
g
lar
g
e
-
s
ca
le,
m
u
ltimo
d
al
d
atasets
.
Fu
r
th
er
m
o
r
e
,
m
an
y
tech
n
iq
u
es
ar
e
s
till
h
in
d
er
ed
b
y
th
e
d
if
f
icu
lty
o
f
f
in
e
-
t
u
n
in
g
LLMs
u
s
in
g
d
o
m
ain
-
s
p
ec
if
i
c
d
ata
[
2
1
]
.
T
h
ese
s
h
o
r
tco
m
in
g
s
h
ig
h
lig
h
t
th
e
n
ee
d
f
o
r
a
m
u
ltimo
d
al
f
r
am
ewo
r
k
th
at
ad
d
r
ess
es
th
e
s
ca
lab
ilit
y
an
d
in
ter
p
r
etab
ilit
y
o
f
AI
s
y
s
tem
s
in
r
ea
l
-
tim
e
s
ce
n
ar
io
s
f
o
r
a
tr
u
s
two
r
th
y
,
s
ca
lab
le,
an
d
m
o
r
ally
s
o
u
n
d
m
en
tal
h
ea
lth
ev
alu
atio
n
.
T
h
e
E
m
o
Vib
e
m
o
d
el,
a
p
o
s
t
-
f
u
s
io
n
m
eth
o
d
o
lo
g
y
,
p
r
eser
v
es
m
o
d
ality
-
s
p
ec
if
ic
p
r
o
p
er
ties
wh
ile
r
eso
lv
in
g
cr
o
s
s
-
m
o
d
al
co
n
f
li
cts
b
y
u
s
in
g
atten
tio
n
.
I
t
co
m
b
in
es
p
r
e
-
tr
ain
ed
m
o
d
els
s
u
ch
as
B
E
R
T
an
d
R
esNet
-
5
0
with
a
cu
s
to
m
ized
L
STM
an
d
C
NN
ar
ch
itectu
r
e
with
h
ig
h
p
er
f
o
r
m
a
n
ce
an
d
lo
w
co
m
p
u
tatio
n
a
l
co
m
p
lex
ity
.
I
n
ter
esti
n
g
ly
,
it
i
n
co
r
p
o
r
ates
em
o
tico
n
s
,
tex
t,
a
n
d
im
ag
es
t
o
p
r
o
v
id
e
h
ig
h
co
n
tex
tu
al
s
en
s
itiv
ity
to
war
d
s
em
o
tio
n
al
cu
es.
T
h
e
f
r
am
ewo
r
k
is
o
p
tim
ized
f
o
r
r
ea
l
-
tim
e
d
ep
lo
y
m
en
t
with
o
p
tim
ized
p
ip
elin
es
a
n
d
lig
h
t
co
n
f
ig
u
r
atio
n
s
,
en
a
b
lin
g
s
ca
lab
le
an
d
r
esp
o
n
s
iv
e
em
o
ti
o
n
tr
ac
k
in
g
.
E
m
o
Vib
e
r
e
m
ed
i
es
d
ef
icits
in
f
u
s
io
n
s
tr
ateg
y
,
m
o
d
ality
f
u
s
io
n
,
in
ter
p
r
etab
ilit
y
,
an
d
s
ca
lab
ilit
y
,
p
r
o
v
i
d
in
g
a
r
o
b
u
s
t,
eth
ical,
an
d
f
u
ll
-
f
led
g
e
d
s
o
lu
tio
n
f
o
r
AI
-
b
ased
m
e
n
tal
h
ea
lth
an
aly
s
is
o
n
th
e
in
ter
n
et.
Mu
ltimo
d
al
em
o
tio
n
r
ec
o
g
n
it
io
n
h
as
co
m
e
a
lo
n
g
way
,
y
et
th
e
ex
is
tin
g
m
o
d
els
h
av
e
m
o
d
ality
co
n
f
lict,
ex
p
en
s
iv
e
co
m
p
u
tati
o
n
al
co
s
t,
lim
ited
r
ea
l
-
tim
e
p
er
f
o
r
m
a
n
ce
,
an
d
in
s
ec
u
r
e
p
r
i
v
a
cy
.
I
n
ea
r
ly
m
eth
o
d
s
o
f
f
u
s
io
n
,
m
o
d
ality
-
s
p
ec
if
ic
k
n
o
wled
g
e
is
lik
ely
to
b
e
d
ilu
te
d
an
d
p
er
f
o
r
m
a
n
ce
d
e
g
r
ad
e
d
.
Mo
s
t
m
o
d
els
d
o
n
o
t
f
o
cu
s
o
n
em
o
tico
n
s
,
wh
ich
a
r
e
n
ee
d
ed
to
d
ec
o
d
e
n
u
an
ce
d
em
o
tio
n
al
r
esp
o
n
s
es
in
s
o
cial
m
ed
ia
p
o
s
ts
.
T
o
m
itig
ate
th
ese
r
estrictio
n
s
,
th
e
p
r
esen
t
r
ep
o
r
t
p
r
o
p
o
s
es
a
s
ca
lab
le,
eth
ical
an
d
r
ea
l
-
tim
e
m
u
ltimo
d
al
m
o
d
el
o
n
late
f
u
s
io
n
,
h
y
b
r
id
t
r
an
s
f
er
lea
r
n
in
g
a
n
d
em
o
tico
n
r
o
b
u
s
t in
te
g
r
atio
n
to
e
n
h
an
ce
em
o
tio
n
al
an
aly
s
is
.
3.
M
E
T
H
O
D
A
m
u
ltimo
d
al
m
en
tal
h
ea
lth
a
s
s
es
s
m
en
t
f
r
am
ewo
r
k
is
en
v
is
io
n
ed
to
tr
a
n
s
ce
n
d
th
e
s
h
o
r
tc
o
m
in
g
s
o
f
u
n
im
o
d
al
s
o
lu
tio
n
s
b
y
c
o
m
b
in
in
g
s
o
cial
m
e
d
ia
tex
t,
i
m
ag
es,
an
d
e
m
o
tico
n
s
.
L
e
v
er
ag
in
g
c
u
ttin
g
-
ed
g
e
m
ac
h
in
e
lear
n
i
n
g
/
d
ee
p
lear
n
in
g
m
eth
o
d
s
,
th
e
s
y
s
tem
will
o
f
f
er
r
ea
l
-
tim
e
u
s
er
m
en
tal
h
ea
lth
in
s
ig
h
ts
in
ter
m
s
o
f
in
d
icato
r
s
o
f
a
n
x
iety
,
a
n
g
er
,
an
d
d
ep
r
ess
io
n
.
T
h
e
r
esu
ltin
g
SME
D
will
v
is
u
ally
r
ep
r
esen
t
em
o
tio
n
al
tr
en
d
s
am
o
n
g
u
s
er
s
an
d
p
r
o
f
ess
io
n
al
s
,
en
ab
lin
g
ea
r
ly
d
etec
tio
n
an
d
o
n
g
o
in
g
m
o
n
ito
r
in
g
.
T
h
e
d
ata
co
llectio
n
m
o
d
u
le
co
llects
m
u
ltimo
d
al
d
ata
th
r
o
u
g
h
s
o
cial
m
e
d
ia
API
s
,
g
r
o
u
p
in
g
it
b
y
e
m
o
tio
n
al
ca
te
g
o
r
ies
an
d
ap
p
ly
in
g
s
tan
d
ar
d
ized
f
o
r
m
ats.
Pre
p
r
o
c
ess
in
g
o
p
er
atio
n
s
ar
e
ap
p
lied
to
all
m
o
d
ality
ty
p
es
an
d
s
tr
ict
q
u
ality
co
n
tr
o
l
g
u
ar
an
tees
d
ata
r
eliab
ilit
y
f
o
r
ac
cu
r
ate
m
o
d
el
p
r
e
d
ictio
n
s
.
T
h
e
o
v
er
all
s
tr
u
ctu
r
e
o
f
th
e
en
v
is
ag
ed
E
m
o
Vib
e
f
r
am
ewo
r
k
is
d
ep
icted
in
Fig
u
r
es 1
an
d
2
,
s
h
o
win
g
two
d
if
f
e
r
en
t stag
es o
f
th
e
p
i
p
elin
e.
Fig
u
r
e
1
illu
s
tr
ates th
e
f
ir
s
t
p
h
ase,
s
tar
tin
g
f
r
o
m
th
e
d
ata
co
llectio
n
m
o
d
u
le.
I
t
is
wh
er
e
tex
t
co
n
te
n
t,
em
o
tico
n
s
,
an
d
im
ag
es
ar
e
m
in
ed
f
r
o
m
s
o
cial
m
ed
ia
u
s
in
g
API
in
teg
r
atio
n
.
T
h
e
p
r
e
-
p
r
o
ce
s
s
in
g
o
p
er
atio
n
s
ar
e
m
o
d
ality
-
d
ep
en
d
en
t:
tex
t
d
ata
is
to
k
en
ized
an
d
p
a
d
d
ed
,
im
ag
es
r
esized
an
d
a
u
g
m
en
t
ed
,
wh
ile
em
o
tico
n
s
ar
e
tr
an
s
l
ated
in
to
s
en
tim
en
t
class
es.
E
ac
h
m
o
d
ality
m
u
s
t b
e
clea
n
ed
an
d
n
o
r
m
alize
d
h
e
r
e
in
p
r
ep
a
r
atio
n
f
o
r
p
ar
allel
p
r
o
ce
s
s
in
g
.
T
h
e
p
r
im
ar
y
an
al
y
s
is
m
o
d
u
l
e
co
m
b
i
n
es
a
n
L
S
T
M
te
x
t
a
n
a
ly
s
is
m
o
d
el
a
n
d
a
C
N
N
i
m
a
g
e
a
n
a
ly
s
is
m
o
d
el.
T
r
ai
n
i
n
g
ea
ch
m
o
d
el
i
n
d
i
v
i
d
u
all
y
a
n
d
f
u
s
i
n
g
t
h
ei
r
o
u
t
p
u
ts
i
n
a
c
r
o
s
s
-
m
o
d
al
i
n
te
r
a
cti
o
n
la
y
e
r
i
m
p
r
o
v
es
th
e
s
y
s
t
em
'
s
a
b
i
lit
y
t
o
m
a
k
e
ac
c
u
r
at
e
e
m
o
ti
o
n
al
p
r
e
d
i
cti
o
n
s
.
Pr
e
-
tr
ai
n
e
d
m
o
d
els
a
r
e
o
p
ti
m
iz
ed
u
s
i
n
g
h
y
b
r
i
d
tr
a
n
s
f
e
r
l
ea
r
n
i
n
g
,
e
n
s
u
r
i
n
g
m
ax
i
m
u
m
s
y
s
t
e
m
p
e
r
f
o
r
m
an
ce
o
n
r
ea
l
-
w
o
r
l
d
d
at
a.
T
h
e
m
o
d
el
c
an
ca
p
t
u
r
e
f
i
n
e
-
g
r
ai
n
e
d
e
m
o
ti
o
n
al
p
att
er
n
s
t
h
at
u
n
i
m
o
d
al
s
y
s
t
em
s
t
en
d
t
o
m
is
s
d
u
e
t
o
c
o
m
b
i
n
i
n
g
s
e
p
a
r
a
te
m
o
d
al
it
y
-
s
p
ec
i
f
i
c
m
o
d
els
an
d
a
f
u
s
i
o
n
c
o
m
p
o
n
en
t
.
Du
e
to
t
h
ei
r
s
tat
e
-
of
-
t
h
e
-
a
r
t
ca
p
a
cit
y
t
o
p
e
r
c
ei
v
e
s
e
m
an
t
ic
n
u
a
n
ce
s
a
n
d
co
n
t
ex
tu
al
r
ela
ti
o
n
s
h
i
p
s
,
p
r
e
-
t
r
ain
e
d
m
o
d
els
li
k
e
B
E
R
T
an
d
Dis
tilB
E
R
T
we
r
e
u
til
iz
ed
f
o
r
t
ex
t
p
r
o
ce
s
s
i
n
g
[
2
6
]
.
T
h
e
m
o
d
els
wer
e
f
in
e
-
tu
n
ed
f
o
r
d
o
m
ain
-
s
p
ec
if
ic
u
s
e
to
en
h
an
ce
th
eir
p
e
r
f
o
r
m
an
ce
o
n
d
at
asets
s
u
ch
as
Go
E
m
o
tio
n
s
.
Pre
-
t
r
ain
ed
m
o
d
els
s
u
ch
as
VGG
-
1
6
a
n
d
R
esNet
-
5
0
wer
e
em
p
lo
y
e
d
f
o
r
im
ag
e
d
ata
to
lev
er
ag
e
th
eir
r
o
b
u
s
t
f
ea
tu
r
e
ex
tr
ac
tio
n
ab
ilit
y
,
p
ar
ticu
lar
ly
f
o
r
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
[
2
7
]
.
C
u
s
to
m
C
NN
an
d
L
STM
m
o
d
els
wer
e
also
cr
ea
ted
to
s
tr
ik
e
a
co
m
p
r
o
m
is
e
b
etwe
en
c
o
m
p
u
tatio
n
a
l
ef
f
ec
tiv
en
ess
an
d
f
lex
ib
ilit
y
to
s
p
ec
if
ic
d
ata
p
r
o
p
er
ties
.
T
o
im
p
r
o
v
e
d
o
m
ain
-
s
p
ec
if
ic
p
er
f
o
r
m
an
ce
,
h
y
b
r
id
c
o
n
f
ig
u
r
atio
n
s
wer
e
cr
ea
ted
b
y
f
u
s
in
g
th
e
ad
v
a
n
tag
es o
f
p
r
e
-
tr
ai
n
ed
m
o
d
els with
u
n
iq
u
e
d
esig
n
s
.
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
E
mo
V
ib
e:
A
I
-
d
r
iven
mu
ltimo
d
a
l e
mo
tio
n
a
n
a
lysi
s
fo
r
men
ta
l h
ea
lth
…
(
Dee
p
a
li V
o
r
a
)
4569
Fig
u
r
e
1
.
Pro
p
o
s
ed
s
y
s
tem
ar
c
h
itectu
r
e
-
p
h
ase
1
Fig
u
r
e
2
.
Pro
p
o
s
ed
s
y
s
tem
ar
c
h
itectu
r
e
-
p
h
ase
2
Fig
u
r
e
2
illu
s
tr
ates
th
e
s
ec
o
n
d
s
tag
e,
i
n
wh
ich
p
r
o
ce
s
s
ed
in
p
u
ts
ar
e
in
p
u
t
to
s
p
ec
ialized
m
o
d
el
s
tr
ea
m
s
:
an
L
STM
-
b
ased
s
tr
ea
m
f
o
r
s
eq
u
en
tial
tex
t
d
ata
,
a
C
NN
-
b
ased
s
tr
ea
m
f
o
r
v
is
u
al
in
p
u
t,
an
d
a
s
y
m
b
o
lic
m
ap
p
in
g
s
tr
ea
m
f
o
r
em
o
tico
n
s
.
T
h
ese
s
tr
ea
m
s
'
o
u
tp
u
ts
ar
e
in
p
u
t
to
a
m
u
ltimo
d
al
f
u
s
io
n
lay
er
,
wh
er
e
atten
tio
n
-
b
ased
late
f
u
s
io
n
is
p
er
f
o
r
m
ed
to
r
eso
l
v
e
co
n
f
licts
ac
r
o
s
s
m
o
d
alities
an
d
d
y
n
am
ically
weig
h
co
n
tr
ib
u
tio
n
s
f
r
o
m
ev
e
r
y
s
t
r
ea
m
.
Ou
r
f
u
s
io
n
lay
er
im
p
lem
en
ts
tr
an
s
f
o
r
m
er
-
b
ased
s
ca
led
d
o
t
-
p
r
o
d
u
ct
atten
tio
n
.
Giv
en
m
o
d
ality
-
s
p
e
cif
ic
q
u
er
y
Q,
k
ey
K,
an
d
v
alu
e
V
m
atr
ices,
we
co
m
p
u
te
as i
n
(
1
)
.
(
,
,
)
=
(
√
)
(
1
)
W
h
er
e
is
th
e
d
im
en
s
io
n
ality
o
f
th
e
k
e
y
v
ec
to
r
s
.
W
e
em
p
lo
y
m
u
lti
-
h
ea
d
atten
tio
n
with
h
p
ar
allel
h
ea
d
s
—
co
n
ca
ten
atin
g
an
d
lin
ea
r
ly
p
r
o
jectin
g
th
eir
o
u
tp
u
ts
—
to
allo
w
th
e
m
o
d
el
to
jo
in
tly
atten
d
to
in
f
o
r
m
atio
n
f
r
o
m
d
if
f
er
en
t
r
e
p
r
esen
tatio
n
s
u
b
s
p
ac
es
[
2
8
]
.
T
h
is
co
m
b
in
ati
o
n
m
ec
h
an
is
m
en
s
u
r
es
th
at
ev
er
y
m
o
d
ality
'
s
h
ig
h
-
lev
el
s
em
an
tic
f
ea
tu
r
es
ar
e
alig
n
ed
co
n
te
x
tu
ally
b
ef
o
r
e
f
in
al
class
if
icatio
n
.
C
u
r
r
en
tly
,
wh
en
co
n
f
lictin
g
em
o
tio
n
s
co
r
es
o
cc
u
r
f
o
r
a
tex
t,
v
is
io
n
,
an
d
em
o
tico
n
s
tr
ea
m
,
E
m
o
Vib
e
ca
lcu
lates
th
eir
f
in
al
p
r
ed
ictio
n
as
th
e
av
er
ag
e
o
f
th
e
th
r
ee
m
o
d
ality
s
co
r
es,
wh
ile
m
ain
tain
in
g
eq
u
al
weig
h
t
f
o
r
ea
ch
s
co
r
e.
Alth
o
u
g
h
th
is
ap
p
r
o
ac
h
p
r
ev
en
ts
o
v
er
-
o
r
u
n
d
e
r
-
em
p
h
asizin
g
an
y
o
n
e
co
n
tr
i
b
u
tio
n
,
it
m
ay
m
in
im
ize
u
s
in
g
th
e
m
o
s
t
d
ep
en
d
ab
le
m
o
d
ality
u
n
d
er
p
ar
ticu
lar
cir
c
u
m
s
tan
ce
s
.
As th
e
s
y
s
tem
'
s
u
s
er
in
ter
f
ac
e,
th
e
d
ash
b
o
ar
d
m
o
d
u
le
co
m
b
i
n
es d
ata
in
s
ig
h
ts
in
to
an
attr
ac
tiv
e
d
esig
n
th
at
d
is
p
lay
s
em
o
tio
n
al
s
h
if
ts
an
d
r
ea
l
-
tim
e
b
eh
av
i
o
r
al
tr
en
d
s
.
I
n
ad
d
itio
n
,
th
e
d
ash
b
o
ar
d
ca
n
tr
ac
k
m
etr
ics
ac
r
o
s
s
s
p
ec
if
ic
tim
ef
r
am
es
lik
e
d
aily
,
wee
k
ly
,
o
r
m
o
n
th
l
y
.
Fo
r
co
n
s
u
m
er
s
a
n
d
m
e
n
tal
h
e
alth
ex
p
er
ts
,
to
o
ls
lik
e
lin
e
ch
ar
ts
,
b
ar
g
r
ap
h
s
,
an
d
tr
en
d
lin
es
d
e
p
ict
em
o
tio
n
al
s
tates
ac
r
o
s
s
tim
e,
m
ak
in
g
th
e
m
an
in
d
is
p
en
s
ab
le
d
ec
is
io
n
-
s
u
p
p
o
r
t to
o
l.
Sin
ce
m
e
n
t
al
h
ea
l
th
i
n
f
o
r
m
at
i
o
n
is
p
e
r
s
o
n
a
l,
t
h
e
a
r
c
h
it
ec
tu
r
e
h
as
r
o
b
u
s
t
p
r
i
v
ac
y
a
n
d
s
e
c
u
r
ity
c
o
n
t
r
o
ls
to
p
r
o
te
ct
u
s
e
r
in
f
o
r
m
a
ti
o
n
.
An
o
n
y
m
i
za
ti
o
n
s
y
s
t
em
s
e
n
s
u
r
e
d
at
a
p
r
iv
ac
y
b
y
eli
m
i
n
at
in
g
p
er
s
o
n
al
i
d
e
n
ti
f
i
er
s
d
u
r
i
n
g
co
lle
cti
o
n
[
2
9
]
.
R
o
le
-
b
ase
d
a
cc
ess
c
o
n
t
r
o
ls
l
im
i
t
a
cc
e
s
s
to
d
ata
t
o
t
h
e
a
p
p
r
o
v
e
d
u
s
e
r
s
,
w
h
il
e
en
cr
y
p
ti
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
5
6
5
-
4
5
7
8
4570
s
ec
u
r
es
d
at
a
w
h
e
n
it
is
tr
a
n
s
m
itte
d
a
n
d
s
t
o
r
e
d
.
Asi
d
e
f
r
o
m
o
f
f
er
in
g
t
ec
h
n
ic
al
s
af
eg
u
ar
d
s
,
t
h
e
s
y
s
t
em
ad
d
r
ess
es
b
r
o
ad
er
et
h
i
ca
l
c
o
n
c
er
n
s
r
eg
a
r
d
in
g
its
u
s
ag
e.
T
h
e
a
p
p
r
o
a
c
h
em
p
l
o
y
s
a
d
v
a
n
c
e
d
p
r
e
p
r
o
c
ess
in
g
m
et
h
o
d
s
th
at
en
s
u
r
e
s
o
u
n
d
d
a
ta
h
a
n
d
li
n
g
to
a
v
er
t
m
is
in
te
r
p
r
eta
ti
o
n
o
f
d
il
ig
en
tl
y
s
el
ec
te
d
c
o
n
te
n
t
o
r
i
n
f
la
te
d
b
ias
.
B
y
s
h
u
n
n
i
n
g
a
n
o
v
e
r
-
r
eli
an
ce
o
n
an
y
m
o
d
a
lit
y
,
m
o
d
al
c
o
n
te
x
t
u
a
l
an
al
y
s
is
d
ec
r
e
ases
t
h
e
li
k
e
li
h
o
o
d
o
f
m
is
c
lass
i
f
i
ca
ti
o
n
.
A
n
e
th
ica
l
a
n
d
t
r
an
s
p
a
r
en
t
en
v
i
r
o
n
m
en
t
f
o
r
d
ea
li
n
g
wit
h
s
e
n
s
it
iv
e
m
en
tal
h
ea
lt
h
i
n
f
o
r
m
a
ti
o
n
is
es
ta
b
lis
h
e
d
b
y
h
o
n
es
t
an
d
et
h
ic
al
s
ta
n
d
a
r
d
s
m
a
in
tai
n
e
d
t
h
r
o
u
g
h
o
p
e
n
u
s
e
r
co
n
s
en
t
p
r
o
ce
s
s
es
[
3
0
]
.
A
s
y
s
tem
atic
ev
alu
atio
n
p
r
o
ce
d
u
r
e
is
ap
p
lied
to
th
e
f
r
am
ew
o
r
k
to
en
s
u
r
e
its
r
eliab
ilit
y
an
d
s
tr
en
g
th
.
A
s
tr
atif
ied
tr
ain
in
g
-
v
alid
atio
n
-
test
in
g
r
atio
en
s
u
r
es
p
r
o
p
er
r
ep
r
esen
tatio
n
o
f
all
em
o
tio
n
a
l
class
es,
an
d
cr
o
s
s
-
v
alid
atio
n
is
em
p
lo
y
ed
t
o
p
r
e
v
en
t
o
v
e
r
f
itti
n
g
a
n
d
p
r
o
v
i
d
e
b
alan
ce
d
m
o
d
el
ev
al
u
atio
n
s
.
C
o
n
f
u
s
io
n
m
atr
ices
ar
e
an
aly
ze
d
to
i
d
en
tify
a
n
d
c
o
r
r
ec
t
co
m
m
o
n
m
is
class
if
icati
o
n
s
.
T
o
co
n
f
ir
m
th
e
s
y
s
tem
'
s
u
s
ef
u
ln
ess
in
ac
tu
al
en
v
ir
o
n
m
en
ts
,
its
r
ea
l
-
tim
e
p
er
f
o
r
m
an
ce
is
also
test
ed
i
n
d
if
f
e
r
en
t
d
ata
l
o
ad
s
,
m
ea
s
u
r
in
g
laten
c
y
an
d
s
ca
lab
ilit
y
[
3
1
]
.
T
h
e
d
esig
n
an
d
ap
p
licatio
n
o
f
th
e
f
r
am
ewo
r
k
r
est
o
n
s
o
m
e
ass
u
m
p
tio
n
s
.
T
h
e
f
r
am
ewo
r
k
ass
u
m
e
s
p
u
b
licly
av
ailab
le
s
o
cial
m
e
d
ia
d
ata
ca
n
in
f
o
r
m
u
s
er
s
'
m
en
tal
h
ea
lth
s
tatu
s
es,
as
u
s
er
-
g
en
er
ate
d
co
n
te
n
t
ca
r
r
ies
u
s
ef
u
l
in
f
o
r
m
atio
n
ab
o
u
t
em
o
tio
n
al
h
ea
lth
.
T
h
e
s
y
s
tem
also
ass
u
m
es
ac
ce
s
s
to
h
ig
h
-
p
er
f
o
r
m
an
ce
co
m
p
u
tin
g
r
eso
u
r
ce
s
in
o
r
d
er
to
d
ea
l
with
lar
g
e
m
u
ltimo
d
al
d
atasets
'
tr
ain
in
g
an
d
r
ea
l
-
tim
e
d
ep
lo
y
m
en
t.
T
h
e
p
r
o
ject
d
e
v
elo
p
ed
o
n
an
HP
Z
8
GPU
wo
r
k
s
tatio
n
with
tw
o
AM
D
R
ad
eo
n
PR
O
W
6
8
0
0
GPUS
an
d
3
2
GB
R
AM
to
ac
ce
ler
ate
tr
ain
in
g
.
T
h
e
s
y
s
tem
p
r
esu
m
es
th
at
m
u
lt
im
o
d
al
d
ata
im
p
r
o
v
es
d
etec
tio
n
o
f
s
u
b
tle
m
en
tal
h
ea
lth
m
ar
k
e
r
s
b
y
co
n
v
ey
i
n
g
m
o
r
e
in
f
o
r
m
atio
n
ab
o
u
t
em
o
ti
o
n
s
th
an
an
y
s
in
g
le
m
o
d
ality
.
L
astl
y
,
th
e
s
y
s
tem
d
ep
en
d
s
o
n
s
tr
ict
p
r
iv
ac
y
r
eg
u
latio
n
s
to
g
u
ar
an
tee
m
o
r
al
c
o
m
p
lian
ce
with
o
u
t
s
ac
r
if
icin
g
o
p
en
n
ess
in
d
at
a
p
r
o
ce
s
s
in
g
.
T
h
e
war
n
in
g
th
r
es
h
o
ld
s
o
f
th
e
d
ash
b
o
ar
d
an
d
p
e
r
s
o
n
aliza
tio
n
f
ea
tu
r
es
en
ab
le
e
x
p
er
ts
an
d
u
s
er
s
to
cu
s
to
m
ize
th
eir
ex
p
e
r
ien
ce
.
4.
I
M
P
L
E
M
E
NT
A
T
I
O
N
D
E
T
AIL
S
T
h
e
p
r
o
p
o
s
ed
m
u
ltimo
d
al
f
r
a
m
ewo
r
k
i
n
teg
r
ates
v
a
r
io
u
s
d
at
a
m
o
d
alities
an
d
late
f
u
s
io
n
is
em
p
lo
y
ed
in
th
e
m
u
ltimo
d
al
f
u
s
io
n
p
r
o
c
ess
to
p
r
eser
v
e
m
o
d
ality
-
s
p
ec
i
f
ic
f
ea
tu
r
es,
allo
win
g
tex
t,
im
a
g
es,
an
d
em
o
tico
n
s
to
b
e
p
r
o
ce
s
s
ed
s
ep
ar
ately
b
ef
o
r
e
th
eir
o
u
tp
u
ts
ar
e
m
er
g
ed
[
3
2
]
.
L
ate
f
u
s
io
n
was
ch
o
s
en
b
ec
au
s
e
it
p
r
ev
en
ts
f
ea
tu
r
e
d
ilu
tio
n
an
d
m
ak
es
m
o
r
e
ac
cu
r
ate
an
d
s
u
b
tle
em
o
ti
o
n
al
p
r
ed
ictio
n
s
[
3
3
]
.
L
STM
s
ar
e
g
o
o
d
at
d
ea
lin
g
with
co
n
tex
tu
al
an
d
s
eq
u
e
n
tial
d
ata
th
eir
u
s
e
f
o
r
tex
t
a
n
al
y
s
is
was
s
u
p
p
o
r
ted
.
Ho
wev
er
,
th
e
B
E
R
T
’
s
h
ig
h
co
m
p
u
tin
g
n
ee
d
s
m
ad
e
it
im
p
r
ac
tical
f
o
r
r
ea
l
-
tim
e
u
s
e.
As
C
NNs
wo
r
k
ef
f
ec
tiv
ely
with
s
m
all
d
atasets
,
th
ey
wer
e
s
elec
ted
f
o
r
im
ag
e
an
aly
s
is
.
No
r
m
aliza
tio
n
in
tex
t p
r
o
c
ess
in
g
is
tak
en
ca
r
e
o
f
b
y
n
atu
r
al
lan
g
u
ag
e
to
o
lk
it
(
NL
T
K
)
an
d
T
en
s
o
r
Flo
w.
Op
en
C
V
tr
an
s
f
o
r
m
s
im
a
g
es
an
d
m
ap
s
em
o
tico
n
s
to
ca
teg
o
r
ies
o
f
s
en
tim
en
ts
.
Usi
n
g
p
ar
allel
p
r
o
ce
s
s
in
g
an
d
h
ig
h
ly
o
p
tim
ize
d
p
ip
elin
e
d
esig
n
s
,
th
e
s
y
s
tem
r
ed
u
ce
s
laten
cy
ev
en
wh
en
wo
r
k
in
g
with
lar
g
e
v
o
lu
m
es o
f
d
ata.
T
h
e
im
p
lem
en
ted
s
y
s
tem
is
o
r
g
an
ized
as f
o
llo
ws
.
4
.
1
.
Da
t
a
c
o
llect
io
n a
nd
prepa
ra
t
io
n
R
eg
ar
d
in
g
d
ataset
s
elec
tio
n
,
th
e
f
r
am
ewo
r
k
em
p
lo
y
s
d
atasets
lik
e
Go
E
m
o
tio
n
s
[
3
4
]
f
o
r
tex
t
-
b
ased
em
o
tio
n
lab
els
an
d
FER
[
3
5
]
/
Af
f
ec
tNet
[
3
6
]
f
o
r
im
ag
e
-
b
ased
em
o
ti
o
n
r
ec
o
g
n
itio
n
.
T
h
e
G
o
E
m
o
tio
n
s
d
ataset,
d
ev
elo
p
e
d
b
y
Go
o
g
le
R
esear
ch
,
in
clu
d
es
tex
t
d
ata
th
at
ca
p
tu
r
es
co
m
p
lex
em
o
tio
n
al
n
u
an
ce
s
.
T
h
e
d
ataset
d
is
tr
ib
u
tio
n
an
d
ex
am
p
les ar
e
illu
s
tr
ated
in
Fig
u
r
e
3
.
Fig
u
r
e
3
.
T
e
x
t a
n
d
em
o
tico
n
d
ata
s
am
p
le
T
h
e
FER
an
d
Af
f
ec
tNet
d
at
asets
co
n
tain
im
ag
es
o
f
f
ac
i
al
ex
p
r
ess
io
n
s
co
r
r
esp
o
n
d
in
g
to
v
ar
io
u
s
em
o
tio
n
al
s
tates.
Fig
u
r
e
4
,
w
h
ich
ap
p
ea
r
s
as
f
o
llo
w
,
p
r
o
v
id
es
r
ep
r
esen
tativ
e
s
am
p
les
o
f
f
ac
ial
ex
p
r
ess
io
n
s
.
T
h
e
Fig
u
r
es
4
(
a
)
to
4
(f)
–
a
n
g
er
,
d
ig
ested
,
f
ea
r
f
u
l,
h
ap
p
y
,
n
eu
tr
al,
an
d
s
ad
,
p
r
esen
t
th
e
r
an
g
e
o
f
em
o
tio
n
s
an
aly
ze
d
,
h
ig
h
lig
h
tin
g
th
e
d
i
v
er
s
ity
o
f
em
o
tio
n
al
ex
p
r
ess
io
n
s
ca
p
tu
r
ed
v
ia
th
e
s
u
b
f
ig
u
r
es.
T
h
e
f
r
am
ewo
r
k
f
u
r
th
er
ex
a
m
in
es e
m
o
tio
n
s
f
r
o
m
a
m
u
ltid
im
en
s
io
n
al
p
er
s
p
ec
tiv
e,
m
ak
in
g
it m
o
r
e
s
en
s
itiv
e
to
s
u
b
tle
em
o
tio
n
al
cu
es
with
in
co
n
ten
t
g
e
n
er
ate
d
o
n
s
o
cial
m
e
d
ia.
T
o
ac
h
iev
e
d
ataset
ef
f
icien
cy
,
it'
s
s
o
r
ted
b
y
em
o
tio
n
a
n
d
m
o
d
ality
,
with
im
ag
es
ca
teg
o
r
ized
b
y
tag
g
ed
ex
p
r
ess
io
n
s
an
d
tex
t
f
iles
b
y
em
o
tio
n
wh
ich
h
elp
s
in
r
ap
id
d
ata
ac
ce
s
s
wh
en
th
e
m
o
d
el
is
b
ein
g
tr
ain
ed
an
d
test
ed
.
T
h
e
s
y
s
tem
u
p
d
ates
th
e
d
ata
with
th
e
latest
r
elev
an
t
m
ater
ial
in
s
o
cial
m
ed
ia
in
r
ea
l
-
tim
e,
k
ee
p
in
g
p
ac
e
with
ch
an
g
in
g
p
atter
n
s
o
f
lan
g
u
ag
e
u
s
ag
e,
s
lan
g
,
an
d
v
is
u
al
m
o
d
es
o
f
ex
p
r
ess
io
n
.
T
im
e
-
b
ased
r
elev
an
ce
e
n
tails
r
eg
u
lar
u
p
d
atin
g
o
f
d
ata
r
e
p
r
esen
tin
g
r
ea
l
-
tim
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
E
mo
V
ib
e:
A
I
-
d
r
iven
mu
ltimo
d
a
l e
mo
tio
n
a
n
a
lysi
s
fo
r
men
ta
l h
ea
lth
…
(
Dee
p
a
li V
o
r
a
)
4571
em
o
tio
n
al
s
tates.
I
t
is
ca
p
ab
le
o
f
h
an
d
lin
g
th
e
m
o
s
t
r
ec
e
n
t
t
r
en
d
s
o
f
u
s
er
ac
tiv
ities
to
r
em
ain
q
u
ic
k
to
en
a
b
le
u
s
ef
u
l
m
o
n
ito
r
in
g
in
r
ea
l
ti
m
e.
Fig
u
r
e
4
p
r
esen
ts
a
v
is
u
al
s
am
p
le
f
r
o
m
th
ese
d
atasets
,
h
ig
h
lig
h
tin
g
t
h
e
d
iv
er
s
ity
o
f
em
o
tio
n
al
ex
p
r
ess
io
n
s
ca
p
tu
r
ed
v
ia
Fig
u
r
es 4
(
a)
to
4
(
f
)
.
I
n
th
is
r
esear
ch
,
all
s
o
cial
m
e
d
ia
d
ata
em
p
l
o
y
ed
we
r
e
g
ath
e
r
ed
f
r
o
m
p
u
b
licly
ac
ce
s
s
ib
le
s
o
u
r
ce
s
b
y
th
e
p
latf
o
r
m
'
s
ter
m
s
o
f
s
er
v
ice.
T
o
en
s
u
r
e
eth
ical
co
n
s
id
er
ati
o
n
s
,
we
en
s
u
r
ed
d
ata
u
s
e
co
m
p
lied
with
r
elev
an
t
g
u
id
elin
es
f
o
r
r
esp
o
n
s
ib
le
AI
r
esear
ch
,
s
u
c
h
as
r
esp
ec
t
f
o
r
u
s
er
p
r
iv
ac
y
an
d
co
n
s
en
t
wh
er
e
n
ec
ess
ar
y
.
I
n
ad
d
itio
n
,
we
test
ed
f
o
r
lan
g
u
a
g
e,
g
eo
g
r
ap
h
y
,
a
n
d
d
e
m
o
g
r
a
p
h
ic
r
ep
r
esen
tatio
n
b
iases
,
an
d
we
will
co
n
tin
u
e
to
r
ed
u
ce
s
u
ch
b
iases
in
s
u
b
s
eq
u
en
t w
o
r
k
th
r
o
u
g
h
m
o
r
e
b
ala
n
c
ed
an
d
d
iv
er
s
e
d
ata
s
am
p
lin
g
ap
p
r
o
ac
h
es.
(
a)
(
b
)
(
c)
(
d
)
(
e)
(f)
Fig
u
r
e
4
.
Sam
p
le
im
a
g
e
d
atas
et
lab
els s
h
o
win
g
em
o
tio
n
s
th
r
o
u
g
h
f
ac
ial
ex
p
r
ess
io
n
s
of
(a
)
an
g
er
,
(
b
)
d
is
g
u
s
ted
,
(
c)
f
ea
r
f
u
l
,
(
d
)
h
ap
p
y
,
(
e)
n
e
u
tr
al
,
an
d
(
f
)
s
ad
4
.
2
.
Da
t
a
prepro
ce
s
s
ing
Sp
ec
ial
p
r
ep
r
o
ce
s
s
in
g
is
n
ee
d
ed
f
o
r
ea
c
h
ty
p
e
o
f
d
ata.
T
h
e
Ker
as
to
k
en
izer
is
u
s
ed
to
co
n
v
er
t
wo
r
d
s
in
to
in
teg
er
in
d
ices.
Fu
r
th
er
p
r
o
ce
s
s
in
g
s
an
itizes
th
e
tex
t,
elim
in
atin
g
s
p
ec
ial
c
h
ar
ac
te
r
s
an
d
s
to
p
wo
r
d
s
.
Seq
u
en
ce
s
ar
e
p
ad
d
e
d
to
a
p
r
e
d
eter
m
in
ed
len
g
th
o
f
1
0
0
to
k
e
n
s
f
o
r
n
o
r
m
alizin
g
in
p
u
t
s
h
ap
e
s
f
o
r
p
r
o
ce
s
s
in
g
in
b
atch
es,
o
p
tim
izin
g
c
o
m
p
u
tin
g
ef
f
icien
c
y
,
a
n
d
en
h
a
n
cin
g
m
o
d
el
p
e
r
f
o
r
m
an
ce
.
Pictu
r
e
d
at
a
is
p
r
ep
r
o
ce
s
s
ed
b
y
r
esizin
g
im
ag
es
to
1
5
0
×
1
5
0
p
ix
els
f
o
r
cu
s
to
m
m
o
d
els
an
d
2
2
4
×
2
2
4
p
i
x
els
f
o
r
R
esNe
t
-
5
0
an
d
VGG
-
1
6
.
Scalin
g
p
ix
el
v
alu
es
b
etwe
en
0
an
d
1
e
n
h
an
ce
s
g
r
ad
ie
n
t
d
es
ce
n
t
ef
f
icien
c
y
a
n
d
m
o
d
el
co
n
v
er
g
en
ce
.
Var
io
u
s
d
ata
au
g
m
en
tatio
n
tech
n
iq
u
es
im
p
r
o
v
e
g
e
n
er
aliza
tio
n
an
d
s
tr
en
g
th
en
th
e
d
ataset,
allo
win
g
o
p
tim
al
m
o
d
el
p
er
f
o
r
m
an
ce
in
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
.
E
m
o
tico
n
s
ar
e
n
ec
ess
ar
y
m
ar
k
er
s
o
f
e
m
o
tio
n
al
co
n
ten
t.
T
o
ad
eq
u
ately
ex
p
r
ess
th
eir
m
ea
n
in
g
,
th
ey
ar
e
ass
o
ciate
d
with
p
r
e
-
d
ef
in
e
d
s
en
tim
en
t ta
g
s
.
Fo
r
in
s
tan
ce
,
th
e
"
"
em
o
tico
n
is
u
s
ed
f
o
r
h
ap
p
in
e
s
s
,
wh
ile
th
e
"
"
em
o
ji
is
u
s
ed
f
o
r
m
ela
n
ch
o
ly
.
B
y
p
r
o
v
i
d
in
g
co
n
tex
t
to
t
h
e
tex
t
d
ata,
th
is
m
ap
p
in
g
allo
ws
th
e
alg
o
r
ith
m
to
co
n
s
id
er
ca
s
es
wh
er
e
em
o
tico
n
s
r
ein
f
o
r
ce
o
r
r
e
f
u
te
th
e
s
en
tim
en
t
ex
p
r
ess
ed
in
t
h
e
t
ex
t.
T
h
e
s
y
s
tem
o
f
f
er
s
a
m
o
r
e
co
m
p
lete
a
n
d
i
n
tr
icate
e
m
o
tio
n
al
p
r
o
f
ile
b
y
in
clu
d
in
g
em
o
tico
n
s
in
th
e
em
o
tio
n
al
an
aly
s
is
,
s
ig
n
if
ican
tly
en
h
an
cin
g
its
ab
ilit
y
to
in
ter
p
r
et
s
o
cial
m
ed
ia
u
s
er
ex
p
r
ess
io
n
s
.
4
.
3
.
M
o
del
a
rc
hite
ct
ure
dev
elo
pm
ent
T
h
e
m
o
d
el
s
tr
u
ct
u
r
e
s
ep
a
r
at
es
te
x
t
a
n
d
im
ag
e
p
r
o
ce
s
s
in
g
b
e
f
o
r
e
f
u
s
io
n
t
o
e
x
tr
ac
t
u
s
e
f
u
l
p
atte
r
n
s
b
y
ea
c
h
s
u
b
m
o
d
e
l,
a
n
d
t
h
e
f
u
s
i
o
n
l
ay
e
r
m
e
r
g
es
i
n
s
i
g
h
ts
to
p
r
o
d
u
ce
t
h
e
o
v
e
r
a
ll
o
u
t
p
u
t.
T
h
e
t
ex
t
m
o
d
e
l
e
m
p
l
o
y
s
L
STM
l
a
y
e
r
s
f
o
r
s
eq
u
en
ce
d
ata
.
A
1
2
8
-
u
n
it
L
STM
lea
r
n
s
s
e
n
t
en
ce
w
o
r
d
r
el
ati
o
n
s
h
i
p
s
as
is
o
p
ti
m
al
f
o
r
r
e
co
g
n
iz
in
g
m
o
o
d
c
h
a
n
g
es
an
d
e
m
o
ti
o
n
al
t
o
n
e
b
e
ca
u
s
e
i
t
c
a
n
m
ai
n
ta
in
lo
n
g
te
x
t
s
eq
u
en
ce
s
.
A
d
e
n
s
e
l
ay
er
wit
h
R
eL
U
a
cti
v
a
ti
o
n
b
o
o
s
ts
in
te
r
p
r
et
ati
v
e
p
o
we
r
o
f
w
ea
k
e
m
o
tio
n
a
l
s
i
g
n
als
.
O
u
t
p
u
t
is
p
r
o
j
e
cte
d
t
o
d
e
p
r
ess
i
o
n
,
an
x
i
et
y
,
an
g
e
r
,
a
n
d
n
eu
tr
al
c
a
teg
o
r
i
es.
A
3
-
la
y
e
r
C
NN
m
o
d
el
wit
h
i
n
c
r
e
asi
n
g
f
i
lte
r
s
i
ze
s
(
3
2
,
6
4
,
1
2
8
)
a
n
d
Ma
x
Po
o
l
in
g
l
a
y
e
r
s
is
u
s
e
d
f
o
r
s
p
ati
al
in
f
o
r
m
a
ti
o
n
.
I
t
p
r
o
d
u
c
es
h
ie
r
a
r
c
h
i
ca
l
f
ac
i
al
ex
p
r
e
s
s
io
n
f
ea
tu
r
es
f
r
o
m
ed
g
es
t
o
l
ar
g
e
r
s
t
r
u
ct
u
r
es
li
k
e
s
m
iles
o
r
f
r
o
wn
s
.
A
d
e
n
s
e
la
y
e
r
wit
h
R
e
L
U
a
cti
v
at
io
n
f
i
n
e
-
t
u
n
es
t
h
e
f
ea
tu
r
es
,
r
es
u
lt
in
g
in
d
at
a
f
o
r
th
e
f
i
n
al
c
lass
i
f
ic
ati
o
n
l
ay
e
r
,
p
r
o
d
u
ci
n
g
t
h
r
ee
e
m
o
ti
o
n
cl
ass
es:
h
a
p
p
y
,
s
ad
,
a
n
d
an
g
r
y
.
T
h
e
h
y
b
r
i
d
B
E
R
T
+
L
ST
M
a
r
c
h
it
ec
tu
r
e
co
m
b
i
n
e
d
t
h
e
s
e
q
u
e
n
ti
al
p
r
o
ce
s
s
i
n
g
p
o
we
r
o
f
L
ST
M
wit
h
t
h
e
co
n
t
ex
tu
al
ad
v
a
n
t
a
g
es o
f
B
E
R
T
.
Fi
g
u
r
e
5
ill
u
s
t
r
a
tes th
e
h
y
b
r
i
d
te
x
t
-
p
r
o
c
ess
in
g
m
o
d
el
,
wh
ic
h
u
s
es
B
E
R
T
e
m
b
e
d
d
i
n
g
s
a
n
d
an
L
S
T
M
la
y
er
t
o
le
ar
n
s
e
m
a
n
tic
r
i
c
h
n
ess
a
n
d
s
e
q
u
e
n
ti
al
d
e
p
en
d
e
n
cies
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
5
6
5
-
4
5
7
8
4572
T
o
ad
ju
s
t
th
e
f
ea
tu
r
es
to
th
e
p
ar
ticu
lar
d
o
m
ain
r
e
q
u
ir
em
e
n
ts
o
f
th
e
FER/
Af
f
ec
tNet
d
atasets
,
th
e
R
esNet
-
5
0
+CNN
h
y
b
r
id
m
o
d
el
f
ir
s
t
ex
tr
ac
ted
h
ig
h
-
lev
el
f
e
atu
r
es
f
r
o
m
p
ictu
r
e
d
ata
u
s
in
g
R
esNe
t
-
5
0
.
T
h
ese
f
ea
tu
r
es
wer
e
t
h
en
s
en
t
v
ia
a
b
esp
o
k
e
C
NN
lay
er
.
Usi
n
g
p
r
e
-
tr
ain
ed
in
f
o
r
m
atio
n
an
d
a
d
ju
s
tin
g
to
d
o
m
ain
-
s
p
ec
if
ic
task
s
wer
e
b
alan
ce
d
b
y
th
ese
h
y
b
r
i
d
ar
r
an
g
em
e
n
ts
.
Fig
u
r
e
6
illu
s
tr
ates
th
e
h
y
b
r
i
d
im
ag
e
-
p
r
o
ce
s
s
in
g
m
o
d
el
th
at
c
o
m
b
in
es
R
esNet
-
5
0
'
s
p
r
e
-
tr
ain
ed
f
ea
t
u
r
es
with
a
cu
s
to
m
C
NN
h
ea
d
t
o
im
p
r
o
v
e
f
ac
ial
em
o
tio
n
d
etec
tio
n
ac
cu
r
ac
y
.
T
h
e
m
u
ltimo
d
al
f
u
s
io
n
lay
e
r
tr
ig
g
er
s
a
h
y
b
r
id
tr
a
n
s
f
er
lear
n
in
g
p
r
o
ce
s
s
to
c
o
m
b
in
e
th
e
im
ag
e
an
d
tex
t
m
o
d
el
o
u
tp
u
ts
th
er
e
b
y
i
m
p
r
o
v
i
n
g
u
n
d
er
s
tan
d
i
n
g
ac
r
o
s
s
th
e
m
o
d
alities
.
Ad
d
itio
n
ally
,
th
e
f
u
s
io
n
lay
e
r
im
p
r
o
v
es
in
te
r
p
r
etab
ilit
y
t
h
r
o
u
g
h
co
m
b
in
in
g
tex
t,
im
a
g
e
s
,
an
d
em
o
tico
n
d
ata
in
to
a
s
in
g
le
en
tity
,
th
u
s
p
r
o
v
id
i
n
g
a
m
o
r
e
in
teg
r
ated
ex
p
lan
atio
n
o
f
em
o
tio
n
al
s
tates
th
at
is
u
s
ef
u
l
in
d
is
t
in
g
u
is
h
in
g
co
m
p
lex
em
o
tio
n
al
p
atter
n
s
.
An
atten
tio
n
lay
er
d
y
n
am
ically
b
ala
n
ce
s
f
ea
tu
r
es
r
eso
lv
in
g
co
n
f
licts
ar
is
in
g
wh
en
m
o
d
alities
p
r
esen
t
co
n
f
lictin
g
v
iews
wh
ile
h
ig
h
lig
h
tin
g
co
n
tex
tu
ally
r
ele
v
an
t
f
ea
tu
r
es.
T
h
e
atten
tio
n
m
ec
h
an
is
m
s
u
p
p
o
r
ts
th
e
s
y
s
tem
in
h
an
d
lin
g
co
m
p
lex
em
o
ti
o
n
al
s
tates,
th
u
s
en
s
u
r
in
g
f
air
p
r
ed
ictio
n
s
ac
r
o
s
s
all
th
e
m
o
d
alities
.
T
h
e
m
u
ltimo
d
al
f
u
s
io
n
ar
ch
itectu
r
e
o
f
Fig
u
r
e
7
f
u
s
es
o
u
tp
u
ts
o
f
th
e
te
x
t
an
d
im
ag
e
m
o
d
els
(
as
well
a
s
s
en
tim
en
t
-
m
ap
p
ed
em
o
tico
n
s
)
b
ased
o
n
atten
tio
n
-
b
ased
late
f
u
s
io
n
.
T
h
is
ar
ch
itectu
r
e
en
ab
les
th
e
s
y
s
tem
to
weig
h
in
p
u
ts
ad
a
p
tiv
ely
b
ased
o
n
c
o
n
tex
tu
al
r
e
lev
an
ce
,
wh
ich
is
m
o
s
t
ef
f
ec
tiv
e
in
co
n
f
lictin
g
em
o
tio
n
al
s
ig
n
als b
etwe
en
m
o
d
alities
.
Fig
u
r
e
5
.
Hy
b
r
id
tex
t
-
b
ased
m
o
d
el
to
d
eter
m
in
e
an
d
m
ap
em
o
tio
n
f
r
o
m
tex
t
Fig
u
r
e
6
.
Hy
b
r
id
im
a
g
e
-
b
ased
m
o
d
el
f
o
r
f
ac
ial
em
o
ti
o
n
d
etec
tio
n
an
d
m
ap
p
in
g
Fig
u
r
e
7
.
Pro
p
o
s
ed
m
u
ltimo
d
a
l f
u
s
io
n
m
o
d
el
to
co
m
b
in
e
tex
t
u
al,
im
ag
e
an
d
em
o
tico
n
m
o
d
el
o
u
tp
u
ts
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
E
mo
V
ib
e:
A
I
-
d
r
iven
mu
ltimo
d
a
l e
mo
tio
n
a
n
a
lysi
s
fo
r
men
ta
l h
ea
lth
…
(
Dee
p
a
li V
o
r
a
)
4573
4
.
4
.
M
o
del
t
ra
ini
ng
a
nd
ev
a
l
ua
t
io
n
T
r
ain
in
g
a
n
d
test
in
g
ar
e
in
p
h
ases
.
I
n
itially
,
all
m
o
d
ality
-
s
p
ec
if
ic
m
o
d
els
ar
e
tr
ain
e
d
in
d
e
p
en
d
en
tly
.
L
ater
,
th
ey
a
r
e
co
llectiv
ely
r
ef
in
ed
th
r
o
u
g
h
f
u
s
io
n
lay
e
r
.
T
h
e
tex
t
m
o
d
el
is
s
u
p
er
v
i
s
ed
f
in
e
-
tu
n
e
d
f
o
r
1
6
ep
o
c
h
s
with
Ad
am
W
o
p
tim
izer
,
in
itial
lear
n
in
g
r
ate
2
×
1
0
⁻⁵,
weig
h
t
d
ec
ay
0
.
0
1
,
b
atch
s
ize
6
4
,
d
r
o
p
o
u
t
0
.
1
,
an
d
v
alid
atio
n
s
p
lit
2
0
%.
E
a
r
ly
s
to
p
p
in
g
with
3
e
p
o
ch
s
o
f
n
o
im
p
r
o
v
e
m
en
t.
T
h
e
im
ag
e
-
b
ased
m
o
d
els
ar
e
tr
ain
ed
f
o
r
u
p
to
3
2
e
p
o
ch
s
u
n
d
er
th
e
s
am
e
co
n
d
itio
n
s
,
in
itia
l
lear
n
in
g
r
ate
o
f
o
p
tim
izer
1
×
1
0
⁻⁴,
b
atch
s
ize
3
2
,
d
r
o
p
o
u
t
0
.
1
,
v
alid
atio
n
s
p
lit
2
0
%,
an
d
ea
r
ly
s
to
p
p
in
g
a
t
3
s
tatic
ep
o
ch
s
.
Hig
h
-
lev
e
l
em
b
ed
d
in
g
s
ar
e
co
n
ca
ten
ated
af
te
r
ev
er
y
b
ac
k
b
o
n
e
c
h
ec
k
p
o
in
t
in
th
e
late
-
f
u
s
io
n
lay
er
.
J
o
in
t
f
in
e
-
t
u
n
in
g
is
co
n
d
u
cte
d
f
o
r
1
0
ep
o
ch
s
u
s
in
g
Ad
am
W
o
p
tim
izer
,
lear
n
in
g
r
ate
1
×1
0
⁻⁵,
weig
h
t
d
ec
ay
0
.
0
1
,
d
r
o
p
o
u
t
0
.
1
,
an
d
v
alid
ated
af
ter
2
ep
o
ch
s
with
n
o
im
p
r
o
v
em
en
t.
Mo
d
el
p
er
f
o
r
m
a
n
ce
is
m
ea
s
u
r
ed
an
d
s
u
p
p
lem
en
te
d
with
tr
ac
k
in
g
u
s
in
g
ap
p
r
o
p
r
iate
m
etr
ics lik
e
m
ea
n
ab
s
o
lu
te
er
r
o
r
,
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
,
ac
cu
r
ac
y
an
d
co
n
f
u
s
io
n
m
atr
ices
.
T
h
is
s
etu
p
ac
h
iev
es
in
d
u
s
tr
y
-
s
tan
d
ar
d
r
eso
u
r
ce
u
tili
za
tio
n
an
d
a
d
ap
tiv
e
m
o
d
el
co
n
v
er
g
e
n
ce
f
o
r
m
u
ltimo
d
al
em
o
tio
n
an
al
y
s
is
.
4
.
5
.
Da
s
hb
o
a
rd
inte
g
ra
t
i
o
n a
nd
re
a
l
-
t
im
e
a
na
l
y
s
is
SME
D
o
f
f
er
s
la
y
u
s
er
s
an
d
m
en
tal
h
ea
lth
p
r
ac
titi
o
n
e
r
s
’
im
m
ed
iate
ac
ce
s
s
to
e
m
o
tio
n
al
t
r
en
d
s
.
T
h
e
d
ash
b
o
ar
d
d
is
p
lay
s
em
o
tio
n
al
d
ata
in
v
ar
io
u
s
ch
ar
ts
t
o
id
en
tify
em
o
tio
n
al
p
atter
n
s
an
d
n
ec
ess
ar
y
in
ter
v
en
tio
n
s
b
y
p
r
ac
titi
o
n
er
s
.
Fig
u
r
e
8
p
r
esen
ts
th
e
SME
D,
th
e
v
is
u
al
in
ter
f
ac
e
lay
er
.
I
t
o
f
f
er
s
r
ea
l
-
tim
e
v
is
u
aliza
tio
n
s
u
s
in
g
lin
e
g
r
a
p
h
s
,
em
o
tio
n
d
is
tr
ib
u
tio
n
ch
ar
ts
,
an
d
aler
t
s
y
s
tem
s
f
o
r
s
i
g
n
if
ican
t
em
o
tio
n
al
s
h
if
ts
.
T
h
e
d
ash
b
o
a
r
d
'
s
ar
ch
itectu
r
e
s
u
p
p
o
r
ts
r
ea
l
-
tim
e
u
p
d
ates,
m
ak
in
g
it
a
p
r
ac
tical
to
o
l
f
o
r
clin
ician
s
an
d
en
d
-
u
s
er
s
.
T
h
is
d
ash
b
o
ar
d
also
in
clu
d
es
aler
t
m
ec
h
an
is
m
s
t
h
at
m
ay
s
ig
n
al
s
ig
n
if
ican
t
em
o
tio
n
al
ch
an
g
es
o
n
s
o
m
e
p
r
ed
e
f
in
ed
th
r
esh
o
ld
s
.
F
o
r
ex
am
p
le,
s
u
p
p
o
s
e
th
e
u
s
er
'
s
em
o
tio
n
al
s
tate
ch
an
g
es
r
a
p
i
d
ly
.
I
f
th
ey
h
av
e
a
s
ig
n
if
ican
t
in
cr
ea
s
e
in
t
h
eir
a
n
x
iety
o
r
d
e
p
r
ess
io
n
lev
els,
t
h
e
s
y
s
tem
aler
ts
th
em
alm
o
s
t
im
m
ed
iately
s
o
th
at
th
ey
ca
n
s
ee
k
a
d
eq
u
ate
s
u
p
p
o
r
t o
r
in
ter
v
e
n
tio
n
in
g
o
o
d
tim
e.
Fig
u
r
e
8
.
SME
D
p
r
o
to
ty
p
e
5.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
wo
o
f
th
e
p
r
e
-
tr
ain
e
d
m
o
d
els
,
B
E
R
T
an
d
Dis
tilB
E
R
T
,
wer
e
em
p
lo
y
e
d
f
o
r
tex
t
d
ata
an
aly
s
is
,
an
d
an
L
STM
m
o
d
el
as
well
as
a
h
y
b
r
id
m
o
d
el
u
s
in
g
an
L
ST
M
ar
ch
itectu
r
e
with
B
E
R
T
em
b
ed
d
in
g
s
.
B
E
R
T
p
er
f
o
r
m
ed
o
u
ts
tan
d
in
g
s
em
an
tic
u
n
d
er
s
tan
d
in
g
an
d
co
n
tex
t
u
al
n
u
an
ce
s
o
n
t
h
e
Go
E
m
o
tio
n
s
d
ataset,
with
an
ac
cu
r
ac
y
o
f
8
7
.
4
%.
Dis
tilB
E
R
T
was
s
lig
h
tly
lo
wer
,
with
an
ac
cu
r
ac
y
o
f
8
5
.
9
%,
b
u
t
wit
h
less
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
.
T
h
e
L
STM
m
o
d
el
i
n
-
h
o
u
s
e
h
ad
r
o
b
u
s
t
s
eq
u
en
tial
p
r
o
ce
s
s
in
g
ca
p
ab
ilit
ies,
ac
h
iev
in
g
an
ac
cu
r
ac
y
o
f
8
3
.
6
%,
b
u
t
was
n
o
t
ca
p
ab
le
o
f
d
etec
tin
g
co
m
p
le
x
s
em
an
tic
r
elatio
n
s
h
ip
s
.
B
len
d
in
g
s
eq
u
e
n
ce
m
o
d
ellin
g
with
p
r
e
-
tr
ain
ed
em
b
e
d
d
in
g
s
,
th
e
h
y
b
r
id
ar
ch
itectu
r
e
o
f
B
E
R
T
em
b
ed
d
in
g
s
co
u
p
led
with
a
n
L
STM
,
en
h
an
ce
d
p
er
f
o
r
m
an
ce
,
lead
i
n
g
to
a
f
in
a
l a
cc
u
r
ac
y
o
f
8
8
.
2
%.
R
esNet
-
5
0
an
d
VGG
-
1
6
wer
e
u
s
ed
to
tr
ain
th
e
d
ataset,
an
d
a
C
NN
lay
er
an
d
a
cu
s
to
m
C
NN
m
o
d
el
wer
e
ex
p
er
im
e
n
ted
with
.
VG
G
-
1
6
,
with
its
p
lain
a
r
ch
itectu
r
e,
f
ar
e
d
p
o
o
r
ly
with
8
2
.
9
%
a
cc
u
r
ac
y
o
n
th
e
FER
d
ataset,
with
R
esNet
-
5
0
,
at
8
4
.
2
%.
T
h
e
c
u
s
to
m
C
NN
o
p
tim
ally
weig
h
ed
f
ea
t
u
r
e
ex
tr
ac
t
io
n
an
d
e
f
f
icien
cy
at
8
1
.
5
%.
T
h
e
h
y
b
r
id
R
esNet
-
5
0
+CNN
co
n
f
ig
u
r
atio
n
,
b
ased
o
n
p
r
e
-
tr
ain
ed
f
ea
tu
r
es
an
d
d
o
m
ain
ad
ap
tatio
n
,
y
ield
ed
8
5
.
1
%.
A
late
f
u
s
io
n
m
eth
o
d
d
y
n
am
ically
co
m
b
in
es
m
u
ltip
le
m
o
d
els
an
d
ac
h
iev
es
b
ette
r
p
er
f
o
r
m
an
ce
.
C
o
n
tex
tu
al
atten
tio
n
n
etwo
r
k
s
b
alan
ce
m
o
d
alities
an
d
s
et
tle
d
is
p
u
tes
an
d
r
ep
o
r
t
a
p
r
ec
is
io
n
o
f
8
9
.
7
%,
h
ig
h
er
th
an
ea
r
ly
a
n
d
in
d
iv
id
u
al
f
u
s
io
n
m
o
d
els
(
7
8
.
4
%).
T
h
e
p
e
r
f
o
r
m
an
ce
m
etr
ics
f
o
r
ea
ch
co
n
f
ig
u
r
atio
n
o
n
th
e
FER
an
d
Go
E
m
o
tio
n
s
d
atasets
ar
e
co
m
p
iled
in
T
ab
le
2
,
an
d
th
e
g
r
a
p
h
ical
r
ep
r
esen
tatio
n
is
in
Fig
u
r
e
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
5
6
5
-
4
5
7
8
4574
T
ab
le
2
.
Mo
d
el
p
er
f
o
r
m
a
n
c
e
o
n
Go
E
m
o
tio
n
s
an
d
FER
d
atasets
M
o
d
e
l
M
o
d
a
l
i
t
y
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
D
a
t
a
s
e
t
Pre
-
t
r
a
i
n
e
d
B
ER
T
Te
x
t
8
7
.
4
0
.
8
5
0
.
6
0
.
8
5
G
o
Em
o
t
i
o
n
s
Pre
-
t
r
a
i
n
e
d
D
i
st
i
l
B
ER
T
Te
x
t
8
5
.
9
0
.
8
3
0
.
5
0
.
8
4
G
o
Em
o
t
i
o
n
s
C
u
s
t
o
m LST
M
Te
x
t
8
3
.
6
0
.
8
1
0
.
2
0
.
8
1
G
o
Em
o
t
i
o
n
s
H
y
b
r
i
d
B
E
R
T+
LST
M
Te
x
t
8
8
.
2
0
.
8
7
0
.
8
0
.
8
7
G
o
Em
o
t
i
o
n
s
Pre
-
t
r
a
i
n
e
d
R
e
sN
e
t
-
50
I
mag
e
8
4
.
2
0
.
8
2
0
.
3
0
.
8
2
F
ER
Pre
-
t
r
a
i
n
e
d
V
G
G
1
6
I
mag
e
8
2
.
9
0
.
8
0
.
1
0
.
8
F
ER
C
u
s
t
o
m C
N
N
I
mag
e
8
1
.
5
0
.
7
8
0
.
7
9
0
.
7
9
F
ER
H
y
b
r
i
d
R
e
sN
e
t
-
5
0
+
C
N
N
I
mag
e
8
5
.
1
0
.
8
4
0
.
8
5
0
.
8
4
F
ER
Ea
r
l
y
f
u
s
i
o
n
M
u
l
t
i
m
o
d
a
l
7
8
.
4
0
.
7
7
0
.
7
6
0
.
7
6
F
u
se
d
D
a
t
a
s
e
t
:
G
o
Em
o
t
i
o
n
s
a
n
d
F
E
R
C
u
s
t
o
m
l
a
t
e
f
u
s
i
o
n
(
P
r
o
p
o
s
e
d
)
M
u
l
t
i
m
o
d
a
l
8
9
.
7
0
.
8
8
0
.
9
0
.
8
9
F
u
se
d
D
a
t
a
s
e
t
:
G
o
Em
o
t
i
o
n
s
a
n
d
F
E
R
Fig
u
r
e
9
.
C
o
m
p
a
r
ativ
e
p
e
r
f
o
r
m
an
ce
o
f
m
o
d
els o
n
Go
E
m
o
ti
o
n
s
an
d
FER
d
atasets
T
o
ass
es
s
g
en
er
aliza
b
ilit
y
,
th
e
f
r
am
ewo
r
k
was
ev
alu
ated
o
n
a
s
ec
o
n
d
d
ataset,
in
clu
d
in
g
Af
f
ec
tNet
f
o
r
im
ag
es
an
d
Go
E
m
o
tio
n
s
f
o
r
tex
t.
C
o
n
s
is
ten
t
p
atter
n
s
em
er
g
ed
in
th
e
r
esu
lts
,
with
th
e
late
f
u
s
io
n
m
o
d
el
co
n
tin
u
in
g
to
p
er
f
o
r
m
b
etter
.
T
h
e
ass
ess
m
en
t m
ea
s
u
r
es f
o
r
th
is
d
ataset
ar
e
s
h
o
wn
in
T
ab
le
3
,
an
d
its
g
r
ap
h
ical
r
ep
r
esen
tatio
n
is
in
Fig
u
r
e
1
0
.
T
ab
le
3
.
Mo
d
el
p
er
f
o
r
m
a
n
ce
o
n
Go
E
m
o
tio
n
s
an
d
Af
f
ec
tNet
d
atasets
M
o
d
e
l
M
o
d
a
l
i
t
y
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
S
c
o
r
e
D
a
t
a
s
e
t
Pre
-
t
r
a
i
n
e
d
B
ER
T
Te
x
t
8
7
.
4
0
.
8
5
0
.
8
6
0
.
8
5
G
o
Em
o
t
i
o
n
s
Pre
-
t
r
a
i
n
e
d
D
i
st
i
l
B
ER
T
Te
x
t
8
5
.
9
0
.
8
3
0
.
8
5
0
.
8
4
G
o
Em
o
t
i
o
n
s
C
u
s
t
o
m LST
M
Te
x
t
8
3
.
6
0
.
8
1
0
.
8
2
0
.
8
1
G
o
Em
o
t
i
o
n
s
H
y
b
r
i
d
B
E
R
T+
LST
M
Te
x
t
8
8
.
2
0
.
8
7
0
.
8
8
0
.
8
7
G
o
Em
o
t
i
o
n
s
Pre
-
t
r
a
i
n
e
d
R
e
sN
e
t
-
50
I
mag
e
8
2
.
3
0
.
8
1
0
.
8
2
0
.
8
1
A
f
f
e
c
t
N
e
t
Pre
-
t
r
a
i
n
e
d
V
G
G
-
16
I
mag
e
8
1
.
2
0
.
7
9
0
.
8
0
.
7
9
A
f
f
e
c
t
N
e
t
C
u
s
t
o
m C
N
N
I
mag
e
7
9
.
8
0
.
7
6
0
.
7
7
0
.
7
7
A
f
f
e
c
t
N
e
t
H
y
b
r
i
d
R
e
sN
e
t
-
5
0
+
C
N
N
I
mag
e
8
3
.
4
0
.
8
2
0
.
8
3
0
.
8
2
A
f
f
e
c
t
N
e
t
Ea
r
l
y
f
u
s
i
o
n
M
u
l
t
i
m
o
d
a
l
7
6
.
5
0
.
7
5
0
.
7
4
0
.
7
5
F
u
se
d
D
a
t
a
s
e
t
:
G
o
Em
o
t
i
o
n
s
a
n
d
A
f
f
e
c
t
N
e
t
C
u
s
t
o
m
l
a
t
e
f
u
s
i
o
n
(
P
r
o
p
o
se
d
)
M
u
l
t
i
m
o
d
a
l
8
8
.
9
0
.
8
7
0
.
8
8
0
.
8
7
F
u
se
d
D
a
t
a
s
e
t
:
G
o
Em
o
t
i
o
n
s
a
n
d
A
f
f
e
c
t
N
e
t
SME
D
is
also
u
s
ef
u
l
in
in
te
r
p
r
etin
g
th
ese
r
esu
lts
.
T
h
e
la
te
f
u
s
io
n
m
o
d
el
allo
ws
an
a
ctio
n
ab
le,
r
elev
an
t,
an
d
ac
cu
r
ate
f
o
r
ec
a
s
t
b
y
r
em
o
v
in
g
tech
n
ical
o
u
tp
u
t
-
v
er
s
u
s
-
r
ea
l
-
wo
r
ld
-
ap
p
licatio
n
s
m
is
m
atch
es.
I
t
r
eg
is
ter
s
f
ea
tu
r
es
b
etwe
en
m
o
d
alities
d
y
n
am
ically
,
b
o
o
s
t
ed
b
y
p
r
e
-
tr
ain
e
d
an
d
task
-
s
p
ec
if
ic
m
o
d
els,
an
d
p
er
f
o
r
m
s
well
co
n
s
is
ten
tly
.
T
h
e
m
eth
o
d
g
ath
er
s
c
o
m
p
lem
e
n
tar
y
a
n
d
co
n
tex
t
u
ally
r
elate
d
in
f
o
r
m
atio
n
f
r
o
m
Evaluation Warning : The document was created with Spire.PDF for Python.