I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
3
9
,
No
.
3
,
Sep
tem
b
er
2
0
2
5
,
p
p
.
1
8
1
5
~
1
825
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
3
9
.i
3
.
pp
1
8
1
5
-
1
8
2
5
1815
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Cha
ra
cter
iza
tion
o
f
bina
rized
neur
a
l net
wo
rks
for e
f
ficient
deplo
y
ment on r
e
so
urce
-
limite
d ed
g
e devices
Ra
m
y
a
B
a
na
v
a
ra
N
a
ra
y
a
na
1
,
Seem
a
Sin
g
h
2
1
D
e
p
a
r
t
me
n
t
o
f
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
M
a
c
h
i
n
e
L
e
a
r
n
i
n
g
,
J
y
o
t
h
y
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
B
a
n
g
a
l
o
r
e
,
A
f
f
i
l
i
a
t
e
d
t
o
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s
a
n
d
T
e
l
e
c
o
mm
u
n
i
c
a
t
i
o
n
E
n
g
i
n
e
e
r
i
n
g
,
B
M
S
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
a
n
d
M
a
n
a
g
e
m
e
n
t
,
B
a
n
g
a
l
o
r
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
Dec
12
,
2
0
2
4
R
ev
is
ed
Ap
r
7
,
2
0
2
5
Acc
ep
ted
J
u
l
2
,
2
0
2
5
Th
is
p
a
p
e
r
d
e
lv
e
s
in
to
b
in
a
riz
e
d
n
e
u
ra
l
n
e
tw
o
rk
s
(BNN
s)
ta
il
o
re
d
f
o
r
re
so
u
rc
e
-
c
o
n
stra
in
e
d
e
d
g
e
d
e
v
i
c
e
s.
BNN
s
h
a
rn
e
ss
b
in
a
r
y
we
ig
h
ts
a
n
d
a
c
ti
v
a
ti
o
n
s
to
a
m
p
li
f
y
e
fficie
n
c
y
wh
il
e
u
p
h
o
ld
i
n
g
a
c
c
u
ra
c
y
.
Ac
ro
ss
d
iv
e
rse
n
e
two
rk
c
o
n
fi
g
u
ra
ti
o
n
s,
BNN
s
c
o
n
siste
n
tl
y
o
u
tsh
in
e
trad
i
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
(NN
s).
A
p
i
o
n
e
e
rin
g
BNN
a
rc
h
it
e
c
tu
re
is
d
e
v
e
l
o
p
e
d
in
LARQ,
a
c
h
iev
in
g
a
n
imp
re
ss
iv
e
6
1
%
a
c
c
u
ra
c
y
o
n
t
h
e
M
NIST
d
a
tas
e
t
th
ro
u
g
h
b
in
a
r
y
q
u
a
n
t
iza
ti
o
n
,
we
ig
h
t
c
li
p
p
i
n
g
,
a
n
d
p
o
i
n
twise
c
o
n
v
o
lu
ti
o
n
s.
Im
p
le
m
e
n
tatio
n
o
n
t
h
e
Xili
n
x
P
YN
QZ2
F
P
G
A
b
o
a
rd
sh
o
ws
fa
r
q
u
ick
e
r
c
las
sifica
ti
o
n
ra
tes
,
with
a
m
a
x
imu
m
in
fe
re
n
c
e
ti
m
e
o
f
0
.
0
0
8
4
1
m
il
li
se
c
o
n
d
s
p
e
r
ima
g
e
,
a
p
p
ro
x
ima
tely
1
0
,
0
0
0
ima
g
e
s
b
e
in
g
c
las
sified
i
n
th
is
le
n
g
t
h
o
f
t
ime
.
Th
e
ti
m
e
tak
e
n
p
e
r
ima
g
e
re
p
re
se
n
ts
a
p
p
ro
x
ima
tely
0
.
0
1
%
o
f
t
h
e
t
o
tal
in
fe
re
n
c
e
ti
m
e
.
Th
is u
n
d
e
rsc
o
re
s BNN
s'
p
o
ten
ti
a
l
to
re
d
e
fin
e
re
a
l
-
ti
m
e
e
d
g
e
c
o
m
p
u
ti
n
g
a
p
p
li
c
a
ti
o
n
s.
Th
e
p
a
p
e
r
m
a
k
e
s
sig
n
ifi
c
a
n
t
strid
e
s
b
y
e
lu
c
i
d
a
ti
n
g
BNN
s'
p
e
rfo
rm
a
n
c
e
su
p
e
rio
ri
ty
,
p
ro
p
o
sin
g
a
n
in
n
o
v
a
ti
v
e
a
rc
h
it
e
c
t
u
re
,
a
n
d
v
a
li
d
a
ti
n
g
it
s
p
ro
we
ss
th
r
o
u
g
h
re
a
l
-
wo
rld
d
e
p
lo
y
m
e
n
t.
T
h
e
se
fin
d
i
n
g
s
u
n
d
e
rsc
o
r
e
BNN
s
a
s
a
g
il
e
,
h
ig
h
-
p
e
rfo
rm
a
n
c
e
m
o
d
e
ls
p
rime
d
fo
r
e
d
g
e
c
o
m
p
u
ti
n
g
,
fo
ste
ri
n
g
a
n
e
w era
o
f
re
a
l
-
ti
m
e
p
ro
c
e
ss
in
g
i
n
n
o
v
a
ti
o
n
s
.
K
ey
w
o
r
d
s
:
Acc
u
r
ac
y
B
in
ar
ized
n
eu
r
al
n
etwo
r
k
FP
GA
Neu
r
al
n
etwo
r
k
Xilin
x
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
:
R
am
y
a
B
an
av
ar
a
Nar
ay
an
a
Dep
ar
tm
en
t o
f
Ar
tific
ial
I
n
tellig
en
ce
an
d
Ma
ch
in
e
L
ea
r
n
in
g
,
J
y
o
th
y
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
B
an
g
alo
r
e
,
Af
f
iliated
to
Vis
v
e
s
v
ar
ay
a
T
ec
h
n
o
lo
g
ical
Un
iv
e
r
s
ity
B
elag
av
i
-
5
9
0
0
1
8
,
Kar
n
atak
a,
I
n
d
ia
E
m
ail:
r
am
y
a.
b
n
wo
r
k
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
(
C
NN
s
)
ar
e
h
ig
h
ly
ef
f
ec
tiv
e
f
o
r
co
m
p
u
ter
v
is
io
n
task
s
b
u
t
co
m
e
with
s
ig
n
if
ican
t
co
m
p
u
tatio
n
al
d
e
m
an
d
s
d
u
e
to
th
eir
lar
g
e
n
u
m
b
er
o
f
p
a
r
am
eter
s
an
d
r
elia
n
ce
o
n
f
lo
atin
g
-
p
o
i
n
t
o
p
er
atio
n
s
[
1
]
.
T
h
is
m
ak
es
t
h
em
r
eso
u
r
ce
-
in
ten
s
iv
e,
r
eq
u
i
r
in
g
s
u
b
s
tan
tial
p
r
o
ce
s
s
in
g
s
p
ee
d
an
d
m
em
o
r
y
,
wh
ich
lim
its
th
eir
d
ep
lo
y
m
en
t
to
h
ig
h
-
p
er
f
o
r
m
an
ce
co
m
p
u
ti
n
g
p
latf
o
r
m
s
lik
e
ce
n
tr
al
p
r
o
c
ess
in
g
u
n
it
(
C
PU)
o
r
g
r
ap
h
ics p
r
o
ce
s
s
in
g
u
n
it (
G
PU)
[
2
]
.
Field
p
r
o
g
r
am
m
ab
le
g
ate
ar
r
ay
s
(
FP
GAs)
o
f
f
er
a
p
r
o
m
is
in
g
alter
n
ativ
e
b
y
p
r
o
v
i
d
in
g
c
u
s
to
m
h
ar
d
w
ar
e
ac
ce
ler
ato
r
s
th
at
en
h
a
n
c
e
co
m
p
u
tatio
n
al
ef
f
icien
cy
a
n
d
en
e
r
g
y
s
av
in
g
s
th
r
o
u
g
h
p
ar
allelis
m
an
d
r
ec
o
n
f
ig
u
r
a
b
ilit
y
[
3
]
.
T
r
ad
itio
n
al
co
m
p
u
tin
g
ar
ch
itectu
r
es
h
a
v
e
lim
itatio
n
s
wh
en
it
co
m
e
s
to
ad
d
r
ess
in
g
th
e
co
m
p
u
tatio
n
al
n
ee
d
s
an
d
ef
f
icien
cy
r
eq
u
ir
e
m
en
ts
o
f
co
n
tem
p
o
r
ar
y
d
ee
p
lea
r
n
in
g
ap
p
licatio
n
s
[
4
]
.
GPUs
h
av
e
ex
ten
d
ed
te
n
s
o
r
-
b
ased
d
ee
p
l
ea
r
n
in
g
m
o
d
els
b
y
(
i)
ac
ce
le
r
atin
g
d
ee
p
n
eu
r
al
n
etwo
r
k
tr
ain
in
g
an
d
test
in
g
a
s
well;
y
et
GPUs
s
till
f
a
ll
s
h
o
r
t
wh
en
it
co
m
es
to
h
an
d
lin
g
v
er
y
c
o
m
p
lex
DNNs
d
u
e
to
tig
h
t
m
em
o
r
y
n
ee
d
s
,
l
o
n
g
tr
an
s
f
er
tim
es
am
o
n
g
m
e
m
o
r
y
h
ier
ar
c
h
y
p
ar
ts
,
an
d
lim
ited
p
ar
allelis
m
o
n
ce
r
tain
ty
p
es
o
f
co
n
v
o
lu
tio
n
al/
d
ee
p
r
esid
u
al
n
etwo
r
k
s
.
A
C
PUs
,
wh
ich
was
d
es
ig
n
ed
to
b
e
g
en
er
al
-
p
u
r
p
o
s
e
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
3
,
Sep
tem
b
er
20
25
:
1
8
1
5
-
1
8
2
5
1816
n
atu
r
e
ar
e
n
o
t
o
p
tim
ized
f
o
r
e
f
f
icien
tly
r
esp
o
n
d
in
g
to
th
e
m
ass
iv
e
p
ar
allelis
m
ass
o
ciate
d
with
d
ee
p
lear
n
in
g
wo
r
k
lo
ad
s
d
u
e
t
o
th
is
r
ea
s
o
n
th
eir
p
er
f
o
r
m
a
n
ce
lag
s
as
co
m
p
ar
ed
to
s
p
ec
ialized
h
ar
d
wa
r
e.
Mo
r
e
o
v
er
,
th
ese
ar
ch
itectu
r
es
ar
e
ty
p
ically
co
n
s
tr
ain
ed
b
y
en
er
g
y
ef
f
icien
cy
r
eq
u
ir
em
e
n
ts
f
o
r
th
e
tr
ain
in
g
o
f
lar
g
e
m
o
d
els
an
d
th
er
ef
o
r
e
m
ay
n
o
t b
e
s
u
itab
le
f
o
r
h
ig
h
-
p
er
f
o
r
m
a
n
ce
d
ee
p
lear
n
in
g
at
a
s
ca
le.
As d
ee
p
-
lear
n
i
n
g
m
o
d
els b
ec
o
m
e
lar
g
er
an
d
m
o
r
e
in
tr
icate
,
h
o
wev
er
,
s
p
ec
ialized
h
ar
d
war
e
b
ec
o
m
es
in
cr
ea
s
in
g
ly
n
ec
ess
ar
y
-
FP
GAs
o
r
cu
s
to
m
ar
tific
ial
in
tellig
en
ce
(
AI
)
ac
c
eler
ato
r
s
lik
e
Go
o
g
le'
s
T
PUs
o
f
f
er
a
n
ar
c
h
itectu
r
e
to
p
r
o
v
id
e
b
etter
th
r
o
u
g
h
p
u
t
f
o
r
m
em
o
r
y
-
b
o
u
n
d
ap
p
licatio
n
s
,
r
ed
u
ce
laten
cy
,
as
well
as
b
ein
g
f
ar
f
aster
at
m
atr
ix
-
h
e
av
y
o
p
er
atio
n
s
th
at
ca
n
u
n
d
er
p
in
m
an
y
o
f
th
ese
ta
s
k
s
.
T
h
ese
ar
e
d
esig
n
ed
s
p
ec
if
ically
f
o
r
d
ee
p
lear
n
in
g
task
s
,
b
u
t
th
ey
to
o
h
av
e
ch
allen
g
es
in
m
ee
tin
g
th
e
f
le
x
ib
ilit
y
,
p
r
o
g
r
am
m
ab
ilit
y
a
n
d
s
ca
lab
ilit
y
n
ee
d
s
o
f
f
ast
-
ev
o
lv
i
n
g
n
aïv
e
n
etwo
r
k
s
.
T
h
ese
co
n
s
tr
ain
ts
h
av
e
b
ee
n
b
r
o
u
g
h
t
to
li
g
h
t
b
y
th
e
g
r
o
win
g
d
em
an
d
f
o
r
AI
-
b
ased
s
o
lu
tio
n
s
.
T
h
ese
co
n
v
en
tio
n
al
m
o
d
els
f
r
eq
u
en
tl
y
s
u
f
f
er
f
r
o
m
lo
w
ac
cu
r
ac
y
a
n
d
lar
g
e
in
f
er
e
n
ce
tim
es,
m
ak
in
g
th
em
in
ef
f
icien
t
f
o
r
n
eu
r
al
n
etwo
r
k
class
if
icatio
n
,
esp
ec
ially
o
n
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
ed
g
e
d
e
v
ices
[
5
]
.
B
in
ar
ized
n
eu
r
a
l
n
etwo
r
k
s
(
B
NNs)
ad
d
r
ess
th
ese
ch
allen
g
es
b
y
u
s
in
g
b
in
a
r
y
weig
h
ts
an
d
ac
tiv
atio
n
s
,
d
r
asti
ca
lly
r
ed
u
cin
g
m
em
o
r
y
u
s
ag
e
a
n
d
s
im
p
lify
in
g
co
m
p
u
tatio
n
s
[
6
]
-
[
1
0
]
.
T
h
is
r
esu
lts
in
f
aster
o
p
er
atio
n
s
an
d
g
r
ea
ter
ef
f
icien
c
y
,
p
ar
ticu
lar
ly
s
u
ited
f
o
r
FP
GA
im
p
lem
en
tatio
n
s
[
1
1
]
.
Ad
v
an
ce
d
tech
n
i
q
u
es
s
u
ch
as
L
AR
Q
im
p
r
o
v
e
th
e
p
er
f
o
r
m
an
ce
o
f
B
NNs
b
y
o
p
tim
izin
g
ac
tiv
atio
n
q
u
an
tizatio
n
,
b
ala
n
cin
g
ac
c
u
r
ac
y
an
d
e
f
f
icien
cy
.
Pra
ctica
l
v
alid
atio
n
s
h
av
e
s
h
o
wn
t
h
at
B
NNs
ca
n
o
u
tp
er
f
o
r
m
tr
a
d
itio
n
al
n
eu
r
al
n
etwo
r
k
s
in
r
ea
l
-
tim
e
an
d
ed
g
e
co
m
p
u
tin
g
s
ce
n
ar
io
s
,
m
ak
i
n
g
t
h
em
a
v
iab
le
s
o
lu
tio
n
f
o
r
d
e
p
lo
y
in
g
AI
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
en
v
ir
o
n
m
en
ts
.
I
n
r
e
ce
n
t
s
tate
-
of
-
th
e
-
ar
t
r
esea
r
ch
,
m
an
y
r
elev
a
n
t
s
tu
d
ies
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
t
o
ac
h
iev
e
a
b
alan
ce
d
t
r
a
d
e
-
o
f
f
b
e
t
w
e
e
n
m
o
d
e
l
a
c
c
u
r
a
c
y
a
n
d
c
o
m
p
u
t
a
t
i
o
n
a
l
e
f
f
i
c
ie
n
c
y
.
T
h
e
r
e
s
e
a
r
c
h
w
o
r
k
c
a
r
r
i
e
d
o
u
t
b
y
P
h
i
p
p
s
e
t
a
l
.
[
1
2
]
f
o
cu
s
es
o
n
im
p
r
o
v
i
n
g
t
h
e
ef
f
icien
cy
o
f
8
la
y
er
s
o
f
B
NNs
u
s
in
g
th
e
L
AR
Q
p
latf
o
r
m
o
f
f
er
in
g
b
o
th
ac
cu
r
ac
y
a
n
d
ef
f
icien
c
y
.
T
h
e
wo
r
k
ca
r
r
ied
o
u
t
b
y
Sak
r
et
a
l.
[
1
3
]
f
o
cu
s
es
o
n
d
e
v
elo
p
i
n
g
a
B
NN
f
r
am
ewo
r
k
u
s
in
g
L
AR
Q
an
d
th
e
GC
C
c
o
m
p
iler
with
in
STM
3
2
C
u
b
eI
DE
,
tar
g
etin
g
AR
M
C
o
r
tex
-
M
m
icr
o
co
n
tr
o
ller
s
.
L
ee
et
a
l.
[
1
4
]
h
a
v
e
in
tr
o
d
u
ce
d
a
f
r
am
ewo
r
k
d
esig
n
ed
to
o
p
tim
ize
f
u
lly
h
o
m
o
m
o
r
p
h
ic
en
cr
y
p
tio
n
(
FHE)
f
o
r
p
r
iv
ate
m
ac
h
i
n
e
lear
n
in
g
in
f
e
r
en
ce
u
s
i
n
g
ter
n
ar
y
n
eu
r
al
n
et
wo
r
k
s
.
Z
h
an
g
et
a
l.
[
1
5
]
d
e
p
lo
y
ed
a
B
NN
o
n
t
h
e
Xilin
x
Z
YNQ
b
o
ar
d
,
f
o
cu
s
in
g
o
n
ac
h
iev
in
g
h
ig
h
p
e
r
f
o
r
m
an
ce
in
ed
g
e
ap
p
licatio
n
s
.
I
n
th
e
s
tu
d
y
m
o
d
el
o
f
Salau
y
o
u
[
1
6
]
,
p
r
esen
ts
a
n
o
v
el
alg
o
r
ith
m
aim
ed
at
o
p
tim
izi
n
g
th
e
ar
ea
a
n
d
d
ep
th
o
f
FP
GA
d
esig
n
s
f
o
cu
s
in
g
o
n
ef
f
icien
tly
m
a
p
p
in
g
lo
g
ic
f
u
n
ctio
n
s
to
FP
GA
r
eso
u
r
ce
s
wh
ile
m
in
im
izin
g
th
e
cr
itical
p
ath
d
ep
th
.
An
o
th
er
s
tu
d
y
b
y
Z
h
u
et
a
l.
[
1
7
]
p
r
esen
ted
a
tech
n
iq
u
e
ca
lled
B
NN
-
Do
R
eFa,
wh
ich
u
s
es
t
er
n
ar
y
weig
h
ts
an
d
ac
tiv
atio
n
s
with
d
y
n
am
ic
r
an
g
e
s
ca
lin
g
f
ac
to
r
s
,
ac
h
iev
in
g
s
im
ilar
ac
cu
r
ac
y
t
o
tr
a
d
itio
n
al
n
eu
r
al
n
etwo
r
k
s
wh
ile
r
ed
u
cin
g
th
e
m
e
m
o
r
y
f
o
o
tp
r
in
t
a
n
d
c
o
m
p
u
tatio
n
r
eq
u
ir
em
en
ts
.
L
u
o
et
a
l.
[
1
8
]
h
a
v
e
in
tr
o
d
u
ce
d
t
h
e
co
n
ce
p
t
o
f
B
in
ar
y
Dilated
Den
s
eNe
t,
wh
ich
is
a
n
eu
r
al
n
etw
o
r
k
m
o
d
el
d
esig
n
ed
f
o
r
e
f
f
ici
en
t
h
u
m
an
ac
tiv
it
y
r
ec
o
g
n
itio
n
(
HAR)
at
t
h
e
n
etwo
r
k
ed
g
e
u
s
in
g
n
etwo
r
k
b
in
ar
izatio
n
f
o
r
m
em
o
r
y
u
s
ag
e
an
d
r
ed
u
cin
g
co
m
p
u
tatio
n
al
co
m
p
lex
ity
.
S
hi
et
a
l.
[
1
9
]
f
o
cu
s
es
o
n
en
h
an
cin
g
th
e
p
er
f
o
r
m
a
n
ce
o
f
b
in
ar
ized
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
B
C
NNs)
th
r
o
u
g
h
th
e
u
s
e
o
f
d
y
n
am
ic
p
ar
ti
al
r
ec
o
n
f
i
g
u
r
atio
n
o
n
d
is
ag
g
r
e
g
ated
FP
GAs
with
s
ig
n
if
ican
t
r
ed
u
ctio
n
o
f
in
f
er
en
ce
tim
e
an
d
en
er
g
y
co
n
s
u
m
p
tio
n
co
m
p
ar
ed
t
o
tr
ad
itio
n
al
f
ix
ed
im
p
lem
en
tatio
n
s
.
A
s
tu
d
y
b
y
L
i
et
a
l.
[
2
0
]
p
r
esen
ted
a
tech
n
iq
u
e
ca
lled
q
u
a
n
tized
C
NN,
wh
ic
h
u
s
es
b
in
ar
y
weig
h
ts
an
d
ac
tiv
atio
n
s
with
q
u
an
tized
s
ca
lin
g
f
ac
t
o
r
s
,
ac
h
iev
in
g
c
o
m
p
ar
a
b
le
ac
cu
r
ac
y
to
tr
ad
itio
n
al
n
e
u
r
a
l
n
etwo
r
k
s
wh
ile
r
e
d
u
cin
g
th
e
m
em
o
r
y
f
o
o
t
p
r
in
t
a
n
d
c
o
m
p
u
t
atio
n
r
eq
u
ir
em
en
ts
.
Gu
id
o
tti
e
t
a
l.
[
2
1
]
p
r
esen
ted
a
n
a
p
p
r
o
a
ch
f
o
r
v
e
r
if
y
in
g
B
N
N
s
u
s
i
n
g
s
a
t
i
s
f
i
ab
i
l
i
t
y
m
o
d
u
l
o
th
e
o
r
i
e
s
(
S
M
T
)
t
h
a
t
in
v
o
l
v
e
s
en
c
o
d
i
n
g
t
h
e
v
e
r
if
i
c
a
t
i
o
n
p
r
o
b
l
em
i
n
to
SM
T
f
o
r
m
u
l
a
s
an
d
le
v
er
a
g
in
g
t
h
e
p
o
w
er
o
f
S
M
T
s
o
l
v
e
r
s
t
o
p
r
o
v
e
p
r
o
p
er
t
i
e
s
a
b
o
u
t
B
N
N
s
.
Fro
m
th
e
v
iewp
o
in
t
o
f
t
h
e
id
en
tifie
d
r
esear
ch
p
r
o
b
lem
,
it
i
s
n
o
ted
th
at
p
ast
r
esear
ch
o
n
B
NNs
h
a
s
laid
a
s
o
lid
f
o
u
n
d
atio
n
f
o
r
im
p
lem
en
tin
g
ef
f
icien
t
n
eu
r
al
n
etwo
r
k
s
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
en
v
ir
o
n
m
en
ts
,
p
ar
ticu
lar
ly
f
o
r
ed
g
e
d
e
v
ices.
Ho
wev
er
,
th
e
tr
ad
e
-
o
f
f
b
etwe
e
n
ac
cu
r
ac
y
an
d
co
m
p
u
tatio
n
al
ef
f
icien
cy
r
em
ain
s
an
o
p
en
p
r
o
b
lem
.
Desp
ite
th
e
p
r
o
m
is
in
g
r
esu
lts
,
th
er
e
a
r
e
s
till
s
ev
er
al
ch
allen
g
es
th
at
n
ee
d
to
b
e
ad
d
r
ess
ed
in
th
e
o
p
tim
izatio
n
o
f
B
NNs
f
o
r
h
ar
d
war
e
im
p
lem
en
tatio
n
o
n
th
e
PYNQ
Z
2
b
o
ar
d
.
T
h
er
e
f
o
r
e,
f
u
r
t
h
er
r
esear
ch
is
n
ee
d
ed
to
d
ev
elo
p
n
ew
ap
p
r
o
x
im
atio
n
tech
n
iq
u
es
th
at
ca
n
ac
h
iev
e
b
etter
ac
cu
r
ac
y
wh
ile
m
ain
tain
in
g
th
e
b
en
ef
its
o
f
B
NNs.
I
n
s
u
m
m
ar
y
,
B
NNs
h
as
b
ee
n
a
m
ajo
r
r
es
ea
r
ch
a
r
ea
in
r
ec
e
n
t
y
ea
r
s
.
Se
v
er
al
ap
p
r
o
x
im
atio
n
tech
n
iq
u
es
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
to
o
v
e
r
co
m
e
th
e
lo
s
s
o
f
a
cc
u
r
ac
y
ass
o
ciate
d
with
u
s
in
g
b
in
ar
y
v
alu
es,
an
d
v
ar
i
o
u
s
tech
n
i
q
u
es h
av
e
b
ee
n
p
r
o
p
o
s
ed
to
o
p
tim
ize
th
e
p
er
f
o
r
m
an
ce
o
f
B
NNs o
n
th
e
PYNQ
Z
2
b
o
ar
d
.
T
h
e
p
r
im
e
aim
o
f
th
e
p
r
o
p
o
s
ed
s
tu
d
y
is
to
d
esig
n
a
n
o
v
el
B
NN
m
o
d
el
to
war
d
s
ass
is
tin
g
th
e
task
p
er
tain
in
g
to
p
atter
n
r
ec
o
g
n
it
io
n
o
v
er
a
s
tan
d
ar
d
FP
GA
b
o
ar
d
.
T
h
is
p
ap
e
r
aim
s
to
an
aly
ze
th
e
im
p
ac
t
o
f
v
ar
io
u
s
lay
er
ar
ch
itectu
r
es
a
n
d
n
e
u
r
o
n
d
en
s
ities
o
n
B
N
N
p
er
f
o
r
m
an
ce
,
p
r
o
v
id
in
g
c
r
u
cial
in
s
ig
h
ts
f
o
r
d
esig
n
in
g
e
f
f
icien
t
B
NNs
f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
d
g
e
d
e
v
ices.
B
y
b
alan
cin
g
ef
f
icien
c
y
an
d
ac
cu
r
ac
y
,
we
s
ee
k
to
d
em
o
n
s
tr
ate
t
h
e
r
ele
v
an
ce
o
f
B
NNs
f
o
r
p
atter
n
r
ec
o
g
n
itio
n
task
s
o
n
p
latf
o
r
m
s
li
k
e
th
e
PYNQ
Z
2
FP
GA
B
o
ar
d
.
T
h
e
v
alu
e
-
ad
d
e
d
co
n
tr
ib
u
tio
n
o
f
th
e
s
tu
d
y
a
r
e:
i)
to
p
er
f
o
r
m
class
if
icatio
n
o
f
an
im
ag
e
u
s
in
g
d
ee
p
lear
n
in
g
m
o
d
el
o
v
e
r
a
r
ea
l
-
tim
e
h
ar
d
war
e
p
latf
o
r
m
,
ii)
to
d
e
v
e
lo
p
a
s
im
p
lifi
ed
an
d
p
r
o
g
r
ess
iv
e
s
eq
u
en
tial
m
o
d
ellin
g
to
war
d
s
ac
co
m
p
lis
h
in
g
b
etter
tr
ain
in
g
s
tab
ilit
y
,
an
d
iii)
to
ca
r
r
y
o
u
t
co
m
p
r
e
h
en
s
iv
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
C
h
a
r
a
cteriz
a
tio
n
o
f b
in
a
r
iz
ed
n
eu
r
a
l n
etw
o
r
ks fo
r
efficien
t d
ep
lo
yme
n
t
…
(
R
a
my
a
B
a
n
a
va
r
a
N
a
r
a
ya
n
a
)
1817
an
aly
s
is
o
f
o
u
tco
m
es
to
war
d
s
ex
h
ib
itin
g
m
o
d
el
r
o
b
u
s
tn
ess
ag
ain
s
t
ex
is
tin
g
s
ch
em
es
f
r
eq
u
en
tly
r
ep
o
r
ted
in
liter
atu
r
es.
I
n
th
e
s
u
g
g
ested
s
y
s
tem
,
a
B
NN
s
u
ited
f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
ed
g
e
d
e
v
ices
is
d
esig
n
ed
,
tr
ain
ed
,
an
d
d
ep
l
o
y
ed
.
B
y
u
tili
zin
g
th
e
L
AR
Q
f
r
am
ewo
r
k
,
t
h
e
s
y
s
tem
in
co
r
p
o
r
ates
weig
h
t
clip
p
in
g
,
b
in
ar
y
q
u
an
tizatio
n
,
an
d
p
o
in
twis
e
co
n
v
o
l
u
tio
n
s
to
d
r
asti
ca
lly
lo
wer
m
em
o
r
y
a
n
d
p
r
o
ce
s
s
in
g
d
em
a
n
d
s
with
o
u
t
s
ac
r
if
icin
g
ac
cu
r
ac
y
.
T
h
e
m
o
d
el
ex
h
ib
its
r
ea
l
-
tim
e
p
ictu
r
e
class
if
icatio
n
s
k
ills
af
ter
b
ein
g
in
s
talled
o
n
th
e
Xilin
x
PYNQ
Z
2
FP
G
A
b
o
ar
d
an
d
tr
ain
ed
o
n
th
e
MN
I
ST
d
a
taset.
Utiliz
in
g
h
ar
d
war
e
ac
ce
l
er
atio
n
,
th
e
s
y
s
tem
p
r
o
c
ess
es
u
p
to
1
0
,
0
0
0
p
ictu
r
es
p
er
s
ec
o
n
d
,
ac
h
iev
in
g
im
p
r
ess
iv
e
s
p
ee
d
g
ain
s
th
at
m
ak
e
it
an
ex
tr
em
ely
ef
f
ec
tiv
e
o
p
tio
n
f
o
r
lo
w
-
laten
c
y
ap
p
licatio
n
s
in
e
d
g
e
co
m
p
u
ti
n
g
en
v
i
r
o
n
m
e
n
ts
.
2.
M
E
T
H
O
D
T
h
e
d
esig
n
o
f
th
e
cu
r
r
en
t r
esear
ch
is
co
m
p
r
e
h
en
s
iv
ely
ex
p
la
in
ed
.
T
h
e
f
ir
s
t
s
tep
o
f
th
e
s
tu
d
y
in
v
o
lv
es
a
co
m
p
r
eh
en
s
iv
e
c
h
ar
ac
ter
iza
tio
n
o
f
tr
ad
itio
n
al
NNs
an
d
B
NNs.
T
h
e
o
b
jectiv
e
is
to
e
v
alu
ate
th
e
im
p
ac
t
o
f
v
ar
y
in
g
n
etwo
r
k
co
n
f
ig
u
r
atio
n
s
o
n
p
er
f
o
r
m
an
ce
m
etr
ics
s
u
ch
as
ac
cu
r
ac
y
an
d
lo
s
s
.
T
h
e
n
ec
ess
ity
to
cr
ea
te
ef
f
ec
tiv
e
n
e
u
r
al
n
etwo
r
k
m
o
d
els
f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
n
v
ir
o
n
m
en
ts
esp
ec
ially
,
e
d
g
e
co
m
p
u
tin
g
s
ce
n
ar
io
s
wh
er
e
m
em
o
r
y
an
d
c
o
m
p
u
te
r
eso
u
r
ce
s
ar
e
s
ca
r
ce
-
is
th
e
m
ain
d
r
i
v
in
g
f
o
r
ce
b
eh
in
d
th
e
u
s
e
o
f
B
NNs
in
th
is
s
tu
d
y
.
B
ec
au
s
e
th
e
y
em
p
lo
y
h
ig
h
-
p
r
ec
is
io
n
weig
h
ts
a
n
d
a
ctiv
atio
n
s
,
tr
ad
itio
n
al
n
e
u
r
al
n
et
wo
r
k
s
,
d
esp
ite
th
eir
p
o
wer
,
f
r
eq
u
en
tly
n
ee
d
a
s
ig
n
if
ican
t
am
o
u
n
t
o
f
m
e
m
o
r
y
an
d
p
r
o
ce
s
s
in
g
p
o
wer
,
m
a
k
in
g
t
h
em
u
n
s
u
itab
le
f
o
r
d
ep
lo
y
m
e
n
t
o
n
d
ev
ices
with
li
m
ited
r
eso
u
r
ce
s
,
s
u
ch
ed
g
e
d
e
v
ices.
B
y
lo
wer
in
g
th
e
ac
cu
r
a
cy
o
f
b
o
th
weig
h
ts
an
d
ac
tiv
atio
n
s
to
b
in
ar
y
v
alu
es
(
-
1
o
r
+1
)
,
B
NNs
p
r
o
v
id
e
a
co
n
v
in
cin
g
way
ar
o
u
n
d
th
ese
r
estrictio
n
s
wh
ile
also
d
r
asti
ca
lly
lo
wer
in
g
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
an
d
m
em
o
r
y
f
o
o
tp
r
i
n
t
o
f
th
e
m
o
d
el.
B
ec
au
s
e
o
f
th
is
,
B
NNs
ca
n
co
n
d
u
ct
r
ea
l
-
tim
e
in
f
er
en
ce
o
n
d
ev
ices
with
co
n
s
tr
ain
ed
p
r
o
ce
s
s
in
g
p
o
we
r
b
ec
au
s
e
th
ey
ar
e
q
u
ick
er
an
d
u
s
e
less
m
em
o
r
y
th
an
s
tan
d
ar
d
n
e
u
r
al
n
etwo
r
k
s
.
Fo
r
t
h
e
f
o
llo
win
g
r
ea
s
o
n
s
,
we
d
ec
id
ed
ag
ain
s
t
u
s
in
g
alter
n
ativ
e
ac
tiv
e
n
et
wo
r
k
s
o
r
b
ac
k
p
r
o
p
ag
atio
n
n
eu
r
al
n
etwo
r
k
s
,
wh
ic
h
ar
e
co
n
v
en
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
tr
ain
ed
u
s
in
g
co
n
v
en
tio
n
al
g
r
ad
ien
t
d
escen
t te
ch
n
iq
u
es:
−
C
o
m
p
u
tatio
n
al
ef
f
icien
c
y
:
w
h
en
im
p
lem
e
n
ted
o
n
ed
g
e
d
ev
ices
with
co
n
s
tr
ain
ed
p
r
o
c
ess
in
g
ca
p
ac
ity
,
b
ac
k
p
r
o
p
ag
atio
n
-
b
ased
n
etw
o
r
k
s
-
esp
ec
ially
th
o
s
e
with
f
u
ll
-
p
r
ec
is
io
n
weig
h
ts
-
f
r
e
q
u
en
tly
h
av
e
lar
g
e
co
m
p
u
tatio
n
al
co
s
ts
.
T
h
e
B
NN
is
a
b
etter
f
it
f
o
r
t
h
e
in
ten
d
ed
d
ep
l
o
y
m
en
t
p
latf
o
r
m
b
ec
a
u
s
e
o
f
its
b
in
a
r
y
f
o
r
m
at,
wh
ic
h
en
ab
les ef
f
ec
tiv
e
tr
ain
in
g
an
d
in
f
e
r
en
ce
.
−
Me
m
o
r
y
c
o
n
s
tr
ain
ts
:
b
ec
au
s
e
B
NNs
r
eq
u
ir
e
less
m
em
o
r
y
,
th
ey
ar
e
m
o
r
e
s
u
ited
f
o
r
ed
g
e
d
ev
ices
with
co
n
s
tr
ain
ed
m
em
o
r
y
.
Fu
ll
-
p
r
e
cisi
o
n
m
o
d
els
a
r
e
less
ap
p
r
o
p
r
iate
f
o
r
d
e
p
lo
y
m
e
n
t
o
n
s
u
c
h
p
latf
o
r
m
s
s
in
ce
th
ey
f
r
e
q
u
en
tly
n
ee
d
a
l
o
t m
o
r
e
R
AM
.
−
R
ea
l
-
tim
e
p
r
o
ce
s
s
in
g
:
th
e
in
f
e
r
en
ce
s
p
ee
d
b
ec
o
m
es
cr
u
cial
as
we
s
tr
iv
e
f
o
r
r
ea
l
-
tim
e
p
er
f
o
r
m
an
ce
o
n
t
h
e
ed
g
e.
I
n
co
m
p
ar
is
o
n
to
co
n
v
e
n
tio
n
al
n
etwo
r
k
s
,
B
NNs
'
h
ar
d
war
e
ac
ce
ler
atio
n
,
esp
ec
ially
o
n
p
latf
o
r
m
s
lik
e
th
e
PYNQ
Z
2
FP
GA,
en
ab
les s
ig
n
if
ican
tly
f
aster
in
f
er
e
n
ce
ti
m
es.
T
h
e
m
ain
r
ea
s
o
n
f
o
r
th
e
a
d
o
p
tio
n
o
f
B
NNs
in
th
is
wo
r
k
is
th
ei
r
ef
f
ec
tiv
en
ess
in
ter
m
s
o
f
co
m
p
u
tatio
n
,
m
em
o
r
y
u
s
ag
e,
an
d
in
f
er
en
c
e
s
p
ee
d
.
T
h
is
m
ak
es
th
em
esp
ec
ially
s
u
itab
le
f
o
r
r
ea
l
-
tim
e
ap
p
licatio
n
s
in
v
o
lv
in
g
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
ed
g
e
d
ev
ices.
E
v
en
wh
ile
th
ey
ar
e
u
s
ef
u
l,
tr
ad
itio
n
al
n
eu
r
a
l
n
etwo
r
k
s
d
o
n
o
t p
r
o
v
i
d
e
th
e
s
am
e
ef
f
icien
cy
t
r
ad
e
-
o
f
f
s
,
p
ar
t
icu
lar
ly
wh
en
aim
in
g
f
o
r
ed
g
e
d
ev
ices.
B
u
ild
in
g
o
n
th
e
in
s
ig
h
ts
f
r
o
m
th
e
ch
ar
ac
ter
izatio
n
p
h
ase,
t
h
e
s
ec
o
n
d
p
h
ase
f
o
cu
s
es
o
n
d
esig
n
in
g
a
n
o
v
el
B
NN
,
wh
ich
in
v
o
lv
es
d
eter
m
in
in
g
th
e
n
u
m
b
er
o
f
l
ay
er
s
,
th
e
n
u
m
b
er
o
f
n
eu
r
o
n
s
p
er
lay
e
r
,
an
d
th
e
ac
tiv
atio
n
f
u
n
ctio
n
in
L
AR
Q.
T
h
e
f
in
al
p
h
ase
in
v
o
lv
es
d
em
o
n
s
tr
atin
g
th
e
p
r
ac
ticality
an
d
ef
f
ec
tiv
en
ess
o
f
th
e
d
esig
n
ed
B
NN
m
o
d
el
b
y
im
p
l
em
en
tin
g
it
in
a
r
ea
l
-
tim
e
a
p
p
l
icatio
n
o
n
t
h
e
PYNQZ
2
FP
GA
b
o
ar
d
f
o
r
p
atter
n
r
ec
o
g
n
itio
n
u
s
in
g
th
e
MN
I
ST
d
ataset.
T
h
is
in
v
o
lv
es
p
r
o
g
r
a
m
m
in
g
th
e
b
o
ar
d
t
o
p
er
f
o
r
m
t
h
e
o
p
er
atio
n
s
o
f
th
e
B
NN
.
T
h
e
PY
NQZ
2
b
o
ar
d
is
a
Xilin
x
b
o
ar
d
th
at
ca
n
b
e
p
r
o
g
r
am
m
ed
u
s
in
g
Py
th
o
n
.
B
in
ar
izatio
n
is
a
cr
itica
l
p
r
ep
r
o
ce
s
s
in
g
s
tep
in
m
an
y
i
m
ag
e
an
al
y
s
is
an
d
p
atter
n
r
e
co
g
n
itio
n
alg
o
r
ith
m
s
.
Fo
r
ex
a
m
p
le,
in
d
o
c
u
m
en
t
im
ag
e
p
r
o
ce
s
s
in
g
,
b
in
a
r
izatio
n
is
u
s
ed
to
co
n
v
er
t
s
ca
n
n
e
d
d
o
cu
m
en
ts
in
to
a
b
in
a
r
y
im
ag
e,
wh
ich
ca
n
th
en
b
e
p
r
o
ce
s
s
ed
to
ex
tr
ac
t
te
x
t
an
d
o
th
er
in
f
o
r
m
atio
n
f
r
o
m
th
e
d
o
cu
m
en
t.
I
n
im
ag
e
s
eg
m
en
tat
io
n
,
b
in
ar
izatio
n
is
u
s
ed
t
o
s
ep
ar
ate
an
im
a
g
e
in
t
o
d
is
tin
ct
r
eg
i
o
n
s
,
s
u
ch
as
s
ep
ar
atin
g
th
e
f
o
r
e
g
r
o
u
n
d
f
r
o
m
t
h
e
b
ac
k
g
r
o
u
n
d
in
a
s
ce
n
e.
I
n
co
m
p
u
ter
v
is
io
n
,
b
in
ar
izatio
n
is
u
s
ed
to
co
n
v
er
t
im
ag
es
to
a
f
o
r
m
at
th
at
ca
n
b
e
p
r
o
ce
s
s
ed
b
y
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
f
o
r
task
s
s
u
ch
as
o
b
jec
t
d
etec
ti
o
n
a
n
d
r
ec
o
g
n
itio
n
.
T
h
e
tr
ain
i
n
g
p
r
o
ce
s
s
s
tar
ts
b
y
im
p
o
r
tin
g
th
e
n
ec
ess
ar
y
lib
r
ar
ies
th
at
ar
e
T
en
s
o
r
Flo
w
an
d
L
AR
Q
as
s
h
o
wn
in
Fig
u
r
e
1
.
T
h
e
MN
I
ST
d
ataset
is
n
ex
t
lo
ad
ed
u
s
in
g
T
e
n
s
o
r
Flo
w'
s
b
u
ilt
-
in
f
u
n
ctio
n
.
T
h
e
d
a
taset
is
s
p
lit
in
to
tr
ain
in
g
an
d
t
est
s
ets,
allo
win
g
th
e
m
o
d
el
to
lear
n
f
r
o
m
o
n
e
s
et
o
f
d
ata
an
d
ev
alu
ate
its
p
er
f
o
r
m
an
ce
o
n
an
o
th
e
r
.
B
ef
o
r
e
f
ee
d
in
g
th
e
d
ata
in
to
th
e
m
o
d
el,
it
n
ee
d
s
to
b
e
p
r
ep
r
o
ce
s
s
ed
.
T
h
e
im
ag
es
ar
e
r
esh
ap
ed
to
a
co
n
s
is
ten
t
s
ize
an
d
n
o
r
m
alize
d
s
o
th
at
th
e
p
ix
el
v
alu
es
r
a
n
g
e
f
r
o
m
-
1
to
1
.
T
h
is
n
o
r
m
aliza
tio
n
h
elp
s
in
b
etter
co
n
v
e
r
g
en
ce
d
u
r
in
g
tr
ai
n
in
g
an
d
en
s
u
r
es th
at
th
e
in
p
u
t d
ata
is
o
n
a
s
im
ilar
s
ca
le.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
3
,
Sep
tem
b
er
20
25
:
1
8
1
5
-
1
8
2
5
1818
Fig
u
r
e
1
.
L
AR
Q
tr
ain
in
g
p
r
o
c
ess
I
n
th
is
p
ap
er
,
a
d
ee
p
lear
n
i
n
g
m
o
d
el
is
im
p
lem
en
ted
f
o
r
im
a
g
e
class
if
icatio
n
u
s
in
g
T
en
s
o
r
Flo
w
an
d
L
AR
Q
,
a
lib
r
ar
y
f
o
r
q
u
a
n
tized
n
eu
r
al
n
etwo
r
k
s
.
Af
te
r
all
th
e
lay
er
s
a
r
e
a
d
d
ed
to
th
e
m
o
d
el
,
a
s
u
m
m
ar
y
o
f
th
e
m
o
d
el
is
p
r
in
ted
u
s
in
g
th
e
L
AR
Q
f
u
n
ctio
n
as
s
h
o
wn
in
T
ab
le
1
.
T
h
e
m
o
d
el
is
tr
ain
e
d
an
d
e
v
alu
ated
o
n
th
e
MN
I
ST
d
ataset,
wh
ich
c
o
n
s
i
s
ts
o
f
g
r
a
y
s
ca
le
im
ag
es
o
f
h
an
d
wr
itten
d
i
g
its
f
r
o
m
0
to
9
[
2
2
]
.
T
h
e
tr
ain
in
g
p
r
o
ce
s
s
s
tar
ts
b
y
im
p
o
r
tin
g
th
e
n
ec
ess
ar
y
lib
r
ar
ies
th
at
ar
e
T
en
s
o
r
Flo
w
an
d
L
AR
Q
.
T
en
s
o
r
Flo
w
is
a
p
o
p
u
lar
d
ee
p
lear
n
in
g
f
r
a
m
ewo
r
k
,
an
d
L
AR
Q
p
r
o
v
id
es
to
o
ls
f
o
r
q
u
an
tizin
g
n
eu
r
al
n
et
wo
r
k
la
y
er
s
,
wh
ich
r
ed
u
ce
s
m
em
o
r
y
an
d
co
m
p
u
tatio
n
r
e
q
u
ir
em
en
ts
.
T
h
e
MN
I
ST
d
at
aset
is
n
ex
t
lo
ad
e
d
u
s
in
g
T
e
n
s
o
r
Flo
w'
s
b
u
ilt
-
in
f
u
n
ctio
n
.
T
h
e
d
ataset
is
s
p
lit
i
n
to
tr
ain
in
g
an
d
test
s
ets,
allo
win
g
th
e
m
o
d
el
to
lear
n
f
r
o
m
o
n
e
s
et
o
f
d
ata
an
d
ev
alu
ate
its
p
er
f
o
r
m
an
ce
o
n
a
n
o
th
er
.
B
ef
o
r
e
f
ee
d
in
g
th
e
d
at
a
in
to
th
e
m
o
d
el,
it
n
ee
d
s
to
b
e
p
r
ep
r
o
ce
s
s
ed
.
T
h
e
im
ag
es
ar
e
r
esh
ap
ed
to
a
c
o
n
s
is
ten
t
s
ize
an
d
n
o
r
m
alize
d
s
o
th
at
th
e
p
ix
el
v
alu
es
r
an
g
e
f
r
o
m
-
1
to
1
.
T
h
is
n
o
r
m
aliza
tio
n
h
elp
s
in
b
etter
c
o
n
v
er
g
en
ce
d
u
r
in
g
tr
ain
in
g
an
d
en
s
u
r
es th
at
th
e
in
p
u
t d
ata
is
o
n
a
s
im
ilar
s
ca
le.
T
h
e
m
o
d
el
is
th
en
cr
ea
ted
as
a
Seq
u
en
tial
m
o
d
el,
wh
ich
all
o
ws
f
o
r
a
lin
ea
r
s
tack
o
f
lay
er
s
.
T
h
e
f
ir
s
t
lay
er
ad
d
ed
to
t
h
e
m
o
d
el
is
a
Qu
an
tC
o
n
v
2
D
la
y
er
.
T
h
is
lay
er
p
er
f
o
r
m
s
co
n
v
o
lu
tio
n
al
o
p
e
r
atio
n
s
,
wh
ic
h
h
elp
in
d
etec
tin
g
d
if
f
er
e
n
t
p
atter
n
s
an
d
f
ea
tu
r
es
in
th
e
im
ag
es.
T
h
e
lay
er
h
as
3
2
f
ilter
s
,
ea
ch
u
s
in
g
a
3
x
3
k
e
r
n
el
to
s
ca
n
th
r
o
u
g
h
th
e
im
ag
es.
T
h
e
weig
h
t
q
u
an
tizatio
n
ap
p
lied
to
th
is
lay
er
r
ed
u
ce
s
th
e
p
r
ec
is
io
n
o
f
th
e
weig
h
ts
,
m
ak
in
g
th
e
m
o
d
el
m
o
r
e
ef
f
ic
ien
t
an
d
f
aster
to
co
m
p
u
te.
Af
ter
th
e
Qu
an
tC
o
n
v
2
D
lay
e
r
,
a
Ma
x
Po
o
lin
g
2
D
lay
er
is
ad
d
ed
.
T
h
is
lay
er
p
er
f
o
r
m
s
d
o
wn
s
am
p
lin
g
b
y
s
elec
tin
g
th
e
m
ax
im
u
m
v
alu
e
with
in
a
s
p
ec
if
ic
r
eg
io
n
o
f
th
e
f
ea
tu
r
e
m
a
p
.
Ma
x
p
o
o
l
in
g
h
elp
s
in
r
ed
u
ci
n
g
th
e
s
p
at
ial
d
im
en
s
io
n
s
an
d
ca
p
tu
r
in
g
th
e
m
o
s
t
im
p
o
r
tan
t
in
f
o
r
m
atio
n
f
r
o
m
th
e
p
r
e
v
io
u
s
lay
er
.
T
o
im
p
r
o
v
e
th
e
m
o
d
el's
tr
ain
in
g
s
tab
ilit
y
an
d
co
n
v
er
g
en
ce
s
p
ee
d
,
a
B
atch
No
r
m
aliza
tio
n
lay
er
is
a
d
d
ed
.
T
h
is
lay
e
r
n
o
r
m
alize
s
t
h
e
o
u
tp
u
ts
o
f
t
h
e
p
r
ev
io
u
s
lay
er
an
d
m
ai
n
tain
s
a
r
u
n
n
in
g
m
ea
n
an
d
v
a
r
ian
ce
d
u
r
in
g
t
r
ain
in
g
,
h
el
p
in
g
in
f
aster
an
d
m
o
r
e
s
tab
le
co
n
v
er
g
e
n
ce
.
T
h
e
m
o
d
el
is
th
en
ex
p
an
d
e
d
with
an
o
th
er
Q
u
an
tC
o
n
v
2
D
lay
er
,
f
o
llo
we
d
b
y
a
Ma
x
Po
o
lin
g
2
D
l
ay
er
an
d
an
o
th
er
B
atch
No
r
m
aliza
tio
n
lay
er
.
T
h
i
s
p
atter
n
is
r
ep
ea
ted
to
cr
ea
te
a
d
ee
p
er
an
d
m
o
r
e
p
o
we
r
f
u
l m
o
d
el,
allo
win
g
it t
o
lear
n
m
o
r
e
co
m
p
le
x
p
atter
n
s
a
n
d
f
ea
tu
r
es f
r
o
m
t
h
e
d
ata.
T
ab
le
1
.
Mo
d
el
s
u
m
m
ar
y
S
e
q
u
e
n
t
i
a
l
s
u
mm
a
r
y
To
t
a
l
p
a
r
a
ms
9
3
.
6
k
Tr
a
i
n
a
b
l
e
p
a
r
a
ms
9
3
.
1
k
N
o
n
-
t
r
a
i
n
a
b
l
e
p
a
r
a
ms
4
6
8
M
o
d
e
l
s
i
z
e
1
3
.
1
9
K
i
B
M
o
d
e
l
s
i
z
e
(
8
-
b
i
t
F
P
w
e
i
g
h
t
s)
1
1
.
8
2
K
i
B
F
l
o
a
t
–
3
2
e
q
u
i
v
a
l
e
n
t
3
6
5
.
4
5
K
i
B
C
o
m
p
r
e
ss
i
o
n
r
a
t
i
o
o
f
mem
o
r
y
0
.
0
4
N
u
mb
e
r
o
f
M
A
C
s
2
.
7
9
M
R
a
t
i
o
o
f
M
A
C
s
t
h
a
t
a
r
e
b
i
n
a
r
i
z
e
d
0
.
9
3
0
3
T
h
e
q
u
an
tizatio
n
o
p
tio
n
s
f
o
r
t
h
e
q
u
an
tized
lay
er
s
ar
e
d
e
f
in
e
d
,
wh
ich
in
clu
d
e
th
e
n
u
m
b
er
o
f
b
its
u
s
ed
f
o
r
q
u
an
tizin
g
weig
h
ts
an
d
ac
tiv
atio
n
s
.
T
h
is
h
elp
s
in
co
n
tr
o
llin
g
th
e
tr
a
d
e
-
o
f
f
b
etwe
en
m
o
d
el
ac
cu
r
ac
y
a
n
d
ef
f
icien
cy
.
Hig
h
e
r
q
u
a
n
tizatio
n
b
it
-
d
e
p
th
ca
n
r
e
tain
m
o
r
e
ac
cu
r
ac
y
b
u
t
r
e
q
u
ir
es
m
o
r
e
m
em
o
r
y
an
d
co
m
p
u
tatio
n
,
wh
ile
l
o
wer
b
it
-
d
ep
th
s
ac
r
if
ices
s
o
m
e
ac
cu
r
ac
y
f
o
r
ef
f
icien
cy
g
ain
s
.
T
h
e
n
e
x
t
lay
er
s
ad
d
ed
t
o
th
e
m
o
d
el
ar
e
Qu
an
tDen
s
e
lay
er
s
,
wh
ich
ar
e
f
u
lly
co
n
n
ec
t
ed
lay
er
s
with
weig
h
t
q
u
an
tizatio
n
ap
p
lied
.
T
h
e
f
ir
s
t
Qu
an
tDen
s
e
lay
er
h
as
6
4
u
n
its
,
an
d
th
e
s
ec
o
n
d
o
n
e
h
a
s
1
0
u
n
its
,
r
ep
r
esen
tin
g
th
e
n
u
m
b
er
o
f
class
es
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
C
h
a
r
a
cteriz
a
tio
n
o
f b
in
a
r
iz
ed
n
eu
r
a
l n
etw
o
r
ks fo
r
efficien
t d
ep
lo
yme
n
t
…
(
R
a
my
a
B
a
n
a
va
r
a
N
a
r
a
ya
n
a
)
1819
th
e
MN
I
ST
d
ataset
(
0
to
9
d
i
g
its
)
.
T
h
e
f
in
al
Qu
an
tDen
s
e
l
ay
er
o
u
tp
u
ts
th
e
p
r
ed
icted
p
r
o
b
ab
ilit
ies
f
o
r
ea
c
h
class
,
an
d
th
e
o
u
tp
u
t w
ith
th
e
h
ig
h
est
p
r
o
b
ab
ilit
y
is
co
n
s
id
er
ed
th
e
p
r
e
d
icted
class
.
3.
RE
SU
L
T
S
T
h
e
ex
p
er
im
en
tal
s
etu
p
u
s
in
g
PYNQZ
2
FP
GA
b
o
ar
d
is
as
s
h
o
wn
i
n
Fig
u
r
e
2
.
Py
th
o
n
a
n
d
th
e
Xilin
x
PYNQ
f
r
am
ewo
r
k
,
wh
ich
allo
ws u
s
er
s
to
cr
ea
te
an
d
im
p
lem
en
t FPGA
ap
p
licatio
n
s
,
wer
e
u
s
ed
to
p
r
o
g
r
am
th
e
PYNQ
Z
2
FP
GA
b
o
ar
d
in
o
u
r
ex
p
er
im
en
ts
.
T
h
e
f
o
llo
win
g
ac
tio
n
s
wer
e
p
ar
t
o
f
th
e
d
ep
l
o
y
m
en
t
p
r
o
ce
s
s
:
i)
u
s
in
g
th
e
s
p
ec
if
ie
d
UR
L
an
d
p
ass
wo
r
d
to
co
n
n
ec
t
t
o
th
e
FP
GA
b
o
ar
d
,
ii)
s
tar
tin
g
a
J
u
p
y
ter
n
o
te
b
o
o
k
o
n
th
e
PYNQ
Z
2
b
o
ar
d
,
wh
ich
m
a
d
e
it
p
o
s
s
ib
le
to
co
m
m
u
n
icate
with
th
e
s
o
f
twar
e
an
d
h
ar
d
war
e
elem
en
ts
,
iii)
u
s
in
g
th
e
MN
I
ST
d
ataset,
th
e
B
NN
m
o
d
el
is
tr
ain
e
d
u
s
in
g
th
e
L
A
R
Q
f
r
am
ewo
r
k
an
d
th
en
u
s
ed
f
o
r
r
ea
l
-
tim
e
im
ag
e
class
if
icatio
n
,
an
d
iv
)
tr
ac
k
in
g
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
d
u
r
i
n
g
d
e
p
lo
y
m
e
n
t,
in
clu
d
in
g
clas
s
if
icatio
n
r
ates
an
d
in
f
er
en
ce
s
p
ee
d
s
,
wh
ich
wer
e
n
o
ted
an
d
d
o
cu
m
e
n
ted
in
th
e
s
tu
d
y
.
T
h
e
s
o
f
twar
e
p
ac
k
a
g
e
d
ep
lo
y
ed
a
r
e
as
f
o
llo
ws:
−
T
en
s
o
r
Flo
w:
th
e
B
NN
m
o
d
el
is
d
ev
elo
p
ed
a
n
d
tr
ai
n
ed
u
s
in
g
T
en
s
o
r
Flo
w,
an
o
p
e
n
-
s
o
u
r
c
e
d
ee
p
lear
n
in
g
p
latf
o
r
m
.
Her
e
,
th
e
B
NN
was
im
p
lem
en
ted
an
d
tr
ain
ed
o
n
th
e
MN
I
ST
d
ataset
u
s
in
g
th
is
v
er
s
atile
f
r
am
ewo
r
k
f
o
r
n
eu
r
al
n
etwo
r
k
d
esig
n
an
d
tr
ain
in
g
.
−
L
AR
Q
:
th
is
lib
r
ar
y
is
s
p
ec
if
ically
d
esig
n
ed
f
o
r
q
u
a
n
tized
a
n
d
B
NN
s
.
I
t
is
p
er
f
ec
t
f
o
r
t
h
e
B
NNs
u
tili
ze
d
in
th
is
s
tu
d
y
s
in
ce
it
o
f
f
er
s
ef
f
ec
tiv
e
im
p
lem
en
tatio
n
s
f
o
r
b
in
a
r
y
weig
h
ts
an
d
ac
tiv
atio
n
s
.
L
A
R
Q
was
lin
k
e
d
with
T
en
s
o
r
Flo
w
to
d
ef
i
n
e
an
d
tr
ain
th
e
m
o
d
el
,
an
d
it
p
lay
ed
a
k
ey
r
o
le
i
n
e
n
ab
lin
g
th
e
b
in
ar
y
q
u
an
tizatio
n
p
r
o
ce
s
s
.
−
PYNQ
f
r
am
ewo
r
k
:
we
u
s
ed
th
e
p
y
th
o
n
f
o
r
Z
y
n
q
(
PYN
Q
)
f
r
am
ewo
r
k
,
wh
ich
en
a
b
le
s
Py
th
o
n
-
b
ased
d
ev
elo
p
m
e
n
t
f
o
r
FP
GA
p
latf
o
r
m
s
b
ased
o
n
Xilin
x
Z
y
n
q
,
f
o
r
FP
GA
-
b
ased
d
ep
lo
y
m
en
t.
T
h
e
tr
ain
ed
B
NN
m
o
d
el
was
d
ep
lo
y
e
d
to
th
e
PYNQ
Z
2
FP
GA
b
o
ar
d
u
s
in
g
th
is
f
r
am
ewo
r
k
,
wh
ich
o
f
f
er
s
an
in
tu
itiv
e
in
ter
f
ac
e
f
o
r
in
ter
f
ac
in
g
with
t
h
e
FP
GA
h
ar
d
war
e.
−
J
u
p
y
ter
No
te
b
o
o
k
:
th
e
m
o
d
e
l
was
d
ev
elo
p
ed
,
tr
ain
ed
,
a
n
d
ass
ess
ed
in
ter
ac
tiv
ely
u
s
in
g
th
e
J
u
p
y
ter
No
teb
o
o
k
e
n
v
ir
o
n
m
en
t.
Fo
r
h
ar
d
war
e
d
ep
l
o
y
m
en
t,
t
h
is
en
v
i
r
o
n
m
en
t
also
en
a
b
les
s
m
o
o
th
i
n
teg
r
atio
n
with
th
e
PYNQ
Z
2
b
o
ar
d
.
I
m
p
lem
en
tin
g
o
b
ject
id
en
tific
atio
n
u
s
in
g
B
NNs
o
n
th
e
PY
NQ
Z
2
FP
GA
b
o
ar
d
is
a
d
etailed
p
r
o
ce
s
s
th
at
b
eg
in
s
with
co
n
n
ec
tin
g
to
th
e
b
o
ar
d
u
s
in
g
a
d
esig
n
ated
UR
L
an
d
p
ass
wo
r
d
.
L
a
u
n
ch
in
g
a
J
u
p
y
ter
n
o
teb
o
o
k
o
n
t
h
e
b
o
a
r
d
a
llo
ws
u
s
er
s
to
tr
an
s
itio
n
f
r
o
m
p
lan
n
in
g
t
o
p
r
ac
tical
m
o
d
e
l
d
ev
elo
p
m
e
n
t
an
d
ex
p
er
im
en
tatio
n
.
T
h
e
co
r
e
o
f
t
h
e
p
r
o
ce
s
s
in
v
o
lv
es
tr
ain
in
g
a
d
ataset
u
s
in
g
th
e
L
AR
Q
B
NN
f
r
am
ewo
r
k
with
in
th
e
J
u
p
y
ter
n
o
te
b
o
o
k
en
v
ir
o
n
m
en
t,
lev
er
ag
in
g
th
e
ca
p
a
b
ilit
ies
o
f
th
e
FP
GA
b
o
ar
d
f
o
r
ef
f
icien
t
co
m
p
u
tatio
n
as.
Mo
n
ito
r
in
g
p
o
wer
co
n
s
u
m
p
tio
n
ac
cu
r
ately
is
cr
u
cial,
n
ec
ess
itatin
g
th
e
u
s
e
o
f
an
ex
ter
n
al
d
ev
ice
f
o
r
m
ea
s
u
r
em
en
t.
T
h
e
im
p
lem
en
t
atio
n
jo
u
r
n
ey
f
o
r
a
n
o
v
el
B
NN
tailo
r
ed
f
o
r
ed
g
e
d
ev
ices
in
itiated
with
B
N
N
d
ev
elo
p
e
d
in
L
AR
Q
as
s
h
o
wn
in
Fig
u
r
e
2
.
B
in
ar
izatio
n
tech
n
iq
u
es
r
ed
u
ce
co
m
p
u
tati
o
n
al
co
m
p
lex
ity
b
y
co
n
v
er
tin
g
weig
h
ts
a
n
d
ac
ti
v
atio
n
s
in
to
b
in
ar
y
v
al
u
es
(
-
1
o
r
+1
)
.
Ad
a
p
tin
g
co
n
v
o
l
u
tio
n
al
lay
er
s
an
d
ad
d
r
ess
in
g
ch
allen
g
es
p
o
s
ed
b
y
b
atc
h
n
o
r
m
aliza
tio
n
ar
e
v
ital
s
tep
s
in
th
e
d
e
s
ig
n
p
r
o
ce
s
s
.
Fin
e
-
tu
n
in
g
th
e
ar
ch
itectu
r
e
th
r
o
u
g
h
tech
n
iq
u
es
lik
e
q
u
an
tizatio
n
-
awa
r
e
tr
a
in
in
g
en
h
an
c
es
m
o
d
el
ac
cu
r
a
cy
.
T
h
e
in
p
u
t
lay
er
p
r
o
ce
s
s
es
im
ag
es
f
o
llo
wed
b
y
L
ay
er
1
with
s
eq
u
en
ce
o
f
o
p
e
r
atio
n
s
lik
e
q
u
an
t_
co
n
v
2
d
,
m
a
x
p
o
o
lin
g
2
D,
b
atch
n
o
r
m
aliza
tio
n
to
s
tab
ilize
th
e
lear
n
in
g
p
r
o
ce
s
s
.
T
h
o
r
o
u
g
h
test
in
g
,
v
alid
atio
n
,
a
n
d
co
n
tin
u
o
u
s
m
o
n
ito
r
in
g
en
s
u
r
e
o
p
tim
al
p
er
f
o
r
m
an
ce
in
r
ea
l
-
wo
r
ld
c
o
n
d
itio
n
s
.
Fig
u
r
e
2
.
E
x
p
er
im
e
n
tal
s
etu
p
u
s
in
g
PYNQZ
2
b
o
ar
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
3
,
Sep
tem
b
er
20
25
:
1
8
1
5
-
1
8
2
5
1820
T
h
e
g
r
ap
h
ical
an
aly
s
is
h
as
b
e
en
d
o
n
e
u
s
in
g
ab
o
v
e
d
ata
an
d
t
h
e
f
o
llo
win
g
g
r
ap
h
s
wer
e
o
b
tain
ed
as
s
h
o
wn
in
Fig
u
r
e
3
f
r
o
m
wh
ich
we
ca
n
in
f
e
r
th
at
B
NN
g
av
e
b
etter
ac
cu
r
ac
y
th
an
NN
f
o
r
s
am
e
4
9
0
n
eu
r
o
n
s
in
L
ay
er
4
u
s
in
g
MN
I
ST
d
ataset
.
T
h
e
p
o
in
t
co
r
r
esp
o
n
d
in
g
t
o
th
e
h
ig
h
est
v
alu
e
o
f
th
e
s
lo
p
e
g
iv
es
th
e
o
p
tim
al
s
o
lu
tio
n
a
s
d
y
/d
x
(
co
n
s
tan
t)
=
0
.
T
h
e
T
ab
le
2
p
r
esen
ts
a
co
m
p
ar
is
o
n
o
f
th
e
p
er
f
o
r
m
an
ce
o
f
NN
an
d
B
NN
o
n
two
b
en
ch
m
a
r
k
d
atasets
:
MN
I
ST
in
T
ab
le
2
(
a)
a
n
d
C
I
F
AR
-
10
in
T
ab
le
2
(
b
)
.
T
h
e
p
er
f
o
r
m
a
n
ce
m
etr
ics
co
n
s
id
er
ed
ar
e
test
lo
s
s
,
test
ac
cu
r
ac
y
,
an
d
tim
e
tak
en
f
o
r
tr
ain
in
g
.
Fo
r
b
o
th
NN
an
d
B
NN,
in
cr
ea
s
in
g
th
e
n
u
m
b
er
o
f
h
id
d
en
lay
er
s
g
en
er
ally
im
p
r
o
v
es
test
ac
cu
r
ac
y
an
d
r
e
d
u
ce
s
test
lo
s
s
.
B
NN
co
n
s
i
s
ten
tly
o
u
tp
er
f
o
r
m
NN
in
ter
m
s
o
f
tes
t a
cc
u
r
ac
y
an
d
tim
e
tak
e
n
.
Fig
u
r
e
3
.
Gr
a
p
h
ical
an
aly
s
is
o
f
n
eu
r
o
n
s
v
er
s
u
s
ac
cu
r
ac
y
f
o
r
NN
an
d
B
NN
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
lay
er
s
,
n
eu
r
o
n
s
test
ac
cu
r
ac
y
a
n
d
test
lo
s
s
f
o
r
(
a)
MN
I
ST
an
d
(
b
)
C
I
FAR
-
1
0
d
ataset
(
a)
(
b
)
M
N
I
S
T
C
I
F
A
R
-
10
N
u
mb
e
r
o
f
h
i
d
d
e
n
l
a
y
e
r
N
u
mb
e
r
o
f
n
e
u
r
o
n
s
Te
st
l
o
ss
Te
st
a
c
c
u
r
a
c
y
Ti
me
(
ms)
N
u
mb
e
r
o
f
h
i
d
d
e
n
l
a
y
e
r
N
u
mb
e
r
o
f
n
e
u
r
o
n
s
Te
st
l
o
ss
Te
st
a
c
c
u
r
a
c
y
Ti
me
(
ms)
1
(
N
N
)
42
0
.
1
3
4
6
4
(
1
3
.
4
6
%)
0
.
9
5
9
6
9
(
9
5
.
9
6
%)
13
1
(
N
N
)
42
0
.
5
0
0
2
(
5
0
.
0
2
%)
0
.
8
6
7
5
0
(
8
6
.
7
5
%)
15
1
(
B
N
N
)
0
.
1
2
0
5
0
(
1
2
.
0
5
%)
0
.
9
6
2
1
9
(
9
6
.
2
1
%)
18
1
(
B
N
N
)
0
.
5
3
6
4
(
5
3
.
6
4
%)
0
.
8
9
6
7
(
8
9
.
6
7
%)
19
2
(
N
N
)
1
0
6
0
.
1
1
8
4
0
(
1
1
.
8
4
%)
0
.
9
6
3
5
0
(
9
6
.
3
5
%)
15
2
(
N
N
)
1
0
6
0
.
6
0
3
4
(
6
0
.
3
4
%)
0
.
8
7
1
1
(
8
7
.
1
1
%)
17
2
(
B
N
N
)
0
.
1
0
8
7
9
(
1
0
.
8
7
%)
0
.
9
6
6
4
9
(
9
6
.
6
4
%)
23
2
(
B
N
N
)
0
.
5
7
8
9
(
5
7
.
8
9
%)
0
.
9
2
2
9
(
9
2
.
2
9
%)
25
3
(
N
N
)
2
3
4
0
.
1
2
0
3
9
(
1
2
.
0
3
%)
0
.
9
6
6
0
0
(
9
6
.
6
0
%)
15
3
(
N
N
)
2
3
4
0
.
5
7
0
5
(
5
7
.
0
5
%)
0
.
9
4
9
4
(
9
4
.
9
4
%)
17
3
(
B
N
N
)
0
.
1
0
8
5
3
(
1
0
.
8
5
%)
0
.
9
6
8
9
0
(
9
6
.
8
9
%)
24
3
(
B
N
N
)
0
.
5
6
6
4
(
5
6
.
6
4
%)
0
.
9
7
5
7
(
9
7
.
5
7
%)
26
4
(
N
N
)
4
9
0
0
.
1
1
0
2
5
(
1
1
.
0
2
%)
0
.
9
6
7
1
9
(
9
6
.
7
1
%)
16
4
(
N
N
)
4
9
0
0
.
4
7
1
5
(
4
7
.
1
5
%)
0
.
9
6
4
2
(
9
6
.
4
2
%)
18
4
(
B
N
N
)
0
.
0
9
2
2
3
(
9
.
2
2
%)
0
.
9
7
0
7
0
(
9
7
.
0
7
%)
28
4
(
B
N
N
)
0
.
4
5
6
5
(
4
5
.
6
5
%)
0
.
9
8
2
0
(
9
8
.
2
0
%)
30
5
(
N
N
)
1
0
0
2
0
.
1
5
6
1
3
(
1
5
.
6
1
%)
0
.
9
5
8
5
0
(
9
5
.
8
5
%)
26
5
(
N
N
)
1
0
0
2
0
.
6
0
9
8
(
6
0
.
9
8
%)
0
.
7
3
4
8
(
7
3
.
4
8
%)
28
5
(
B
N
N
)
0
.
1
1
5
6
0
(
1
1
.
5
6
%)
0
.
9
6
7
9
9
(
9
6
.
7
9
%)
46
5
(
B
N
N
)
0
.
6
1
4
1
(
6
1
.
4
8
%)
0
.
7
8
8
5
(
7
8
.
8
5
%)
48
6
(
N
N
)
2
0
2
6
0
.
1
6
6
5
9
(
1
6
.
6
5
%)
0
.
9
5
5
9
0
(
9
5
.
5
9
%)
70
6
(
N
N
)
2
0
2
6
0
.
6
0
9
5
(
6
0
.
9
5
%)
0
.
7
4
4
4
(
7
4
.
4
4
%)
72
6
(
B
N
N
)
0
.
1
2
3
1
9
(
1
2
.
1
3
%)
0
.
9
6
7
4
0
(
9
6
.
7
4
%)
1
1
3
6
(
B
N
N
)
0
.
5
9
4
1
(
5
9
.
4
1
%)
0
.
8
9
5
5
(
8
9
.
5
5
%)
1
1
8
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
C
h
a
r
a
cteriz
a
tio
n
o
f b
in
a
r
iz
ed
n
eu
r
a
l n
etw
o
r
ks fo
r
efficien
t d
ep
lo
yme
n
t
…
(
R
a
my
a
B
a
n
a
va
r
a
N
a
r
a
ya
n
a
)
1821
T
ab
le
2
p
r
esen
ts
a
co
m
p
ar
at
iv
e
an
aly
s
is
o
f
lay
er
s
,
n
eu
r
o
n
s
,
test
ac
cu
r
ac
y
,
an
d
test
lo
s
s
f
o
r
th
e
MN
I
ST
an
d
C
I
FAR
-
1
0
d
atas
ets.
T
h
e
r
esu
lts
h
ig
h
lig
h
t
th
e
r
elatio
n
s
h
ip
b
etwe
en
n
etwo
r
k
co
m
p
lex
ity
an
d
p
er
f
o
r
m
an
ce
,
d
em
o
n
s
tr
atin
g
th
at
in
cr
ea
s
in
g
lay
er
s
an
d
n
e
u
r
o
n
s
g
en
er
ally
im
p
r
o
v
es
ac
cu
r
ac
y
,
alb
eit
wit
h
d
im
in
is
h
in
g
r
etu
r
n
s
.
N
o
tab
ly
,
th
e
C
I
FAR
-
1
0
d
ataset
ex
h
ib
it
s
a
m
o
r
e
p
r
o
n
o
u
n
ce
d
tr
ad
e
-
o
f
f
b
etwe
en
ac
cu
r
ac
y
an
d
test
lo
s
s
co
m
p
ar
e
d
to
MN
I
ST,
r
ef
lectin
g
its
h
ig
h
er
co
m
p
lex
ity
.
T
h
e
b
est
p
e
r
f
o
r
m
an
ce
is
ac
h
iev
ed
b
y
t
h
e
f
o
u
r
-
la
y
er
B
NN,
wh
ich
ac
h
iev
es
a
test
ac
cu
r
ac
y
o
f
9
7
.
0
7
%
with
a
te
s
t
lo
s
s
o
f
0
.
0
9
2
2
3
.
T
h
e
tr
en
d
f
o
r
b
o
th
NN
an
d
B
NN
is
s
im
ilar
to
MN
I
ST
,
with
in
cr
ea
s
in
g
h
id
d
en
lay
er
s
lead
in
g
to
im
p
r
o
v
ed
p
er
f
o
r
m
a
n
ce
.
Ho
wev
er
,
th
e
d
if
f
er
en
ce
b
etwe
en
NN
an
d
B
NN
is
les
s
p
r
o
n
o
u
n
ce
d
o
n
C
I
FAR
-
1
0
.
T
h
e
f
o
u
r
-
lay
e
r
B
NN
ag
ain
ac
h
iev
es
th
e
b
est
p
er
f
o
r
m
an
ce
,
with
a
test
ac
cu
r
ac
y
o
f
9
8
.
2
0
%
a
n
d
a
t
est
lo
s
s
o
f
0
.
4
5
6
5
.
T
h
e
s
tu
d
y
d
em
o
n
s
tr
ates
th
is
ad
v
an
tag
e
with
a
n
im
p
r
ess
iv
e
9
7
.
6
1
%
ac
cu
r
ac
y
r
ate
ac
h
i
ev
ed
o
n
th
e
MN
I
ST
d
ataset,
s
h
o
wca
s
in
g
B
NNs
'
ca
p
ab
ilit
y
f
o
r
b
o
th
r
ap
id
p
r
o
ce
s
s
in
g
an
d
h
ig
h
class
if
icatio
n
ac
cu
r
ac
y
.
T
ab
le
3
s
h
o
wca
s
es
th
e
o
u
tco
m
e
ac
co
m
p
lis
h
ed
b
y
im
p
lem
en
tin
g
s
p
ec
if
ic
d
ataset
an
d
it
s
s
co
r
e
o
f
im
p
r
o
v
izatio
n
in
c
o
n
tr
ast
to
[
2
3
]
-
[
2
6
]
.
T
ab
le
4
s
h
o
wca
s
es
r
ed
u
ctio
n
o
f
tim
e
f
r
o
m
h
ar
d
war
e
p
latf
o
r
m
p
er
s
p
ec
tiv
e
f
o
r
b
o
th
NN
an
d
B
NN
u
s
e
-
ca
s
es.
W
ith
a
class
i
f
icatio
n
tim
e
o
f
0
.
0
0
8
4
1
m
illi
s
ec
o
n
d
s
p
er
im
a
g
e,
ap
p
r
o
x
im
ately
1
0
,
0
0
0
im
ag
es
wer
e
class
if
ied
d
u
r
in
g
th
e
i
n
f
e
r
en
ce
tim
e.
T
h
e
tim
e
ta
k
en
p
e
r
im
ag
e
r
ep
r
esen
ts
ap
p
r
o
x
im
ately
0
.
0
1
%
o
f
th
e
to
tal
in
f
er
en
ce
tim
e
.
W
h
en
co
m
p
ar
in
g
th
e
class
if
icatio
n
r
ates
o
f
d
i
f
f
er
e
n
t
in
f
er
en
ce
c
o
n
f
ig
u
r
atio
n
s
,
with
W
1
A1
as
th
e
b
aselin
e
at
1
0
0
%,
th
e
f
ir
s
t
r
u
n
o
f
W
1
A2
s
h
o
ws
a
s
ig
n
if
ican
t
d
r
o
p
in
p
er
f
o
r
m
an
ce
,
ac
h
iev
i
n
g
o
n
l
y
2
0
.
8
0
%
o
f
W
1
A1
'
s
s
p
ee
d
(
2
.
5
9
im
ag
es
p
er
s
ec
o
n
d
co
m
p
ar
ed
to
1
2
.
4
5
im
ag
es
p
er
s
ec
o
n
d
f
o
r
W
1
A1
)
.
Ho
wev
er
,
in
a
s
u
b
s
eq
u
e
n
t
r
u
n
,
W
1
A2
ex
h
ib
ited
a
m
ass
iv
e
in
cr
ea
s
e
in
s
p
ee
d
,
p
r
o
ce
s
s
in
g
at
an
asto
n
is
h
in
g
r
ate
o
f
9
5
5
,
1
9
4
.
7
8
%
f
aster
th
a
n
W
1
A1
as.
I
n
co
n
clu
s
io
n
,
th
i
s
r
esear
ch
p
av
es
th
e
way
f
o
r
a
tr
an
s
f
o
r
m
ativ
e
a
p
p
r
o
ac
h
to
im
ag
e
class
if
icatio
n
.
B
y
lev
er
ag
in
g
B
NNs
o
n
h
ar
d
war
e
p
la
tf
o
r
m
s
,
we
ca
n
ac
h
iev
e
s
ig
n
if
ican
t im
p
r
o
v
em
en
ts
in
p
r
o
c
ess
in
g
s
p
ee
d
,
ac
cu
r
ac
y
,
a
n
d
o
v
er
all
co
m
p
u
ta
tio
n
al
ef
f
ec
tiv
en
ess
,
p
ar
ticu
lar
ly
in
s
ce
n
ar
io
s
wh
er
e
th
ese
f
ac
to
r
s
ar
e
cr
itical.
T
h
e
f
in
d
in
g
s
d
em
o
n
s
tr
ate
a
d
r
am
atic
d
is
p
ar
ity
in
p
r
o
ce
s
s
in
g
s
p
ee
d
b
etwe
en
s
o
f
twar
e
an
d
h
ar
d
war
e.
So
f
twa
r
e
p
r
o
ce
s
s
in
g
s
tr
u
g
g
les
at
a
r
ate
o
f
r
o
u
g
h
ly
1
2
im
ag
es
p
er
s
ec
o
n
d
,
w
h
er
ea
s
h
ar
d
war
e
lik
e
th
e
PYNQ
Z
2
FP
GA
b
o
ar
d
ac
h
iev
es
a
p
r
o
ce
s
s
in
g
r
ate
o
f
a
s
tag
g
er
in
g
4
1
,
6
6
7
im
ag
es
p
e
r
s
ec
o
n
d
.
T
h
e
tim
e
tak
en
f
o
r
p
r
o
ce
s
s
in
g
a
s
in
g
le
im
ag
e
o
n
th
e
PYNQ
p
latf
o
r
m
h
ig
h
lig
h
ts
its
ef
f
icien
c
y
i
n
h
an
d
lin
g
co
m
p
u
tatio
n
al
wo
r
k
l
o
ad
s
.
T
h
is
m
etr
ic
is
cr
u
cial
f
o
r
e
v
alu
atin
g
th
e
p
latf
o
r
m
'
s
s
u
itab
ilit
y
f
o
r
r
ea
l
-
t
im
e
ap
p
licatio
n
s
.
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
b
etwe
en
n
eu
r
al
n
etwo
r
k
an
d
B
NN
o
n
P
YNQZ
2
FP
GA
b
o
ar
d
P
l
a
t
f
o
r
m
D
a
t
a
s
e
t
A
C
C
(
%)
To
p
o
l
o
g
y
N
o
o
f
l
a
y
e
r
s
P
a
p
e
r
2
0
2
3
-
LA
R
Q
M
N
I
S
T
9
6
.
1
1
B
N
N
8
[
2
3
]
LA
R
Q
M
N
I
S
T
9
7
.
6
1
B
N
N
4
I
n
c
r
e
a
se
d
b
y
1
.
0
1
I
n
c
r
e
a
se
d
b
y
2
2
0
2
4
-
LA
R
Q
C
I
F
A
R
-
10
7
7
.
4
2
B
N
N
8
[
2
4
]
LA
R
Q
C
I
F
A
R
-
10
9
7
.
0
3
B
N
N
4
I
n
c
r
e
a
se
d
b
y
1
9
.
6
I
n
c
r
e
a
se
d
b
y
2
2
0
2
1
-
LA
R
Q
M
N
I
S
T
90
B
N
N
10
[
2
5
]
LA
R
Q
M
N
I
S
T
9
7
.
6
1
B
N
N
4
I
n
c
r
e
a
se
d
b
y
7
.
6
1
I
n
c
r
e
a
se
d
b
y
2
.
5
2
0
2
1
-
LA
R
Q
C
I
F
A
R
-
10
8
8
.
5
B
N
N
9
[
2
6
]
LA
R
Q
C
I
F
A
R
-
10
9
7
.
0
3
B
N
N
4
I
n
c
r
e
a
se
d
b
y
8
.
5
3
I
n
c
r
e
a
se
d
b
y
2
.
2
5
T
ab
le
4
.
T
im
e
ta
k
en
f
o
r
s
in
g
le
im
ag
e
o
n
PYNQ
Ty
p
e
o
f
n
e
t
w
o
r
k
S
o
f
t
w
a
r
e
H
a
r
d
w
a
r
e
NN
2
9
.
1
5
1
ms
2
.
2
2
5
9
ms
B
N
N
8
0
.
2
9
6
ms
0
.
0
0
8
4
1
ms
An
an
aly
s
is
h
as
b
ee
n
ca
r
r
ied
o
u
t
tab
le
th
at
co
n
tr
asts
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
with
o
th
e
r
cu
r
r
en
t
r
esear
ch
o
n
B
NNs,
p
ar
ticu
lar
ly
with
r
eg
ar
d
to
p
er
f
o
r
m
an
ce
m
ea
s
u
r
es
lik
e
h
ar
d
war
e
p
latf
o
r
m
,
in
f
er
en
ce
tim
e,
an
d
test
ac
cu
r
ac
y
.
T
ab
le
5
will
p
r
o
v
id
e
u
s
a
clea
r
p
ictu
r
e
o
f
h
o
w
o
u
r
f
in
d
i
n
g
s
co
m
p
ar
e
to
th
o
s
e
o
f
o
th
er
s
tu
d
ies
o
f
a
s
im
ilar
n
at
u
r
e.
I
n
th
e
p
er
s
p
ec
tiv
e
o
f
ac
c
u
r
ac
y
c
o
m
p
ar
is
o
n
,
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
o
u
tp
er
f
o
r
m
s
a
n
u
m
b
er
o
f
p
r
ev
io
u
s
s
tu
d
ies
th
at
r
ec
o
r
d
lo
wer
ac
c
u
r
ac
y
le
v
els,
ac
h
iev
in
g
9
7
.
6
1
%
ac
cu
r
ac
y
o
n
t
h
e
MN
I
ST
d
ataset
an
d
9
7
.
0
3
%
ac
cu
r
ac
y
o
n
th
e
C
I
FAR
-
1
0
d
ataset.
Fo
r
in
s
tan
ce
,
th
e
[
2
0
2
4
-
L
AR
Q]
s
tu
d
y
r
ec
o
r
d
ed
ju
s
t
7
7
.
4
2
%
o
n
C
I
FAR
-
1
0
,
wh
ile
th
e
[
2
0
2
3
-
L
AR
Q]
s
tu
d
y
ac
h
ie
v
ed
9
6
.
1
1
%
o
n
MN
I
ST.
B
ec
au
s
e
o
f
th
e
n
etwo
r
k
ar
ch
itectu
r
e'
s
in
n
o
v
ativ
e
elem
en
ts
an
d
ef
f
icien
t
d
esig
n
,
o
u
r
m
o
d
el'
s
ac
cu
r
ac
y
is
n
o
ticea
b
l
y
h
ig
h
er
,
s
u
g
g
esti
n
g
im
p
r
o
v
e
d
g
en
e
r
aliza
tio
n
ca
p
ab
ilit
ies.
I
n
th
e
p
er
s
p
ec
tiv
e
o
f
in
f
er
e
n
ce
tim
e,
co
m
p
ar
ed
to
th
e
p
u
b
lis
h
ed
n
u
m
b
er
s
f
r
o
m
o
th
er
s
tu
d
ies,
o
u
r
s
u
g
g
ested
s
y
s
tem
'
s
in
f
er
e
n
ce
tim
e
is
n
o
ticea
b
ly
s
h
o
r
ter
.
W
h
ile
th
e
FP
GA
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
3
,
Sep
tem
b
er
20
25
:
1
8
1
5
-
1
8
2
5
1822
im
p
lem
en
tatio
n
u
s
ed
in
th
e
[
2
0
2
3
-
L
AR
Q]
s
tu
d
y
to
o
k
2
.
2
2
5
9
m
s
p
er
im
ag
e,
o
u
r
m
o
d
el
p
r
o
ce
s
s
es
an
im
ag
e
in
0
.
0
0
8
4
1
m
s
.
T
h
is
illu
s
tr
ates
h
o
w
h
ar
d
war
e
ac
ce
ler
atio
n
,
b
in
ar
y
q
u
an
tizatio
n
,
an
d
m
o
d
e
l
o
p
tim
izatio
n
s
ca
n
s
ig
n
if
ican
tly
in
cr
ea
s
e
p
r
o
ce
s
s
in
g
s
p
ee
d
f
o
r
ed
g
e
d
ev
ic
es.
I
n
p
er
s
p
ec
tiv
e
o
f
h
ar
d
war
e
p
l
atf
o
r
m
,
t
h
e
Xilin
x
PYNQ
Z
2
FP
GA
is
u
s
ed
in
all
o
f
th
e
a
f
o
r
em
e
n
tio
n
ed
in
v
esti
g
atio
n
s
,
g
u
ar
an
teein
g
an
eq
u
itab
le
co
m
p
ar
is
o
n
o
f
h
ar
d
war
e
p
e
r
f
o
r
m
an
ce
.
Ou
r
r
esu
lts
s
h
o
w
th
at
th
e
m
o
d
el
o
u
tp
er
f
o
r
m
s
m
o
r
e
s
o
p
h
is
ticated
m
o
d
els
in
ter
m
s
o
f
ac
cu
r
ac
y
an
d
in
f
e
r
en
ce
tim
e
w
ith
f
ewe
r
lay
er
s
an
d
n
eu
r
o
n
s
(
4
lay
er
s
an
d
f
ewe
r
p
a
r
am
eter
s
)
.
T
ab
le
5
.
C
o
m
p
a
r
is
o
n
tab
le
o
f
r
elate
d
wo
r
k
s
an
d
th
e
p
r
o
p
o
s
ed
m
o
d
el
M
o
d
e
l
s
D
a
t
a
s
e
t
M
o
d
e
l
t
y
p
e
N
u
mb
e
r
o
f
l
a
y
e
r
s
Te
st
a
c
c
u
r
a
c
y
(
%)
I
n
f
e
r
e
n
c
e
t
i
me
(
ms/i
ma
g
e
)
[
2
0
2
3
-
LA
R
Q
]
M
N
I
S
T
B
N
N
8
9
6
.
1
1
2
.
2
2
5
9
[
2
0
2
4
-
LA
R
Q
]
C
I
F
A
R
-
10
B
N
N
8
7
7
.
4
2
2
.
4
1
6
7
[
2
0
2
1
-
LA
R
Q
]
M
N
I
S
T
B
N
N
10
90
4
.
5
7
[
2
0
2
1
-
LA
R
Q
]
C
I
F
A
R
-
10
B
N
N
9
8
8
.
5
5
.
2
P
r
o
p
o
se
d
M
N
I
S
T
B
N
N
4
9
7
.
6
1
0
.
0
0
8
4
1
P
r
o
p
o
se
d
C
I
F
A
R
-
10
B
N
N
4
9
7
.
0
3
0
.
0
0
8
4
1
T
h
e
lear
n
in
g
o
u
tco
m
e
o
f
e
x
ten
d
ed
an
al
y
s
is
as
s
h
o
wn
in
Fig
u
r
e
4
is
f
o
llo
win
g
:
o
u
r
B
NN
m
o
d
el
ef
f
ec
tiv
ely
ac
h
iev
es
h
ig
h
ac
c
u
r
ac
y
lev
els
with
f
ewe
r
p
ar
a
m
eter
s
th
an
ty
p
ical
NNs,
wh
ich
f
r
eq
u
e
n
tly
n
ee
d
m
o
r
e
lay
er
s
an
d
n
eu
r
o
n
s
to
attain
h
ig
h
er
ac
cu
r
ac
y
.
T
h
e
s
u
g
g
ested
m
o
d
el'
s
h
ar
d
war
e
ac
ce
ler
atio
n
an
d
b
in
ar
y
q
u
an
tizatio
n
s
tr
ateg
ies
en
ab
le
f
aste
r
p
r
o
ce
s
s
in
g
r
ates
with
o
u
t
s
ac
r
if
icin
g
o
r
ev
en
ex
ce
ed
i
n
g
th
e
ac
cu
r
ac
y
o
f
m
o
r
e
i
n
tr
icate
n
etwo
r
k
s
.
Ou
r
s
y
s
tem
's
r
ea
l
-
tim
e
p
er
f
o
r
m
a
n
ce
(
in
f
e
r
en
ce
tim
es
o
f
0
.
0
0
8
4
1
m
s
/im
ag
e)
m
ak
es
it
id
ea
l
f
o
r
ed
g
e
co
m
p
u
tin
g
ap
p
licatio
n
s
wh
er
e
lo
w
laten
cy
an
d
co
m
p
u
tatio
n
al
ef
f
icien
c
y
ar
e
cr
u
cial,
s
u
ch
au
to
n
o
m
o
u
s
s
y
s
tem
s
an
d
in
te
r
n
et
o
f
th
in
g
s
(
I
o
T
)
s
en
s
o
r
s
.
T
h
e
s
u
g
g
ested
a
p
p
r
o
ac
h
is
s
c
alab
le
an
d
a
d
ap
tab
le
to
in
cr
ea
s
in
g
ly
co
m
p
licated
d
atasets
o
r
r
ea
l
-
wo
r
ld
ap
p
licati
o
n
s
th
at
r
e
q
u
ir
e
h
ig
h
-
th
r
o
u
g
h
p
u
t
p
r
o
ce
s
s
in
g
with
co
n
s
tr
ain
ed
r
e
s
o
u
r
ce
s
,
as
s
ee
n
b
y
th
e
n
o
tab
le
im
p
r
o
v
em
e
n
t
i
n
ac
cu
r
ac
y
an
d
p
er
f
o
r
m
an
ce
o
v
er
ea
r
lier
ef
f
o
r
ts
.
I
n
th
is
wo
r
k
,
we
u
s
ed
th
e
Xili
n
x
PYNQ
Z
2
FP
GA
to
in
tr
o
d
u
ce
a
B
NN
d
esig
n
e
d
esp
ec
iall
y
f
o
r
d
ep
lo
y
m
en
t o
n
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
d
g
e
d
e
v
ices.
T
h
e
f
o
llo
win
g
is
a
s
u
m
m
a
r
y
o
f
th
e
m
ain
co
n
clu
s
i
o
n
s
d
r
awn
f
r
o
m
o
u
r
ex
p
er
im
en
ts
:
i)
B
NNs
p
r
o
v
id
e
r
ed
u
ce
d
c
o
m
p
l
ex
ity
an
d
co
m
p
etitiv
e
ac
cu
r
ac
y
:
o
u
r
test
s
s
h
o
wed
th
at
th
e
s
u
g
g
ested
B
NN
d
esig
n
o
u
tp
er
f
o
r
m
e
d
co
n
v
en
ti
o
n
al
NNs with
s
im
ilar
o
r
g
r
ea
ter
co
m
p
lex
ity
,
ac
h
iev
in
g
9
7
.
6
1
%
ac
cu
r
ac
y
o
n
th
e
MN
I
ST
d
ataset
with
o
n
l
y
f
o
u
r
lay
er
s
.
T
h
is
r
esu
lt
d
em
o
n
s
tr
ates
h
o
w
ef
f
ec
tiv
el
y
b
i
n
ar
y
q
u
an
tizatio
n
ca
n
p
r
eser
v
e
ex
ce
llen
t
ac
cu
r
a
cy
wh
ile
s
ig
n
if
ican
tly
lo
wer
in
g
th
e
m
em
o
r
y
a
n
d
p
r
o
ce
s
s
in
g
d
em
an
d
s
o
f
th
e
m
o
d
el.
ii)
R
ea
l
-
tim
e
in
f
er
en
ce
with
h
ar
d
war
e
ac
ce
ler
atio
n
:
1
0
,
0
0
0
p
h
o
to
s
co
u
ld
b
e
p
r
o
ce
s
s
ed
in
alm
o
s
t
th
e
s
am
e
am
o
u
n
t
o
f
tim
e
as
p
r
o
ce
s
s
in
g
a
s
in
g
le
im
ag
e
u
s
in
g
a
s
o
f
twar
e
-
b
ased
m
eth
o
d
th
an
k
s
to
th
e
PYNQ
Z
2
FP
GA
im
p
lem
en
tatio
n
,
wh
ich
p
r
o
d
u
ce
d
a
n
in
cr
ed
i
b
le
class
i
f
icatio
n
tim
e
o
f
0
.
0
0
8
4
1
m
s
p
er
im
ag
e
.
T
h
e
p
o
ten
tial
o
f
h
a
r
d
war
e
ac
ce
ler
atio
n
to
en
ab
le
r
ea
l
-
tim
e
ed
g
e
co
m
p
u
tin
g
ap
p
licatio
n
s
is
s
h
o
wn
b
y
th
is
s
ig
n
if
ican
t sp
ee
d
u
p
.
Fig
u
r
e
4
.
E
x
ten
d
e
d
co
m
p
ar
ativ
e
an
aly
s
is
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
C
h
a
r
a
cteriz
a
tio
n
o
f b
in
a
r
iz
ed
n
eu
r
a
l n
etw
o
r
ks fo
r
efficien
t d
ep
lo
yme
n
t
…
(
R
a
my
a
B
a
n
a
va
r
a
N
a
r
a
ya
n
a
)
1823
B
etter
ef
f
icien
cy
in
ed
g
e
d
e
v
i
ce
s
:
in
ter
m
s
o
f
m
em
o
r
y
a
n
d
co
m
p
u
tatio
n
al
ef
f
icien
c
y
,
th
e
s
u
g
g
ested
B
NN
p
er
f
o
r
m
ed
b
etter
th
an
ex
is
tin
g
B
NN
im
p
lem
en
tatio
n
s
as
wel
l
as
co
n
v
en
tio
n
al
NNs.
T
h
e
s
y
s
tem
's
u
s
e
o
f
b
in
ar
y
weig
h
ts
an
d
p
o
in
twis
e
co
n
v
o
l
u
tio
n
s
allo
wed
it
to
ac
h
iev
e
g
r
ea
t
p
er
f
o
r
m
an
c
e
with
l
o
wer
h
ar
d
war
e
a
n
d
m
em
o
r
y
n
ee
d
s
,
wh
ich
m
a
k
es
i
t
p
er
f
ec
t
f
o
r
d
ep
lo
y
m
en
t
in
s
it
u
atio
n
s
with
lim
ited
r
eso
u
r
ce
s
,
s
u
ch
as
em
b
e
d
d
ed
s
y
s
tem
s
an
d
I
o
T
d
e
v
ices
.
T
h
e
co
n
te
x
tu
alizin
g
with
p
r
e
v
io
u
s
s
tu
d
ies
is
as
f
o
llo
ws
:
it
is
clea
r
f
r
o
m
c
o
m
p
ar
i
n
g
o
u
r
f
in
d
in
g
s
with
th
o
s
e
o
f
o
th
er
s
tu
d
ies
th
at
t
h
e
s
u
g
g
ested
B
NN
ar
ch
i
tectu
r
e
p
er
f
o
r
m
s
b
etter
th
a
n
m
a
n
y
o
f
th
e
p
r
ec
ed
i
n
g
m
o
d
els
in
ter
m
s
o
f
ac
c
u
r
ac
y
an
d
i
n
f
er
en
ce
s
p
ee
d
.
Prio
r
r
esear
ch
o
n
B
NNs
s
h
o
wed
g
a
in
s
in
ac
cu
r
ac
y
o
n
s
im
p
ler
d
atasets
,
s
u
ch
as
M
NI
ST,
b
u
t
d
id
n
o
t
p
er
f
o
r
m
s
im
ilar
ly
o
n
m
o
r
e
co
m
p
licate
d
d
atasets
,
s
u
ch
a
s
C
I
FA
R
-
1
0
,
o
r
ap
p
ly
m
o
d
els
in
a
r
ea
l
-
tim
e
h
ar
d
war
e
en
v
ir
o
n
m
en
t,
lik
e
FP
GA.
T
h
er
ef
o
r
e,
o
u
r
wo
r
k
f
ills
a
g
ap
b
y
s
h
o
wca
s
in
g
th
e
s
ca
lab
ilit
y
an
d
p
r
ac
tical
ap
p
licatio
n
o
f
B
NNs
in
ed
g
e
co
m
p
u
tin
g
,
in
ad
d
itio
n
to
ex
h
ib
itin
g
h
ig
h
er
p
er
f
o
r
m
an
ce
ac
r
o
s
s
m
u
ltip
le
d
atasets
.
T
h
e
tak
e
-
awa
y
s
tatem
en
t
o
f
p
r
o
p
o
s
ed
s
tu
d
y
is
as
f
o
llo
w:
t
o
s
u
m
u
p
,
th
is
s
tu
d
y
d
em
o
n
s
tr
ates
th
e
am
az
in
g
p
o
te
n
tial
o
f
B
NNs
f
o
r
r
ea
l
-
tim
e
ed
g
e
co
m
p
u
tin
g
,
p
r
o
v
id
i
n
g
a
p
o
ten
t
s
o
lu
tio
n
t
h
at
b
len
d
s
m
em
o
r
y
ef
f
icien
cy
,
h
ig
h
ac
c
u
r
ac
y
,
a
n
d
q
u
ick
in
f
e
r
en
ce
s
p
ee
d
s
.
B
NN
s
'
ca
p
ac
ity
to
m
ee
t
th
e
u
r
g
en
t
n
ee
d
s
o
f
co
m
p
u
tatio
n
ally
lim
ited
co
n
te
x
ts
is
d
em
o
n
s
tr
ated
b
y
th
eir
s
u
cc
ess
f
u
l
d
ep
lo
y
m
e
n
t
o
n
th
e
PYNQ
Z
2
FP
GA,
o
p
en
in
g
th
e
d
o
o
r
f
o
r
t
h
eir
p
o
ten
tial
u
s
e
in
au
t
o
n
o
m
o
u
s
s
y
s
tem
s
,
m
ed
ical
d
ev
ices,
an
d
t
h
e
I
o
T
.
B
NNs
ar
e
ex
p
ec
ted
to
b
e
cr
u
cial
to
th
e
d
ev
elo
p
m
e
n
t
o
f
th
e
n
e
x
t
g
en
e
r
atio
n
o
f
ed
g
e
c
o
m
p
u
tin
g
tech
n
o
lo
g
ies
s
in
ce
th
e
y
o
f
f
er
e
f
f
ec
tiv
e
p
r
o
ce
s
s
in
g
ca
p
ab
ilit
ies with
lo
w
r
eso
u
r
ce
u
s
ag
e.
4.
CO
NCLU
SI
O
N
I
n
th
is
wo
r
k
,
we
in
tr
o
d
u
ce
d
a
B
NN
th
at
is
s
p
ec
if
ically
d
e
s
ig
n
ed
an
d
im
p
lem
en
ted
o
n
th
e
Xilin
x
PYNQ
Z
2
FP
GA
b
o
ar
d
,
an
d
is
o
p
tim
ized
f
o
r
ed
g
e
d
ev
ices.
T
h
e
s
u
g
g
ested
B
NN
ar
ch
itectu
r
e
o
u
tp
e
r
f
o
r
m
e
d
ea
r
lier
B
NN
im
p
lem
en
tatio
n
s
th
at
em
p
lo
y
e
d
m
o
r
e
lev
els
in
ter
m
s
o
f
test
ac
cu
r
ac
y
,
e
v
en
t
h
o
u
g
h
it
o
n
ly
u
s
ed
f
o
u
r
lay
er
s
(
8
–
1
0
)
.
I
n
p
ar
ticu
l
ar
,
o
u
r
m
o
d
el
o
b
tain
ed
a
test
ac
cu
r
ac
y
o
f
9
7
.
6
1
%
f
o
r
th
e
MN
I
ST
d
ataset
an
d
9
7
.
0
3
%
f
o
r
C
I
FAR
-
1
0
.
T
h
ese
f
in
d
in
g
s
im
p
ly
th
at
g
r
ea
t
p
er
f
o
r
m
an
ce
ca
n
b
e
attain
ed
wit
h
f
ewe
r
lay
er
s
an
d
o
p
tim
al
d
esig
n
s
,
in
cr
ea
s
in
g
t
h
e
co
m
p
u
tatio
n
al
ef
f
icien
cy
o
f
th
e
m
o
d
el.
Ou
r
m
eth
o
d
'
s
s
ig
n
if
ican
t
in
f
er
en
ce
tim
e
r
ed
u
ctio
n
is
o
n
e
o
f
its
m
a
in
b
en
e
f
its
.
C
o
m
p
ar
ed
to
s
o
f
t
war
e
-
b
ased
m
eth
o
d
s
th
at
to
o
k
s
ig
n
if
ican
tly
lo
n
g
er
(
2
.
2
2
5
9
m
s
ea
ch
im
ag
e
)
,
th
e
h
ar
d
war
e
im
p
lem
en
tatio
n
o
n
t
h
e
PYNQ
Z
2
FP
GA
s
h
o
wed
a
class
if
icatio
n
tim
e
o
f
o
n
ly
0
.
0
0
8
4
1
m
s
p
er
im
a
g
e.
Ou
r
tech
n
o
lo
g
y
is
well
s
u
ited
f
o
r
r
ea
l
-
tim
e
ed
g
e
co
m
p
u
tin
g
a
p
p
licatio
n
s
wh
er
e
lo
w
-
laten
cy
in
f
e
r
en
ce
i
s
cr
u
cial
d
u
e
to
th
is
s
p
ee
d
b
o
o
s
t.
T
h
e
s
u
g
g
ested
m
eth
o
d
s
h
o
wed
b
etter
m
em
o
r
y
ec
o
n
o
m
y
b
y
u
tili
zin
g
b
in
ar
y
q
u
an
tizatio
n
an
d
im
p
r
o
v
em
en
ts
lik
e
weig
h
t c
lip
p
in
g
an
d
p
o
in
t
wis
e
co
n
v
o
lu
tio
n
s
.
T
h
ese
im
p
r
o
v
em
e
n
ts
lo
wer
th
e
m
em
o
r
y
a
n
d
p
ar
am
ete
r
r
e
q
u
ir
em
en
ts
,
wh
ich
is
ess
en
tia
l
f
o
r
d
ep
lo
y
m
en
t
in
s
itu
atio
n
s
with
lim
ited
r
eso
u
r
c
es,
in
clu
d
in
g
em
b
ed
d
e
d
s
y
s
te
m
s
an
d
I
o
T
d
ev
ices.
W
ith
its
s
tr
ea
m
lin
ed
d
esig
n
,
th
e
s
u
g
g
ested
B
NN
m
o
d
el
d
e
m
o
n
s
tr
ated
r
esil
ien
ce
wh
en
h
an
d
lin
g
co
m
p
licated
d
atasets
lik
e
C
I
FAR
-
1
0
in
ad
d
itio
n
to
o
u
tp
er
f
o
r
m
in
g
ea
r
lier
s
tu
d
ies
in
ter
m
s
o
f
ac
cu
r
ac
y
an
d
in
f
e
r
en
ce
tim
e.
T
h
is
illu
s
tr
ates
h
o
w
th
e
m
o
d
el
ca
n
g
r
o
w
t
o
in
cr
ea
s
i
n
g
ly
d
if
f
icu
lt
jo
b
s
with
o
u
t
n
ee
d
in
g
u
n
n
ec
ess
ar
ily
b
ig
h
ar
d
war
e
o
r
n
etwo
r
k
r
eso
u
r
ce
s
.
T
h
e
f
i
n
d
in
g
s
d
e
m
o
n
s
tr
ate
th
e
u
s
ef
u
ln
ess
o
f
im
p
lem
en
tin
g
B
NNs
in
r
e
al
-
tim
e
ap
p
licatio
n
s
,
in
clu
d
in
g
s
m
ar
t
I
o
T
d
e
v
ices,
au
to
n
o
m
o
u
s
s
y
s
tem
s
,
an
d
m
ed
ical
d
iag
n
o
s
tics
.
T
h
e
s
y
s
tem
ca
n
m
ee
t
th
e
p
er
f
o
r
m
an
ce
r
e
q
u
ir
em
e
n
ts
o
f
ed
g
e
co
m
p
u
tin
g
s
ce
n
a
r
io
s
wh
er
e
en
er
g
y
an
d
co
m
p
u
te
r
eso
u
r
ce
s
ar
e
co
n
s
tr
ain
ed
b
y
in
cr
ea
s
in
g
ef
f
icien
cy
,
d
ec
r
e
asin
g
in
f
er
en
ce
tim
e,
an
d
i
m
p
r
o
v
in
g
ac
c
u
r
ac
y
.
Fu
tu
r
e
r
esear
ch
ca
n
co
n
ce
n
tr
ate
o
n
ex
p
a
n
d
in
g
th
e
B
NN
ar
ch
itectu
r
e
to
m
o
r
e
co
m
p
licated
d
atasets
an
d
ap
p
licatio
n
s
,
ev
en
th
o
u
g
h
th
e
s
u
g
g
ested
s
y
s
tem
s
h
o
w
n
r
em
ar
k
ab
le
r
esu
lts
.
T
o
p
r
o
v
i
d
e
r
esil
ien
ce
ag
ain
s
t
s
u
ch
v
u
ln
er
ab
ilit
ies,
it
wo
u
ld
also
b
e
ess
en
tia
l
to
in
v
esti
g
at
e
s
ec
u
r
ity
an
d
p
r
iv
ac
y
is
s
u
es
wh
en
im
p
lem
en
tin
g
B
NNs
in
r
ea
l
-
tim
e
ap
p
licatio
n
s
.
T
o
in
cr
ea
s
e
th
e
s
y
s
tem
'
s
f
lex
ib
ilit
y
an
d
s
ca
lab
ilit
y
,
m
o
r
e
r
esear
ch
o
n
ad
ap
tiv
e
B
NNs
—
wh
ich
ca
n
d
y
n
am
ically
ad
ap
t
to
v
a
r
io
u
s
e
d
g
e
d
ev
ices
with
d
iv
er
s
e
p
r
o
c
ess
in
g
ca
p
ac
ities
—
will
b
e
cr
u
cial.
I
n
s
u
m
m
a
r
y
,
th
e
ac
cu
r
ac
y
a
n
d
ef
f
icien
cy
o
f
th
e
s
u
g
g
ested
B
NN
s
y
s
t
em
o
n
th
e
PYNQ
Z
2
FP
GA
h
av
e
s
ig
n
if
ican
tly
im
p
r
o
v
ed
.
T
h
e
o
u
tco
m
es
h
ig
h
lig
h
t BNNs
'
p
r
o
m
is
e
f
o
r
e
d
g
e
c
o
m
p
u
tin
g
ap
p
licatio
n
s
,
wh
er
e
h
ig
h
ac
cu
r
ac
y
,
q
u
ick
in
f
er
en
ce
s
p
ee
d
s
,
a
n
d
lo
w
p
o
we
r
co
n
s
u
m
p
ti
o
n
ar
e
c
r
itical.
ACK
NO
WL
E
DG
E
M
E
NT
S
W
e
wis
h
to
co
n
f
ir
m
th
at
n
o
k
n
o
wn
co
n
f
licts
o
f
in
ter
est
ar
e
ass
o
ciate
d
with
th
is
p
u
b
licati
o
n
an
d
al
l
th
e
au
th
o
r
s
h
av
e
co
n
tr
ib
u
ted
e
q
u
ally
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
3
,
Sep
tem
b
er
20
25
:
1
8
1
5
-
1
8
2
5
1824
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
t
r
ib
u
to
r
R
o
les
T
a
x
o
n
o
m
y
(
C
R
ed
iT
)
to
r
ec
o
g
n
ize
in
d
iv
i
d
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
R
am
y
a
B
an
av
ar
a
Nar
ay
an
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Seem
a
Sin
g
h
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t
th
e
f
i
n
d
in
g
s
o
f
th
is
s
tu
d
y
ar
e
av
aila
b
le
o
n
r
eq
u
est
f
r
o
m
th
e
co
r
r
esp
o
n
d
in
g
au
th
o
r
.
RE
F
E
R
E
NC
E
S
[
1
]
X
.
Z
h
a
o
,
L
.
W
a
n
g
,
Y
.
Z
h
a
n
g
,
X
.
H
a
n
,
M
.
D
e
v
e
c
i
,
a
n
d
M
.
P
a
r
m
a
r
,
“
A
r
e
v
i
e
w
o
f
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
i
n
c
o
m
p
u
t
e
r
v
i
s
i
o
n
,
”
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
R
e
v
i
e
w
,
v
o
l
.
5
7
,
n
o
.
4
,
p
.
9
9
,
M
a
r
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
0
4
6
2
-
0
2
4
-
1
0
7
2
1
-
6.
[
2
]
D
.
B
h
a
t
t
e
t
a
l
.
,
“
C
N
N
v
a
r
i
a
n
t
s
f
o
r
c
o
mp
u
t
e
r
v
i
si
o
n
:
h
i
s
t
o
r
y
,
a
r
c
h
i
t
e
c
t
u
r
e
,
a
p
p
l
i
c
a
t
i
o
n
,
c
h
a
l
l
e
n
g
e
s
a
n
d
f
u
t
u
r
e
sc
o
p
e
,
”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
0
,
n
o
.
2
0
,
p
.
2
4
7
0
,
O
c
t
.
2
0
2
1
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s
1
0
2
0
2
4
7
0
.
[
3
]
H
.
L
v
a
n
d
Q
.
W
u
,
“
A
n
e
n
e
r
g
y
-
e
f
f
i
c
i
e
n
t
f
i
e
l
d
-
p
r
o
g
r
a
m
ma
b
l
e
g
a
t
e
a
r
r
a
y
(
F
P
G
A
)
i
mp
l
e
m
e
n
t
a
t
i
o
n
o
f
a
r
e
a
l
-
t
i
m
e
p
e
r
sp
e
c
t
i
v
e
-
n
-
p
o
i
n
t
so
l
v
e
r
,
”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
3
,
n
o
.
1
9
,
p
.
3
8
1
5
,
S
e
p
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s
1
3
1
9
3
8
1
5
.
[
4
]
I
.
D
.
M
i
e
n
y
e
a
n
d
T.
G
.
S
w
a
r
t
,
“
A
c
o
mp
r
e
h
e
n
s
i
v
e
r
e
v
i
e
w
o
f
d
e
e
p
l
e
a
r
n
i
n
g
:
a
r
c
h
i
t
e
c
t
u
r
e
s
,
r
e
c
e
n
t
a
d
v
a
n
c
e
s
,
a
n
d
a
p
p
l
i
c
a
t
i
o
n
s,
”
I
n
f
o
rm
a
t
i
o
n
,
v
o
l
.
1
5
,
n
o
.
1
2
,
p
.
7
5
5
,
N
o
v
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
i
n
f
o
1
5
1
2
0
7
5
5
.
[
5
]
R
.
B
u
e
n
r
o
s
t
r
o
-
M
a
r
i
sca
l
,
P
.
C
.
S
a
n
t
a
n
a
-
M
a
n
c
i
l
l
a
,
O
.
A
.
M
o
n
t
e
s
i
n
o
s
-
L
o
p
e
z
,
J.
I
.
N
i
e
t
o
H
i
p
o
l
i
t
o
,
a
n
d
L.
E.
A
n
i
d
o
-
R
i
f
o
n
,
“
A
r
e
v
i
e
w
o
f
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
l
i
c
a
t
i
o
n
s
f
o
r
t
h
e
n
e
x
t
g
e
n
e
r
a
t
i
o
n
o
f
c
o
g
n
i
t
i
v
e
n
e
t
w
o
r
k
s,
”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
2
,
n
o
.
1
2
,
p
.
6
2
6
2
,
J
u
n
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
2
1
2
6
2
6
2
.
[
6
]
S
.
C
h
o
i
,
Y
.
Je
o
n
,
a
n
d
Y
.
S
e
o
,
“
H
i
g
h
-
p
e
r
f
o
r
ma
n
c
e
a
n
d
r
o
b
u
st
b
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
a
c
c
e
l
e
r
a
t
o
r
b
a
se
d
o
n
mo
d
i
f
i
e
d
c
o
n
t
e
n
t
-
a
d
d
r
e
ssa
b
l
e
m
e
m
o
r
y
,
”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
1
,
n
o
.
1
7
,
p
.
2
7
8
0
,
S
e
p
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s
1
1
1
7
2
7
8
0
.
[
7
]
L.
M
.
M
e
y
e
r
,
M
.
Z
a
ma
n
i
a
n
d
A
.
D
e
m
o
st
h
e
n
o
u
s
,
"
B
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
s
f
o
r
r
e
so
u
r
c
e
-
e
f
f
i
c
i
e
n
t
sp
i
k
e
so
r
t
i
n
g
,
"
i
n
I
EE
E
Ac
c
e
ss,
v
o
l
.
1
3
,
p
p
.
6
0
2
5
8
-
6
0
2
6
9
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
a
c
c
e
s
s.
2
0
2
5
.
3
5
5
2
0
0
8
.
[
8
]
T.
R
o
z
e
n
,
M
.
K
i
mh
i
,
B
.
C
h
mi
e
l
,
A
.
M
e
n
d
e
l
s
o
n
,
a
n
d
C
.
B
a
s
k
i
n
,
“
B
i
mo
d
a
l
-
d
i
st
r
i
b
u
t
e
d
b
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
M
a
t
h
e
m
a
t
i
c
s
,
v
o
l
.
1
0
,
n
o
.
2
1
,
p
.
4
1
0
7
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
ma
t
h
1
0
2
1
4
1
0
7
.
[
9
]
Q
.
H
.
V
o
,
F
.
A
s
i
m
,
B
.
A
l
i
m
k
h
a
n
u
l
y
,
S
.
L
e
e
,
a
n
d
L
.
K
i
m
,
“
H
a
r
d
w
a
r
e
p
l
a
t
f
o
r
m
-
a
w
a
r
e
b
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
m
o
d
e
l
o
p
t
i
m
i
z
a
t
i
o
n
,”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
2
,
n
o
.
3
,
p
.
1
2
9
6
,
J
a
n
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
2
0
3
1
2
9
6
.
[
1
0
]
M
.
S
.
L
e
,
T
.
N
.
P
h
a
m,
T
.
D
.
N
g
u
y
e
n
,
a
n
d
I
.
J.
C
h
a
n
g
,
“
A
v
a
r
i
a
t
i
o
n
-
a
w
a
r
e
b
i
n
a
r
y
n
e
u
r
a
l
n
e
t
w
o
r
k
f
r
a
m
e
w
o
r
k
f
o
r
p
r
o
c
e
ss
r
e
s
i
l
i
e
n
t
in
-
me
mo
r
y
c
o
m
p
u
t
a
t
i
o
n
s
,
”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
3
,
n
o
.
1
9
,
p
.
3
8
4
7
,
S
e
p
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s1
3
1
9
3
8
4
7
.
[
1
1
]
Y
.
S
u
,
K
.
P
.
S
e
n
g
,
L
.
M
.
A
n
g
,
a
n
d
J
.
S
m
i
t
h
,
“
B
i
n
a
r
y
n
e
u
r
a
l
n
e
t
w
o
r
k
s
i
n
F
P
G
A
s
:
a
r
c
h
i
t
e
c
t
u
r
e
s
,
t
o
o
l
f
l
o
w
s
a
n
d
h
a
r
d
w
a
r
e
c
o
m
p
a
r
i
s
o
n
s
,
”
S
e
n
s
o
r
s
,
v
o
l
.
2
3
,
n
o
.
2
2
,
p
.
9
2
5
4
,
N
o
v
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
s
2
3
2
2
9
2
5
4
.
[
1
2
]
N
.
P
h
i
p
p
s,
J
.
J.
S
h
a
n
g
,
T.
H
.
Te
o
,
a
n
d
I
.
C
.
W
e
y
,
“
P
r
e
-
c
o
m
p
u
t
i
n
g
b
a
t
c
h
n
o
r
mal
i
sat
i
o
n
p
a
r
a
me
t
e
r
s
f
o
r
e
d
g
e
d
e
v
i
c
e
s
o
n
a
b
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
,
”
S
e
n
s
o
rs
,
v
o
l
.
2
3
,
n
o
.
1
2
,
p
.
5
5
5
6
,
J
u
n
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
s2
3
1
2
5
5
5
6
.
[
1
3
]
F
.
S
a
k
r
e
t
a
l
.
,
“
C
B
i
n
-
N
N
:
a
n
i
n
f
e
r
e
n
c
e
e
n
g
i
n
e
f
o
r
b
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
3
,
n
o
.
9
,
p
.
1
6
2
4
,
A
p
r
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s1
3
0
9
1
6
2
4
.
[
1
4
]
J.
W
.
Le
e
e
t
a
l
.
,
“
P
r
i
v
a
c
y
-
p
r
e
ser
v
i
n
g
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
w
i
t
h
f
u
l
l
y
h
o
mo
mo
r
p
h
i
c
e
n
c
r
y
p
t
i
o
n
f
o
r
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
,
”
I
EEE
A
c
c
e
ss
,
v
o
l
.
1
0
,
p
p
.
3
0
0
3
9
–
3
0
0
5
4
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
2
.
3
1
5
9
6
9
4
.
[
1
5
]
Y
.
Zh
a
n
g
,
J.
P
a
n
,
X
.
Li
u
,
H
.
C
h
e
n
,
D
.
C
h
e
n
,
a
n
d
Z.
Z
h
a
n
g
,
“
F
r
a
c
B
N
N
:
a
c
c
u
r
a
t
e
a
n
d
F
P
G
A
-
e
f
f
i
c
i
e
n
t
b
i
n
a
r
y
n
e
u
r
a
l
n
e
t
w
o
r
k
s
w
i
t
h
f
r
a
c
t
i
o
n
a
l
a
c
t
i
v
a
t
i
o
n
s,
”
i
n
T
h
e
2
0
2
1
A
C
M/
S
I
G
D
A
I
n
t
e
rn
a
t
i
o
n
a
l
S
y
m
p
o
si
u
m
o
n
Fi
e
l
d
-
Pr
o
g
r
a
m
m
a
b
l
e
G
a
t
e
Arr
a
y
s
,
N
e
w
Y
o
r
k
,
N
Y
,
U
S
A
:
A
C
M
,
F
e
b
.
2
0
2
1
,
p
p
.
1
7
1
–
1
8
2
.
d
o
i
:
1
0
.
1
1
4
5
/
3
4
3
1
9
2
0
.
3
4
3
9
2
9
6
.
[
1
6
]
V
.
S
a
l
a
u
y
o
u
,
“
A
r
e
a
a
n
d
p
e
r
f
o
r
ma
n
c
e
e
st
i
mat
e
s
o
f
f
i
n
i
t
e
st
a
t
e
m
a
c
h
i
n
e
s
i
n
r
e
c
o
n
f
i
g
u
r
a
b
l
e
sy
s
t
e
ms
,
”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
4
,
n
o
.
2
4
,
p
.
1
1
8
3
3
,
D
e
c
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
4
2
4
1
1
8
3
3
.
[
1
7
]
S
.
Z
h
u
,
L.
H
.
K
.
D
u
o
n
g
,
a
n
d
W
.
Li
u
,
“
TA
B
:
u
n
i
f
i
e
d
a
n
d
o
p
t
i
mi
z
e
d
t
e
r
n
a
r
y
,
b
i
n
a
r
y
,
a
n
d
mi
x
e
d
-
p
r
e
c
i
si
o
n
n
e
u
r
a
l
n
e
t
w
o
r
k
i
n
f
e
r
e
n
c
e
o
n
t
h
e
e
d
g
e
,
”
A
C
M
T
r
a
n
sa
c
t
i
o
n
s
o
n
E
m
b
e
d
d
e
d
C
o
m
p
u
t
i
n
g
S
y
st
e
m
s
,
v
o
l
.
2
1
,
n
o
.
5
,
p
p
.
1
–
2
6
,
S
e
p
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
4
5
/
3
5
0
8
3
9
0
.
[1
8]
F
.
Lu
o
,
S
.
K
h
a
n
,
Y
.
H
u
a
n
g
,
a
n
d
K
.
W
u
,
“
B
i
n
a
r
i
z
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
e
d
g
e
i
n
t
e
l
l
i
g
e
n
c
e
o
f
se
n
s
o
r
-
b
a
se
d
h
u
m
a
n
a
c
t
i
v
i
t
y
re
c
o
g
n
i
t
i
o
n
,
”
I
EEE
T
r
a
n
sa
c
t
i
o
n
s
o
n
Mo
b
i
l
e
C
o
m
p
u
t
i
n
g
,
v
o
l
.
2
2
,
n
o
.
3
,
p
p
.
1
3
5
6
–
1
3
6
8
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TM
C
.
2
0
2
1
.
3
1
0
9
9
4
0
.
[
1
9
]
K
.
S
h
i
,
M
.
W
a
n
g
,
X
.
Ta
n
,
Q
.
L
i
,
a
n
d
T.
Le
i
,
“
Ef
f
i
c
i
e
n
t
d
y
n
a
m
i
c
r
e
c
o
n
f
i
g
u
r
a
b
l
e
C
N
N
a
c
c
e
l
e
r
a
t
o
r
f
o
r
e
d
g
e
i
n
t
e
l
l
i
g
e
n
c
e
c
o
mp
u
t
i
n
g
o
n
F
P
G
A
,
”
I
n
f
o
rm
a
t
i
o
n
,
v
o
l
.
1
4
,
n
o
.
3
,
p
.
1
9
4
,
M
a
r
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
i
n
f
o
1
4
0
3
0
1
9
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.