Ad
alin
e a
nd M
adal
ine
Sign
al p
roce
ssin
g de
velo
ped
as a
n en
gine
erin
g di
scip
line
wit
h th
e ad
vent
of
elec
tron
ic c
omm
unic
atio
n. I
niti
ally
, ana
log
filt
ers
usin
gito
r (R
LC
) ci
rcui
ts w
ere
desi
gned
to
rem
ove
nois
e fr
om t
he c
omm
unic
atio
nsi
gnal
s. T
oday
, si
gnal
pro
cess
ing
has
evol
ved
into
a m
any-
face
ted
tech
nolo
gy,
wit
h th
e em
phas
is h
avin
g sh
ifte
d fr
om t
uned
cir
cuit
impl
emen
tatio
n to
dig
ital
sign
al p
roce
ssor
s (D
SPs)
tha
t ca
n pe
rfor
m t
he s
ame
type
s of
fil
teri
ng a
pplic
a-tio
ns b
y ex
ecut
ing
conv
olut
ion
filte
rs i
mpl
emen
ted
in s
oftw
are.
Th
e ba
sis
for
the
indu
stry
rem
ains
the
des
ign
and
impl
emen
tatio
n of
filt
ers
to p
erfo
rm n
oise
rem
oval
fro
m i
nfor
mat
ion-
bear
ing
sign
als.
In t
his
chap
ter,
we
will
foc
us o
n a
spec
ific
typ
e of
filt
er,
calle
d th
e A
da-
line
(and
the
mul
tiple
-Ada
line,
or
Mad
alin
e) d
evel
oped
by
Ber
nard
Wid
row
of
Stan
ford
Uni
vers
ity.
As
we
wil
l se
e, t
he A
dalin
e m
odel
is
sim
ilar
to t
hat
of a
sing
le P
E i
n an
AN
S.
2.1
RE
VIE
W O
F S
IGN
AL
PR
OC
ES
SIN
G
We
begi
n ou
r di
scus
sion
of
the
Ada
line
and
Mad
alin
e ne
twor
ks w
ith
a re
view
of b
asic
sig
nal-
proc
essi
ng t
heor
y.
An
unde
rsta
ndin
g of
thi
s m
ater
ial
is e
ssen
-tia
l if
we
are
to a
ppre
ciat
e th
e op
erat
ion
and
appl
icat
ions
of
thes
e ne
twor
ks.
How
ever
, th
is m
ater
ial
is a
lso
typi
call
y co
vere
d as
par
t of
an
unde
rgra
duat
ecu
rric
ulum
in
info
rmat
ion
codi
ng a
nd d
ata
com
mun
icat
ion.
T
here
fore
, re
ader
sal
read
y co
mfo
rtab
le w
ith
sign
al-p
roce
ssin
g co
ncep
ts m
ay s
kip
this
fir
st s
ectio
nw
itho
ut f
ear
of m
issi
ng m
ater
ial
rele
vant
to
the
Ada
line
and
Mad
alin
e to
pics
.Fo
r th
ose
read
ers
who
ar
e no
t fa
mil
iar
wit
h th
e te
chni
ques
com
mon
ly u
sed
to i
mpl
emen
t el
ectr
onic
com
mun
icat
ions
and
sig
nal
proc
essi
ng,
we
shal
l be
-gi
n by
des
crib
ing
brie
fly
the
data
-enc
odin
g an
d m
odul
atio
n sc
hem
es u
sed
in a
nam
plit
ude-
mod
ulat
ion
(AM
) ra
dio
tran
smis
sion
. A
s pa
rt o
f th
is d
iscu
ssio
n, w
esh
all
illu
stra
te t
he n
eed
for
filte
rs i
n th
e co
mm
unic
atio
ns i
ndus
try.
We
will
the
n
45
46
Ad
alin
e a
nd
revi
ew t
he c
once
pts
of t
he f
requ
ency
dom
ain,
the
fou
r ba
sic
filt
er t
ypes
, an
dFo
urie
r an
alys
is.
Thi
s pr
elim
inar
y se
ctio
n co
nclu
des
with
a b
rief
ove
rvie
w o
fdi
gita
l si
gnal
pro
cess
ing,
bec
ause
man
y of
the
conc
epts
rea
lize
d in
dig
ital
fil
ters
are
dire
ctly
app
lica
ble
to t
he A
dali
ne a
nd M
adal
ine
(and
man
y ot
her)
neu
ral
netw
orks
.
Si
gnal
Pro
cess
ing
and
Filte
rs
Sig
nal
proc
essi
ng i
s an
eng
inee
ring
dis
cipl
ine
that
dea
ls p
rim
aril
y w
ith
the
im-
plem
enta
tion
of f
ilte
rs t
o re
mov
e or
red
uce
unw
ante
d fr
eque
ncy
com
pone
nts
from
an
info
rmat
ion-
bear
ing
sign
al.
Let
's c
onsi
der,
for
exa
mpl
e, a
n A
M r
a-di
o br
oadc
ast.
Ele
ctro
nic
com
mun
icat
ion
tech
niqu
es, w
heth
er f
or a
udio
sig
nals
or o
ther
dat
a, c
onsi
st o
f si
gnal
enc
odin
g an
d m
odul
atio
n.
Info
rmat
ion
to b
e t
his
case
, au
dibl
e so
unds
, su
ch a
s vo
ice
or b
e en
-co
ded
elec
tron
ical
ly b
y an
ana
log
sign
al t
hat
exac
tly r
epro
duce
s th
e fr
eque
ncie
san
d am
plit
udes
of t
he o
rigi
nal s
ound
s. S
ince
the
soun
ds b
eing
enc
oded
rep
rese
nta
cont
inuu
m f
rom
sile
nce
thro
ugh
voic
e to
mus
ic,
the
inst
anta
neou
s fr
eque
ncy
of t
he e
ncod
ed s
igna
l w
ill
vary
wit
h ti
me,
ra
ngin
g fr
om 0
to
appr
oxim
atel
y10
,000
her
tz (
Hz)
.R
athe
r th
an a
ttem
pt t
o tr
ansm
it th
is e
ncod
ed s
igna
l di
rect
ly,
we
tran
sfor
mth
e si
gnal
int
o a
form
mor
e su
itabl
e fo
r ra
dio
tran
smis
sion
. W
e ac
com
plis
h th
istr
ansf
orm
atio
n by
mod
ulat
ing
the
ampl
itud
e of
a h
igh-
freq
uenc
y ca
rrie
r si
gnal
wit
h th
e an
alog
inf
orm
atio
n si
gnal
. T
his
proc
ess
is i
llus
trat
ed i
n Fi
gure
2.1
.H
ere,
th
e ca
rrie
r is
not
hing
mor
e th
an a
sin
e w
ave
wit
h a
freq
uenc
y m
uch
grea
ter
than
the
inf
orm
atio
n si
gnal
. Fo
r A
M r
adio
, th
e ca
rrie
r fr
eque
ncy
wil
l be
in t
he r
ange
of
550
to (
KH
z).
Sinc
e th
e fr
eque
ncy
of t
he c
arri
eris
sig
nifi
cant
ly g
reat
er t
han
is t
he m
axim
um f
requ
ency
of t
he i
nfor
mat
ion
sign
al,
litt
le i
nfor
mat
ion
is l
ost
by t
his
mod
ulat
ion.
T
he m
odul
ated
sig
nal
can
then
be
tran
smit
ted
to a
rec
eivi
ng s
tati
on (
or b
road
cast
to a
nyon
e w
ith
a ra
dio
rece
iver
),w
here
the
sig
nal
is d
emod
ulat
ed a
nd i
s re
prod
uced
as
soun
d.T
he m
ost o
bvio
us r
easo
n fo
r a
filt
er in
AM
rad
io is
tha
t dif
fere
nt p
eopl
e ha
vedi
ffer
ent p
refe
renc
es i
n m
usic
and
ent
erta
inm
ent.
The
refo
re,
the
gove
rnm
ent
and
the
com
mun
icat
ion
indu
stry
hav
e al
low
ed m
any
diff
eren
t ra
dio
stat
ions
to
op-
erat
e in
the
sam
e ge
ogra
phic
al a
rea,
so
that
eve
ryon
e's
tast
es i
n en
tert
ainm
ent
can
be a
ccom
mod
ated
. W
ith s
o m
any
diff
eren
t ra
dio
stat
ions
all
broa
dcas
ting
in c
lose
pro
xim
ity,
how
is
it th
at w
e ca
n li
sten
to
only
one
sta
tion
at a
tim
e?Th
e an
swer
is
to a
llow
eac
h re
ceiv
er t
o be
tun
ed b
y th
e us
er t
o a
sele
ctab
le f
re-
quen
cy.
In t
unin
g th
e ra
dio,
we
are
esse
ntia
lly c
hang
ing
the
freq
uenc
y-re
spon
sech
arac
teri
stic
s of
a b
andp
ass
filt
er i
nsid
e th
e ra
dio.
T
his
filt
er a
llow
s on
ly t
hesi
gnal
s fr
om t
he s
tatio
n in
whi
ch w
e ar
e in
tere
sted
to
pass
, w
hile
eli
min
atin
gal
l th
e ot
her
sign
als
bein
g br
oadc
ast
wit
hin
the
spec
trum
of
the
AM
rad
io.
To i
llus
trat
e ho
w t
he b
andp
ass
filt
er o
pera
tes,
we
wil
l ch
ange
our
ref
eren
cefr
om t
he t
ime
dom
ain
to t
he f
requ
ency
dom
ain.
W
e be
gin
by c
onst
ruct
ing
atw
o-ax
is g
raph
, w
here
the
x a
xis
repr
esen
ts i
ncre
asin
g fr
eque
ncie
s an
d th
e y
axis
rep
rese
nts
decr
easi
ng a
tten
uati
on i
n a
unit
cal
led
the
deci
bel
(dB
). S
uch
a
2.1
Rev
iew
of
Sig
nal
Pro
cess
ing
47
1.0
0.5
(a)
-0.5
-1.0
0.05
0.
10.15
0.2
Car
rier
wav
e
(b)
(c)
0.05
0.1
0.15
Wav
e co
ntai
ning
info
rmat
ion
0.05
0.1
0.15
0.2
1.1
0.9
0.8
0.7
1.5
1.0
0.5
-0.5
-1.0
-1.5
Am
plitu
de-m
odul
ated
wav
e
Figu
re 2
.1
Typ
ical
in
form
atio
n-en
codi
ng
and
ampl
itude
-mod
ulat
ion
tech
niq
ues
for
ele
ctro
nic
com
munic
atio
n a
re
(a)
The
carr
ier
wave
has
a f
requency
muc
h h
ighe
r th
an t
hat
of
(b)
the
info
rmatio
n-b
earing
(c
) T
he c
arr
ier
wave
is
mod
ulat
edby
th
e i
nfo
rma
tion
-be
ari
ng s
ignal.
0.2
48
Ad
alin
e a
nd
Mad
alin
e
grap
h is
ill
ustr
ated
in
Figu
re 2
.2(a
). F
or t
he A
M r
adio
exa
mpl
e, l
et u
s im
agin
eth
at t
here
are
sev
en A
M r
adio
sta
tions
, la
bele
d A
thr
ough
G,
oper
atin
g in
the
area
whe
re w
e ar
e li
sten
ing.
T
he f
requ
enci
es a
t w
hich
the
se s
tatio
ns t
rans
mit
are
grap
hed
as v
erti
cal
line
s lo
cate
d on
the
fre
quen
cy a
xis
at t
he p
oint
cor
re-
spon
ding
to
thei
r tr
ansm
itti
ng,
or c
arri
er,
freq
uenc
y. T
he a
mpl
itude
of
the
lines
,as
ill
ustr
ated
in
Figu
re 2
.2(a
), i
s al
mos
t 0
dB,
indi
catin
g th
at e
ach
stat
ion
istr
ansm
itti
ng a
t fu
ll p
ower
, an
d ea
ch c
an b
e re
ceiv
ed e
qual
ly w
ell.
Now
we
wil
l a
ban
dpas
s fi
lter
to s
elec
t on
e of
the
sev
en s
tatio
ns.
The
freq
uenc
y re
spon
se o
f a
typi
cal
band
pass
filt
er i
s ill
ustr
ated
in
Figu
re 2
.2(b
).N
otic
e th
at t
he f
requ
ency
-res
pons
e cu
rve
is s
uch
that
all
freq
uenc
ies
that
fal
lou
tsid
e th
e in
vert
ed n
otch
are
atte
nuat
ed t
o ve
ry s
mal
l m
agni
tude
s, w
here
asfr
eque
ncie
s w
ithin
the
pass
band
are
allo
wed
to p
ass
with
ver
y lit
tlehe
nce
the
nam
e "b
andp
ass
filte
r."
To t
une
our
radi
o re
ceiv
er t
o an
y on
e of
the
seve
n br
oadc
astin
g st
atio
ns,
we
sim
ply
adju
st t
he f
requ
ency
res
pons
e of
the
filte
r su
ch t
hat
the
carr
ier
freq
uenc
y of
the
des
ired
sta
tion
is w
ithin
the
pass
band
. CD C o i
0-
f
[ 3
C
[
EE
F- C B
500
700
900
1300
Frequency
(KHz)
1500
17
00
(b)
CO
Figu
re 2
.2
-20
500
700
900
1300
1500
1700
Fre
quen
cy (
KH
z)
Th
ese
are
fre
qu
en
cy d
omai
n g
rap
hs
of (
a) A
M r
ad
io r
ecep
tion
of s
even
d
iffe
ren
t st
atio
ns,
(b)
the
fre
qu
en
cy r
espo
nse
of t
hetu
nin
g fil
ter
and
the
mag
nitu
de o
f th
e re
ceiv
ed
sig
na
ls a
fte
rfilte
rin
g.
2.1
Rev
iew
of
Sig
nal
Pro
cess
ing
49
As
anot
her
exam
ple
of t
he u
se o
f fi
lters
in
the
com
mun
icat
ion
indu
stry
,co
nsid
er t
he p
robl
em o
f ec
ho s
uppr
essi
on i
n lo
ng-d
ista
nce
tele
phon
e co
mm
u-ni
catio
n.
As
indi
cate
d in
Fig
ure
2.3,
the
pro
blem
is
caus
ed b
y th
e in
tera
ctio
nbe
twee
n th
e am
plif
iers
and
ser
ies
coup
ling
use
d on
bot
h en
ds o
f th
e li
ne,
and
the
dela
y tim
e re
quir
ed t
o tr
ansm
it t
he v
oice
inf
orm
atio
n be
twee
n th
e sw
itch
-in
g of
fice
and
the
com
mun
icat
ions
sat
ellit
e in
geo
stat
iona
ry o
rbit
, 23
,000
mil
esab
ove
the
eart
h.
Spec
ific
ally
, yo
u he
ar a
n ec
ho o
f yo
ur o
wn
voic
e in
the
tel
e-ph
one
whe
n yo
u sp
eak.
The
sig
nal
carr
ying
you
r vo
ice
arri
ves
at t
he r
ecei
ving
tele
phon
e ap
prox
imat
ely
270
mill
isec
onds
aft
er y
ou s
peak
. Th
is d
elay
is
the
amou
nt o
f tim
e re
quir
ed b
y th
e m
icro
wav
e si
gnal
to
trav
el t
he 4
6,00
0 m
iles
betw
een
the
tran
smitt
ing
stat
ion,
the
sat
ellit
e, a
nd t
he r
ecei
ving
sta
tion
on t
hegr
ound
. O
nce
rece
ived
and
rou
ted
to t
he d
estin
atio
n te
leph
one,
the
sig
nal
isag
ain
ampl
ifie
d an
d re
prod
uced
as
soun
d on
the
rec
eivi
ng h
ands
et.
Unf
ortu
-na
tely
, it
is a
lso
ofte
n pi
cked
up
by t
he t
rans
mitt
er a
t th
e re
ceiv
ing
end,
due
to i
mpe
rfec
tions
in
the
devi
ces
used
to
deco
uple
the
inc
omin
g si
gnal
s.
It c
anth
en b
e a
nd f
ed b
ack
to y
ou a
ppro
xim
atel
y 1/
2 se
cond
aft
er y
ousp
oke.
The
res
ult
is e
cho.
O
bvio
usly
, a
sim
ple
band
pass
filt
er c
anno
t be
use
dto
rem
ove
the
echo
, be
caus
e th
ere
is n
o w
ay t
o di
stin
guis
h th
e ec
hoed
sig
nal
from
val
id s
igna
ls.
To s
olve
pro
blem
s su
ch a
s th
ese,
the
com
mun
icat
ions
ind
ustr
y ha
s de
vel-
oped
man
y di
ffer
ent
type
s of
filt
ers.
Th
ese
filte
rs n
ot o
nly
are
used
in
23,3
00 m
iles
Inco
min
g si
gnal
retra
nsm
itted
due
to c
oupl
ing
leak
age
Ret
urni
ng s
igna
l is
orig
inal
dela
yed
by 5
00 m
illis
econ
ds,
resu
lting
in a
n "e
cho"
2.3
E
cho
ca
n o
ccu
r in
lo
ng
-dis
tan
ce t
ele
com
mu
nic
atio
ns.
50A
dal
ine a
nd
Mad
alin
e
tron
ic c
omm
unic
atio
ns,
but
also
hav
e an
app
licat
ion
base
tha
t in
clud
es r
adar
and
sona
r im
agin
g, e
lect
roni
c w
arfa
re,
and
med
ical
tec
hnol
ogy.
H
owev
er,
all
the
appl
icat
ion-
spec
ific
fi
lter
im
plem
enta
tions
can
be
grou
ped
into
fou
r ge
n-er
al f
ilte
r ty
pes:
lo
wpa
ss,
ban
dpas
s, a
nd b
ands
top.
The
cha
ract
eris
ticfr
eque
ncy
resp
onse
of
the
se
filt
ers
is
depi
cted
in
Fi
gure
2.
4.
The
ad
aptiv
efi
lter
, w
hich
is
the
su
bjec
t of
the
rem
aind
er o
f th
e ch
apte
r,
has
char
acte
ris-
tics
uni
que
to t
he a
ppli
cati
on i
t se
rves
. It
can
rep
rodu
ce t
he c
hara
cter
istic
s of
any
of t
he f
our
basi
c fi
lter
typ
es,
alon
e or
in
com
bina
tion.
A
s w
e sh
all
show
late
r, th
e ad
aptiv
e fi
lter
is i
deal
ly s
uite
d to
the
tel
epho
ne-e
cho
prob
lem
jus
tdi
scus
sed.
Low
pass
filt
er
Hig
hpas
s fil
ter
Ban
dpas
s fil
ter
Ban
dsto
p fil
ter
c t t c CD
Freq
uenc
y
Freq
uenc
y
Freq
uenc
y
Fre
quen
cy
Figu
re 2
.4
Fre
qu
en
cy-r
esp
on
se c
ha
ract
eri
stic
s of
the
fou
r ba
sic
filte
r ty
pes
are
sh
ow
n.
2.1
Rev
iew
of
Sig
nal
Pro
cess
ing
51
2.1.
2 F
ou
rier
An
alys
is a
nd
th
e F
req
uen
cy D
omai
n
To
anal
yze
a si
gnal
-pro
cess
ing
prob
lem
tha
t re
quir
es a
filt
er,
we
mus
t le
ave
the
tim
e do
mai
n an
d fi
nd a
too
l fo
r tr
ansl
atin
g ou
r fi
lter
mod
els
into
the
fre
quen
cydo
mai
n,
beca
use
mos
t of
the
si
gnal
s w
e w
ill
anal
yze
cann
ot b
e co
mpl
etel
yun
ders
tood
in
the
tim
e do
mai
n.
For
exam
ple,
mos
t si
gnal
s co
nsis
t no
t on
ly o
fa
fund
amen
tal
freq
uenc
y, b
ut a
lso
harm
onic
s th
at m
ust
be c
onsi
dere
d, o
r th
eyco
nsis
t of
man
y di
scre
te f
requ
ency
com
pone
nts
that
mus
t be
acc
ount
ed f
or b
yth
e fi
lters
we
desi
gn.
The
re a
re m
any
tool
s th
at w
e ca
n us
e to
hel
p un
ders
tand
the
freq
uenc
y-do
mai
n na
ture
of
sign
als.
One
of t
he m
ost
com
mon
ly u
sed
is t
heFo
urie
r se
ries.
It
has
bee
n sh
own
that
any
per
iodi
c si
gnal
can
be
mod
eled
as
an i
nfin
ite
serie
s of
sin
es a
nd c
osin
es.
The
Fou
rier
ser
ies,
whi
ch d
escr
ibes
the
freq
uenc
y-do
mai
n na
ture
of
perio
dic
sign
als,
is
give
n by
the
equ
atio
n
x(t)
= +
whe
re i
s th
e fu
ndam
enta
l fr
eque
ncy
of th
e si
gnal
in
the
time
dom
ain,
and
the
coef
fici
ents
, a
nd a
re n
eede
d to
mod
ulat
e th
e am
plitu
de o
f th
e in
divi
dual
term
s of
the
serie
s.T
his
serie
s is
use
ful
for
desc
ribi
ng t
he d
iscr
ete
freq
uenc
y co
mpo
nent
s th
atco
mpr
ise
a no
ntri
vial
per
iodi
c si
gnal
. A
s an
illu
stra
tion,
a s
quar
e w
ave
can
bede
com
pose
d in
to a
sum
mat
ion
of f
requ
ency
ele
men
ts c
onta
inin
g no
thin
g m
ore
than
sin
e w
aves
of
diff
eren
t am
plitu
de a
nd f
requ
ency
, as
is
illus
trat
ed i
n Fi
g-ur
e 2.
5. S
ince
a s
quar
e w
ave
is u
sefu
l for
repr
esen
ting
bina
ry in
form
atio
n in
dat
atr
ansm
issi
on,
it is
im
port
ant
that
we
unde
rsta
nd th
e fr
eque
ncy-
dom
ain
natu
re o
fsu
ch a
sig
nal.
From
ins
pect
ion
in t
he t
ime
dom
ain,
we
can
obse
rve
that
the
squa
re w
ave
is i
deal
ly s
uite
d to
bin
ary
data
rep
rese
ntat
ion
beca
use
ther
e ar
e tw
odi
stin
ct s
tate
s (a
1 a
nd a
0),
and
the
tran
sitio
n ti
me
betw
een
stat
es is
neg
ligib
le.
It i
s di
ffic
ult,
how
ever
, to
obt
ain
a pe
rfec
t sq
uare
wav
e in
any
pra
ctic
alel
ectr
onic
cir
cuit,
due
in
part
to
the
effe
cts
of t
he t
rans
mitt
ing
med
ia o
n th
esi
gnal
. T
o ill
ustr
ate
why
thi
s is
so,
con
side
r th
e Fo
urie
r se
ries
expa
nsio
n
x(t)
= +
- +
- +
whi
ch d
escr
ibes
a t
ypic
al s
quar
e w
ave.
As
illus
trat
ed i
n Fi
gure
2.5
, if
we
alge
brai
cally
add
tog
ethe
r th
e fi
rst
thre
esi
nuso
idal
com
pone
nts
of t
his
Four
ier
serie
s, w
e pr
oduc
e a
sign
al t
hat
alre
ady
stro
ngly
res
embl
es t
he s
quar
e w
ave.
How
ever
, w
e sh
ould
not
ice
that
the
res
ul-
tant
sig
nal
also
exh
ibit
s ri
pple
s in
bot
h ac
tive
regi
ons.
The
se r
ippl
es w
ill r
emai
nto
som
e ex
tent
, un
less
we
com
plet
e th
e in
fini
te s
erie
s.
Sinc
e th
at i
s ob
viou
sly
not
prac
tical
, w
e m
ust
even
tual
ly t
runc
ate
the
seri
es a
nd s
ettle
for
som
e am
ount
of r
ippl
e in
the
res
ulti
ng s
igna
l.It
tur
ns
out
that
thi
s tr
unca
tion
exa
ctly
cor
resp
onds
to
the
beha
vior
we
obse
rved
whe
n tr
ansm
itti
ng a
squ
are
wav
e ac
ross
an
elec
trom
agne
tic m
edia
. A
s i
s im
poss
ible
to
have
a m
ediu
m o
f in
fini
te b
andw
idth
, it
follo
ws
that
it
is t
o tr
ansm
it al
l th
e fr
eque
ncy
com
pone
nts
of a
squ
are
wav
e.
Thu
s,
52
Ad
alin
e a
nd
1.0J I
:
1
;5
0.1!-1
.5-1
.0-0
.5
0.25
0.5
1.0
1.5
0.5
1.0
1.5
-0.5
-0.7
5
0.2
0.1
0.5
1.0
1.5
Figu
re 2
.5
The
fir
st t
hree
fre
quen
cy-d
omai
n co
mpo
nent
s of
a s
quar
e w
ave
are
show
n.
Not
ice
that
the
sin
e w
aves
eac
h ha
ve d
iffer
ent
mag
nitu
des,
as
indi
cate
d by
the
coor
dina
tes
on th
e e
ven
thou
gh t
hey
are g
raph
ed t
o the
sam
e h
eigh
t.
whe
n w
e tr
ansm
it a
peri
odic
squ
are
wav
e, w
e ca
n ob
serv
e th
e fr
eque
ncy-
dom
ain
effe
cts
in t
he t
ime-
dom
ain
sign
al a
s ov
ersh
oot,
unde
rsho
ot,
and
ripp
le.
Thi
s ex
ampl
e sh
ows
that
the
Four
ier s
erie
s ca
n be
a p
ower
ful t
ool i
n he
lpin
gus
to
unde
rsta
nd t
he f
requ
ency
-dom
ain
natu
re o
f an
y pe
riodi
c si
gnal
, an
d to
pred
ict
ahea
d of
tim
e w
hat
tran
smis
sion
eff
ects
we
mus
t co
nsid
er a
s w
e de
sign
filte
rs f
or o
ur s
igna
l-pr
oces
sing
app
licat
ions
.W
e ca
n al
so a
pply
Fou
rier
ana
lysi
s to
ape
riod
ic s
igna
ls,
by e
valu
atin
g th
eFo
urie
r in
tegr
al,
whi
ch i
s gi
ven
by
2.1
Rev
iew
of
Sig
nal
Pro
cess
ing
53
We
wil
l no
t, ho
wev
er,
bela
bor
this
poi
nt.
Our
pur
pose
her
e is
mer
ely
to u
nder
-st
and
the
freq
uenc
y-do
mai
n na
ture
of s
igna
ls.
Rea
ders
inte
rest
ed in
inv
esti
gati
ngFo
urie
r an
alys
is f
urth
er a
re r
efer
red
to K
apla
n
F
ilter
Im
ple
men
tatio
n a
nd
Dig
ital
Sig
nal
Pro
cess
ing
Ear
ly i
mpl
emen
tatio
ns o
f th
e fo
ur b
asic
filt
ers
wer
e pr
edom
inan
tly t
uned
RL
Cci
rcui
ts.
Thi
s ap
proa
ch h
ad a
bas
ic l
imita
tion,
how
ever
, in
tha
t th
e fi
lters
had
only
a v
ery
smal
l ra
nge
of a
djus
tabi
lity.
A
side
fro
m o
ur b
eing
abl
e to
cha
nge
the
reso
nant
fre
quen
cy o
f th
e fi
lter
by a
djus
ting
a va
riab
le c
apac
itor
or i
nduc
tor,
the
filte
rs w
ere
pret
ty m
uch
fixe
d on
ce i
mpl
emen
ted,
lea
ving
litt
le r
oom
for
chan
ge a
s ap
plic
atio
ns b
ecam
e m
ore
soph
istic
ated
.Th
e ne
xt s
tep
in t
he e
volu
tion
of f
ilter
des
ign
cam
e ab
out w
ith t
he a
dven
t of
digi
tal
com
pute
r sy
stem
s, a
nd, j
ust
rece
ntly
, w
ith t
he a
vaila
bilit
y of
mic
roco
m-
pute
r ch
ips
with
arc
hite
ctur
es c
usto
m-t
ailo
red
for
sign
al-p
roce
ssin
g ap
plic
atio
ns.
The
basi
c co
ncep
t und
erly
ing
digi
tal
filte
r im
plem
enta
tion
is t
he i
dea
that
a c
on-
tinuo
us a
nalo
g si
gnal
can
be
sam
pled
per
iodi
cally
, qu
antiz
ed,
and
proc
esse
dby
a f
airl
y st
anda
rd c
ompu
ter
syst
em.
This
app
roac
h, i
llust
rate
d in
Fig
ure
2.6,
over
cam
e th
e lim
itatio
n of
fix
ed im
plem
enta
tion,
bec
ause
cha
ngin
g th
e fi
lter
was
sim
ply
a m
atte
r of
rew
ritin
g th
e so
ftw
are
for
the
com
pute
r. W
e w
ill t
here
fore
conc
entra
te o
n w
hat
goes
on
with
in t
he s
oftw
are
sim
ulat
ion
of th
e an
alog
fil
ter.
We
assu
me
that
the
com
pute
r im
plem
enta
tion
of t
he f
ilter
is
a di
scre
te-
time,
lin
ear,
time-
inva
rian
t sy
stem
. Sy
stem
s th
at s
atis
fy t
hese
con
stra
ints
can
perf
orm
a t
rans
form
atio
n on
an
inpu
t si
gnal
, ba
sed
on s
ome
pred
efin
ed c
rite
ria,
Orig
inal
sig
nal
Tim
e
t
Dis
cret
e sa
mpl
es
Tim
e
Fig
ure
2.6
D
iscr
ete-
time
sam
plin
g of
a c
on
tinu
ou
s si
gn
al
is s
ho
wn
.
54A
dal
ine a
nd
to p
rodu
ce a
n ou
tput
tha
t co
rres
pond
s to
the
inp
ut a
s th
ough
it
had
pass
edth
roug
h an
ana
log
filt
er.
Thu
s, a
com
pute
r, e
xecu
ting
a p
rogr
am t
hat
appl
ies
a gi
ven
tran
sfor
mat
ion
oper
atio
n, t
o di
scre
te,
digi
tize
d ap
prox
imat
ions
of
aco
ntin
uous
inp
ut s
igna
l, c
an p
rodu
ce a
n ou
tput
val
ue y
(n)
for
each
inp
utsa
mpl
e, w
here
is
the
disc
rete
tim
este
p va
riab
le.
In i
ts r
ole
in p
erfo
rmin
g th
istr
ansf
orm
atio
n, t
he c
ompu
ter
can
be t
houg
ht o
f as
a d
igita
l fi
lter
. M
oreo
ver,
any
filt
er c
an b
e co
mpl
etel
y ch
arac
teri
zed
by i
ts r
espo
nse,
h(n
), t
o th
e un
it i
mpu
lse
func
tion,
rep
rese
nted
as
M
ore
prec
isel
y,
=
The
bene
fit
of th
is f
orm
ulat
ion
is t
hat,
once
the
sys
tem
res
pons
e to
the
uni
tim
puls
e is
kno
wn,
the
sys
tem
out
put
for
any
inpu
t is
giv
en b
y
y(n)
=
—
whe
re i
s th
e sy
stem
inp
ut.
Thi
s eq
uatio
n is
mea
ning
ful
to u
s in
tha
t it
desc
ribe
s a
conv
olut
ion
sum
betw
een
the
inpu
t si
gnal
and
the
unit
impu
lse
resp
onse
of
the
syst
em.
The
pro
- c
an b
e pi
ctur
ed a
s a
win
dow
slid
ing
past
a s
cene
of
inte
rest
. A
s ill
ustr
ated
in F
igur
e 2.
7, f
or e
ach
time
step
, th
e sy
stem
out
put
is p
rodu
ced
by t
rans
posi
ngan
d sh
iftin
g o
ne p
ositi
on t
o th
e ri
ght.
The
sum
mat
ion
is t
hen
perf
orm
edov
er a
ll no
nzer
o va
lues
of
for
the
fin
ite l
engt
h of
the
filte
r. In
thi
s m
anne
r,w
e ca
n re
aliz
e th
e fi
lter
by
repe
titiv
ely
perf
orm
ing
floa
ting-
poin
t m
ultip
licat
ions
and
addi
tions
, co
uple
d w
ith s
ampl
e tim
e de
lays
and
shi
ft o
pera
tions
. R
epet
itive
,m
athe
mat
ical
ope
ratio
ns a
re w
hat
com
pute
rs d
o be
st;
ther
efor
e, t
he c
onvo
lutio
nsu
m p
rovi
des
us w
ith
a m
echa
nism
for
bui
ldin
g th
e di
gita
l eq
uiva
lent
of
anal
ogfi
lters
. R
eade
rs i
nter
este
d in
lea
rnin
g m
ore
abou
t di
gita
l si
gnal
pro
cess
ing
are
refe
rred
to
Opp
enhe
im a
nd S
chaf
er [
5] o
r H
amm
ing
It is
suf
fici
ent
for
our
purp
oses
to n
ote
that
the
con
volu
tion
sum
is
apr
oduc
ts o
pera
tion
sim
ilar
to t
he t
ype
of o
pera
tion
an A
NS
PE p
erfo
rms
whe
nco
mpu
ting
its i
nput
act
ivat
ion
sign
al.
Spec
ific
ally
, th
e A
dalin
e us
es e
xact
ly t
his
cal
cula
tion,
with
out t
he s
ampl
e tim
e de
lays
and
shi
ft o
pera
tions
,to
det
erm
ine
how
muc
h in
put
stim
ulat
ion
it re
ceiv
es f
rom
an
inst
anta
neou
s in
put
sign
al.
As
we
shal
l se
e in
the
nex
t se
ctio
n, t
he A
dalin
e ex
tend
s th
e ba
sic
filt
erop
erat
ion
one
step
fur
ther
, in
tha
t it
has
impl
emen
ted
wit
hin
itsel
f a
mea
nsof
ada
ptin
g th
e w
eigh
ting
coef
fici
ents
to
allo
w i
t to
inc
reas
e or
dec
reas
e th
est
imul
atio
n it
rece
ives
the
nex
t tim
e it
is p
rese
nted
wit
h th
e sa
me
sign
al.
The
abi
lity
of t
he A
dalin
e to
ada
pt i
ts w
eigh
ting
coef
fici
ents
is
extr
emel
yus
eful
. W
hen
wri
ting
a di
gita
l fi
lter
pro
gram
on
a co
mpu
ter,
the
pro
gram
mer
mus
t kn
ow e
xact
ly h
ow t
o sp
ecif
y th
e fi
lteri
ng a
lgor
ithm
and
wha
t th
e de
tails
of
the
sign
al c
hara
cter
istic
s ar
e. I
f mod
ific
atio
ns a
re d
esir
ed, o
r if
the
sign
al c
hara
c-te
rist
ics
chan
ge, r
epro
gram
min
g is
requ
ired
. W
hen
the
prog
ram
mer
use
s an
Ada
-lin
e, t
he p
robl
em s
hift
s to
one
of
bein
g ab
le t
o sp
ecif
y th
e de
sire
d ou
tput
sig
nal,
2.2
Ad
alin
e a
nd
th
e A
dap
tive
Lin
ear
Com
bin
er55
(a)
(b)
0.2 0.1
I
I I
•-
•
" , .
I T
, ,
01
23
45
67
n
x(n)
1.00
0.50
- • • •
• • • •
nx(
n)1
1 1
1 0.
5
5
0.5
6
0.5
7
05
(c)
|.05
|.10
I .05
.10
|.2b
.30
.25
y(0)
=
|.1
5 =
=
0.30
= =
Fig
ure
2.7
C
on
volu
tion s
um
ca
lcu
latio
n i
s (
a) T
he p
roce
ss b
egin
sby
de
term
inin
g t
he d
esi
red
resp
on
se o
f th
e f
ilte
r to
th
e u
nit
imp
uls
e f
un
ctio
n
at e
igh
t d
iscr
ete
tim
est
ep
s.
(b)
The
in
put
sign
al i
s sa
mpl
ed a
nd q
uant
ized
eig
ht
(c)
The
out
put
of
the f
ilte
r is
pro
duce
d f
or
ea
ch b
y m
ulti
plic
atio
n o
fea
ch te
rm in
(a)
with
the
corr
espo
ndin
g va
lue
of (b
) fo
r a
ll va
lidtim
este
ps.
give
n a
part
icul
ar i
nput
sig
nal.
The
Ada
line
take
s th
e in
put
and
the
desi
red
out-
put,
and
adju
sts
itse
lf s
o th
at i
t ca
n pe
rfor
m t
he d
esir
ed t
rans
form
atio
n. F
urth
er-
mor
e, t
he s
igna
l ch
arac
teri
stic
s ch
ange
, th
e A
dalin
e ca
n ad
apt
auto
mat
ical
ly.
We
shal
l no
w e
xpan
d th
ese
idea
s, a
nd b
egin
our
inve
stig
atio
n of
the
Ada
line.
2.2
AD
AL
INE
A
ND
T
HE
A
DA
PT
IVE
LIN
EA
R C
OM
BIN
ER
Ada
line
is
a de
vice
con
sist
ing
of a
sin
gle
proc
essi
ng e
lem
ent;
as
such
, it
is t
echn
ical
ly a
neu
ral
netw
ork.
N
ever
thel
ess,
it
is a
ver
y im
port
ant
stru
ctur
eth
at d
eser
ves
clos
e st
udy.
M
oreo
ver,
we
wil
l sh
ow h
ow i
t ca
n fo
rm t
he b
asis
°f a
net
wor
k in
a l
ater
sec
tion.
56A
dal
ine a
nd
Mad
alin
e
The
ter
m A
dalin
e is
an
acro
nym
; ho
wev
er,
its
mea
ning
has
cha
nged
som
e-w
hat
over
the
yea
rs.
Initi
ally
cal
led
the
AD
Apt
ive
Lin
ear
NE
uron
, it
beca
me
the
AD
Apt
ive
LIN
ear
Ele
men
t, w
hen
neur
al n
etw
orks
fel
l ou
t of
fav
or i
n th
ela
te 1
960s
. It
is
alm
ost
iden
tica
l in
str
uctu
re t
o th
e ge
nera
l PE
des
crib
ed i
nC
hapt
er 1
. F
igur
e 2.
8 sh
ows
the
Ada
line
str
uctu
re.
The
re a
re t
wo
basi
c m
od-
ific
atio
ns r
equi
red
to m
ake
the
gene
ral
PE s
truc
ture
int
o an
Ada
line.
T
he f
irst
mod
ific
atio
n is
the
add
itio
n of
a c
onne
ctio
n w
ith
wei
ght,
whi
ch w
e re
fer
to a
s th
e bi
as t
erm
. T
his
term
is
a w
eigh
t on
a c
onne
ctio
n th
at h
as i
ts i
nput
valu
e al
way
s eq
ual
to
1.
The
inc
lusi
on o
f su
ch a
ter
m i
s la
rgel
y a
mat
ter
ofex
peri
ence
. W
e sh
ow i
t he
re f
or c
ompl
eten
ess,
but
it
wil
l no
t ap
pear
in
the
disc
ussi
on o
f th
e ne
xt s
ectio
ns.
We
shal
l re
surr
ect
the
idea
of
a bi
as t
erm
in
Cha
pter
3,
on t
he b
ackp
ropa
gatio
n ne
twor
k.T
he s
econ
d m
odif
icat
ion
is t
he a
ddit
ion
of a
bip
olar
con
diti
on o
n th
e ou
tput
.T
he d
ashe
d bo
x in
Fig
ure
2.8
encl
oses
a p
art
of th
e A
dalin
e ca
lled
the
adap
tive
linea
r co
mbi
ner
(AL
C).
If th
e ou
tput
of t
he A
LC
is p
ositi
ve, t
he A
dalin
e ou
tput
is +
1. I
f th
e A
LC
out
put
is n
egat
ive,
the
Ada
line
out
put
is —
1.
Bec
ause
muc
hof
the
int
eres
ting
pro
cess
ing
take
s pl
ace
in t
he A
LC
por
tion
of
the
Ada
line
,w
e sh
all
conc
entr
ate
on t
he A
LC.
Lat
er,
we
shal
l ad
d ba
ck t
he b
inar
y ou
tput
cond
itio
n.Th
e pr
oces
sing
don
e by
the
ALC
is
that
of
the
typi
cal
proc
essi
ng e
lem
ent
desc
ribe
d in
the
pre
viou
s ch
apte
r. T
he A
LC
per
form
s a
y
+1
outp
ut-
Ada
ptiv
e lin
ear
com
bine
rI
Figu
re 2
.8
The
com
plet
e A
da
line
cons
ists
of t
he a
dapt
ive
linea
r co
mbi
ner,
in
the
dash
ed
box,
and
a
bip
ola
r o
utp
ut
fun
ctio
n.
Th
eadaptiv
e l
inear
com
bine
r re
sem
bles
the
gen
eral
PE
desc
ribed
in C
ha
pte
r
2.2
Ad
alin
e a
nd
th
e A
dap
tive
Lin
ear
Co
mb
iner
5
7
lati
on u
sing
the
inp
ut a
nd w
eigh
t ve
ctor
s, a
nd a
pplie
s an
out
put
func
tion
to
get
a si
ngle
out
put
valu
e. U
sing
the
not
atio
n in
Fig
ure
2.8,
y =
whe
re i
s th
e bi
as w
eigh
t. If
we
mak
e th
e id
enti
fica
tion
, =
1,
we
can
rew
rite
the
pre
cedi
ng e
quat
ion
as
or,
in v
ecto
r no
tati
on,
y =
(2.1)
The
out
put
func
tion
in
this
cas
e is
the
ide
ntity
fun
ctio
n,
as i
s th
e ac
tiva
tion
func
tion
. Th
e us
e of
the
iden
tity
func
tion
as
both
out
put
and
acti
vati
on f
unct
ions
mea
ns t
hat t
he o
utpu
t is
the
sam
e as
the
act
ivat
ion,
whi
ch i
s th
e sa
me
as t
he n
etin
put
to t
he u
nit.
The
Ada
line
(or
the
AL
C)
is A
DA
ptiv
e in
the
sen
se t
hat
ther
e ex
ists
aw
ell-
defi
ned
proc
edur
e fo
r m
odif
ying
the
wei
ghts
in
orde
r to
all
ow t
he d
evic
eto
giv
e th
e co
rrec
t ou
tput
val
ue f
or t
he g
iven
inp
ut.
Wha
t ou
tput
val
ue i
sco
rrec
t de
pend
s on
the
par
ticu
lar
proc
essi
ng f
unct
ion
bein
g pe
rfor
med
by
the
devi
ce.
The
Ada
line
(or
the
AL
C)
is L
inea
r be
caus
e th
e ou
tput
is a
sim
ple
line
arfu
ncti
on o
f th
e in
put
valu
es.
It i
s a
NE
uron
onl
y in
the
ver
y li
mit
ed s
ense
of
the
PEs
desc
ribe
d in
the
pre
viou
s ch
apte
r. T
he A
dali
ne c
ould
als
o be
sai
d to
be
a E
lem
ent,
avoi
ding
the
NE
uron
iss
ue a
ltoge
ther
. In
the
nex
t se
ctio
n, w
elo
ok a
t a
met
hod
to t
rain
the
Ada
line
to p
erfo
rm a
giv
en p
roce
ssin
g fu
ncti
on.
2.2.
1 T
he
LMS
Lea
rnin
g R
ule
Giv
en a
n in
put
vect
or,
x, i
t is
str
aigh
tfor
war
d to
det
erm
ine
a se
t of
wei
ghts
, w
hich
wil
l re
sult
in a
par
ticu
lar
outp
ut v
alue
, y.
Su
ppos
e w
e ha
ve a
set
of in
put
vect
ors,
XL
}, e
ach
havi
ng i
ts o
wn,
per
haps
uni
que,
cor
rect
or d
esir
ed o
utpu
t va
lue,
k =
The
pro
blem
of
find
ing
a si
ngle
wei
ght
vect
or t
hat
can
succ
essf
ully
ass
ocia
te e
ach
inpu
t ve
ctor
wit
h it
s de
sire
d ou
tput
valu
e is
no
long
er s
impl
e. I
n th
is s
ectio
n, w
e de
velo
p a
met
hod
calle
d th
e le
ast-
mea
n-sq
uare
(L
MS)
lea
rnin
g ru
le,
whi
ch i
s on
e m
etho
d of
fin
ding
the
des
ired
wei
ght
vect
or.
We
refe
r to
thi
s pr
oces
s of
fin
ding
the
wei
ght
vect
or a
s tr
aini
ngth
e A
LC.
The
lear
ning
rul
e ca
n be
em
bedd
ed i
n th
e de
vice
its
elf,
whi
ch c
an t
hen
self-
adap
t as
inp
uts
and
desi
red
outp
uts
are
pres
ente
d to
it.
Smal
l ad
just
men
tsar
e m
ade
to t
he w
eigh
t va
lues
as
each
com
bina
tion
is
proc
esse
dun
til
the
AL
C g
ives
cor
rect
out
puts
. In
a s
ense
, th
is p
roce
dure
is
a tr
ue t
rain
ing
proc
edur
e, b
ecau
se w
e do
not
nee
d to
cal
cula
te t
he v
alue
of
the
wei
ght
vect
orex
plic
itly
. B
efor
e de
scri
bing
the
tra
inin
g pr
oces
s in
det
ail,
let
's p
erfo
rm t
heca
lcul
atio
n m
anua
lly.
58
Ad
alin
e a
nd
Cal
cula
tion
of w
*.
To b
egin
, le
t's s
tate
the
pro
blem
a l
ittle
dif
fere
ntly
: G
iven
exam
ples
, o
f so
me
proc
essi
ng f
unct
ion
that
ass
o-ci
ates
inp
ut v
ecto
rs,
wit
h (o
r m
aps
to)
the
desi
red
outp
ut v
alue
s, w
hat
is t
he b
est
wei
ght
vect
or,
for
an
AL
C t
hat
perf
orm
s th
is m
appi
ng?
To a
nsw
er t
his
ques
tion,
we
mus
t fi
rst
defi
ne w
hat
it is
tha
t co
nstit
utes
the
best
wei
ght
vect
or.
Cle
arly
, on
ce t
he b
est
wei
ght
vect
or i
s fo
und,
we
wou
ldlik
e th
e ap
plic
atio
n of
eac
h in
put
vect
or t
o re
sult
in t
he p
reci
se,
corr
espo
ndin
gou
tput
val
ue.
Thu
s, w
e w
ant t
o el
imin
ate,
or
at le
ast t
o m
inim
ize,
the
diff
eren
cebe
twee
n th
e de
sire
d ou
tput
and
the
act
ual
outp
ut f
or e
ach
inpu
t ve
ctor
. Th
eap
proa
ch w
e se
lect
her
e is
to
min
imiz
e th
e m
ean
squa
red
erro
r fo
r th
e se
t of
inpu
t ve
ctor
s.If
the
act
ual
outp
ut v
alue
is
fo
r th
e i
nput
vec
tor,
the
n th
e co
rre-
spon
ding
err
or t
erm
is
—
The
mea
n sq
uare
d er
ror,
or
expe
ctat
ion
valu
e of
the
err
or,
is d
efin
ed b
y
(2.2)
k=\
whe
re L
is
the
num
ber
of i
nput
vec
tors
in
the
trai
ning
Usi
ng E
q. w
e ca
n ex
pand
the
mea
n sq
uare
d er
ror
as f
ollo
ws:
= =•
(2.3)
(2.4)
In g
oing
fro
m E
q. (
2.3)
to
Eq.
(2.4
), w
e ha
ve m
ade
the
assu
mpt
ion
that
the
trai
ning
set
is
stat
istic
ally
sta
tiona
ry,
mea
ning
tha
t an
y ex
pect
atio
n va
lues
var
ysl
owly
wit
h re
spec
t to
tim
e. T
his
assu
mpt
ion
allo
ws
us t
o fa
ctor
out
the
wei
ght
vect
ors
from
the
exp
ecta
tion
valu
e te
rms
in E
q. (
2.4)
.
Exe
rcis
e 2.
1: G
ive
the
deta
ils o
f th
e de
riva
tion
that
lea
ds f
rom
Eq.
(2.
3),
toE
q. (
2.4)
alo
ng w
ith
the
just
ific
atio
n fo
r ea
ch s
tep.
Why
are
the
fac
tors
dk
and
lef
t to
geth
er i
n th
e la
st t
erm
in
Eq.
(2.
4),
rath
er t
han
show
n as
the
pro
duct
of t
he t
wo
sepa
rate
exp
ecta
tion
val
ues?
Def
ine
a m
atri
x R
= c
alle
d th
e in
put
corr
elat
ion
mat
rix,
and
ave
ctor
p
Furt
her,
mak
e th
e id
entif
icat
ion
£ =
U
sing
the
sede
fini
tion
s, w
e ca
n re
wri
te E
q. (
2.4)
as
(2.5
)
Thi
s eq
uatio
n sh
ows
as
an e
xpli
cit
func
tion
of
the
wei
ght
vect
or,
w.
In o
ther
wor
ds,
=
and
Ste
arns
use
the
not
atio
n, ]
, fo
r th
e ex
pect
atio
n va
lue;
als
o, t
he t
erm
exe
mpl
ars
wil
l so
met
imes
be
seen
as
a sy
nony
m f
or t
rain
ing
set.
2.2
Ad
alin
e a
nd
th
e A
dap
tive
Lin
ear
Co
mb
iner
59
To f
ind
the
wei
ght
vect
or c
orre
spon
ding
to
the
min
imum
m
ean
squa
red
erro
r, w
e di
ffer
enti
ate
Eq.
(2
.5),
eval
uate
the
res
ult
at a
nd s
et t
he r
esul
teq
ual
to z
ero:
= -
2p
- 2
p =
0
=
p
=
(2.6)
(2.7) (2.8)
Not
ice
that
, al
thou
gh i
s a
scal
ar,
is
a v
ecto
r. Eq
uatio
n (2
.6)
is a
nex
pres
sion
of
the
grad
ient
of
whi
ch i
s th
e ve
ctor
(2.9)
_
All
tha
t w
e ha
ve d
one
by t
he p
roce
dure
is
to s
how
tha
t w
e ca
n fi
nd a
poi
ntw
here
the
slo
pe o
f th
e fu
ncti
on,
is
zero
. In
gen
eral
, th
at p
oint
may
be
am
inim
um o
r a
max
imum
poi
nt.
In t
he e
xam
ple
that
fol
low
s, w
e sh
ow a
sim
ple
case
whe
re t
he A
LC
has
onl
y tw
o w
eigh
ts.
In t
hat
situ
atio
n, t
he g
raph
of
is a
par
abol
oid.
Fur
ther
mor
e, i
t mus
t be
conc
ave
upw
ard,
sin
ce a
ll co
mbi
nati
ons
of w
eigh
ts m
ust
resu
lt in
a n
onne
gativ
e va
lue
for
the
mea
n sq
uare
d er
ror,
Thi
s re
sult
is g
ener
al a
nd i
s ob
tain
ed r
egar
dles
s of
the
dim
ensi
on o
f th
e w
eigh
tve
ctor
. In
the
cas
e of
dim
ensi
ons
high
er t
han
two,
the
par
abol
oid
is k
now
n as
aSu
ppos
e w
e ha
ve
an
AL
C
wit
h tw
o in
puts
and
var
ious
oth
er q
uant
itie
sde
fine
d as
fol
low
s:
R =
3 1
1 4]
= 1
0
Rat
her
than
inv
erti
ng R
, w
e us
e E
q. (
2.7)
to
find
the
opt
imum
wei
ght
vect
or:
3 1
1 4
Thi
s eq
uati
on r
esul
ts i
n tw
o eq
uati
ons
for
w*
and
w*
+
= 5
The
solu
tion
is w
* (
1,
The
gra
ph o
f a
s a
func
tion
of
the
two
wei
ghts
is s
how
n in
Fig
ure
2.9.
2.2
: Sh
ow t
hat
the
min
imum
val
ue o
f th
e m
ean
squa
red
erro
r ca
n be
as
L
60 a
nd
Fig
ure
2.9
F
or
an
AL
C
with
on
ly
two
wei
ghts
, th
e er
ror
surf
ace
is
a
para
bolo
id.
The
wei
ghts
tha
t m
inim
ize
the
err
or
occ
ur
at t
hebo
ttom
of
the p
arab
oloi
dal
surf
ace.
Exe
rcis
e 2.
3:
Det
erm
ine
an e
xpli
cit
equa
tion
for
as
a fu
ncti
on o
f a
ndus
ing
the
exam
ple
in t
he t
ext.
Use
it
to f
ind
the
opt
imum
wei
ght
vect
or,
the
min
imum
mea
n sq
uare
d er
ror,
and
pro
ve t
hat
the
para
bolo
id i
sco
ncav
e up
war
d.
In t
he n
ext
sect
ion,
we
shal
l ex
amin
e a
met
hod
for
find
ing
the
optim
umw
eigh
t ve
ctor
by
an i
tera
tive
pro
cedu
re.
Thi
s pr
oced
ure
allo
ws
us t
o av
oid
the
ofte
n-di
ffic
ult
calc
ulat
ions
nec
essa
ry t
o de
term
ine
the
wei
ghts
man
uall
y.
Fin
ding
w*
by t
he M
etho
d of
Ste
epes
t D
esce
nt.
As
you
mig
ht i
mag
ine,
the
anal
ytic
al c
alcu
lati
on t
o de
term
ine
the
opti
mum
wei
ghts
for
a p
robl
em i
s ra
ther
diff
icul
t in
gen
eral
. N
ot o
nly
does
the
mat
rix
man
ipul
atio
n ge
t cu
mbe
rsom
e fo
rla
rge
dim
ensi
ons,
but
als
o ea
ch c
ompo
nent
of
R a
nd p
is
itsel
f an
exp
ecta
tion
valu
e. T
hus,
exp
lici
t ca
lcul
atio
ns o
f R
and
p r
equi
re k
now
ledg
e of
the
stat
isti
csof
the
inpu
t si
gnal
s. A
bet
ter
appr
oach
wou
ld b
e to
let
the
AL
C f
ind
the
opti
mum
wei
ghts
its
elf
by h
avin
g it
sear
ch o
ver
the
wei
ght
surf
ace
to f
ind
the
min
imum
.A
pur
ely
rand
om s
earc
h m
ight
not
be
prod
ucti
ve o
r ef
fici
ent,
so w
e sh
all
add
som
e in
tell
igen
ce t
o th
e pr
oced
ure.
2.2
an
d t
he A
dap
tive L
inea
r C
om
bin
er61
Beg
in b
y as
sign
ing
arbi
trar
y va
lues
to
the
wei
ghts
. Fr
om t
hat
poin
t on
the
wei
ght
surf
ace,
det
erm
ine
the
dire
ctio
n of
the
ste
epes
t sl
ope
in t
he d
ownw
ard
dire
ctio
n. C
hang
e th
e w
eigh
ts s
ligh
tly
so t
hat
the
new
wei
ght
vect
or l
ies
fart
her
dow
n th
e su
rfac
e. R
epea
t th
e pr
oces
s un
til
the
min
imum
has
bee
n re
ache
d. T
his
proc
edur
e is
ill
ustr
ated
in
Figu
re I
mpl
icit
in t
his
met
hod
is t
he a
ssum
ptio
nth
at w
e kn
ow w
hat
the
wei
ght
surf
ace
look
s li
ke i
n ad
vanc
e. W
e do
not
kno
w,
but
we
wil
l se
e sh
ortl
y ho
w t
o ge
t ar
ound
thi
s pr
oble
m.
Typ
ical
ly,
the
wei
ght
vect
or d
oes
not
init
iall
y m
ove
dire
ctly
tow
ard
the
min
imum
poi
nt.
The
cros
s-se
ctio
n of
the
par
abol
oida
l w
eigh
t su
rfac
e is
usu
ally
elli
ptic
al,
so t
he n
egat
ive
grad
ient
may
not
poi
nt d
irec
tly
at t
he m
inim
um p
oint
,at
lea
st i
niti
ally
. Th
e si
tuat
ion
is i
llust
rate
d m
ore
clea
rly
in t
he c
onto
ur p
lot
ofth
e w
eigh
t su
rfac
e in
Fig
ure
Figu
re 2
.10
We c
an
use t
his
di
agra
m
to vi
sua
lize t
he st
eepe
st-d
esce
ntm
etho
d.
An
initi
al
sele
ctio
n fo
r th
e w
eig
ht
vect
or
resu
ltsin
a
n
err
or,
T
he st
ee
pe
st-d
esc
en
t m
etho
d co
nsi
sts
of
slid
ing th
is p
oin
t dow
n the s
urf
ace
tow
ard
the b
otto
m,
alw
ays
mo
vin
g i
n t
he d
ire
ctio
n o
f th
e s
teepest
do
wn
wa
rd s
lope.
62 a
nd
2.3.
Fig
ure
In
the c
onto
ur
plot
of
the w
eig
ht
surf
ace
of
Fig
ure
the
dir
ect
ion
of s
teep
est
de
sce
nt
is p
erp
en
dic
ula
r to
the
con
tour
line
s at
ea
ch p
oint
, an
d th
is d
ire
ctio
n do
es n
ot a
lwa
ys p
oin
tto
the m
inim
um
poin
t.
Bec
ause
the
wei
ght
vect
or i
s va
riab
le i
n th
is p
roce
dure
, w
e w
rite
it
as a
nex
plic
it fu
ncti
on o
f th
e ti
mes
tep,
t.
The
init
ial
wei
ght
vect
or i
s de
note
dan
d th
e w
eigh
t ve
ctor
at
times
tep
t is
At e
ach
step
, th
e ne
xt w
eigh
t ve
ctor
is c
alcu
late
d ac
cord
ing
to
1) =
+(2
.10)
whe
re i
s th
e ch
ange
in
at
the
tim
este
p.W
e ar
e lo
okin
g fo
r th
e di
rect
ion
of t
he s
teep
est
desc
ent
at e
ach
poin
t on
the
surf
ace,
so
we
need
to
calc
ulat
e th
e gr
adie
nt o
f th
e su
rfac
e (w
hich
giv
es t
hedi
rect
ion
of t
he s
teep
est
slo
pe).
T
he n
egat
ive
of t
he g
radi
ent
is i
n th
edi
rect
ion
of s
teep
est
desc
ent.
To g
et t
he m
agni
tude
of
the
chan
ge,
mul
tipl
y th
egr
adie
nt b
y a
suita
ble
cons
tant
, T
he a
ppro
pria
te v
alue
for
wil
l be
dis
cuss
edla
ter.
Thi
s pr
oced
ure
resu
lts
in t
he f
ollo
win
g ex
pres
sion
:
(2.1
1)
2.2
A
dal
ine
and
th
e A
dap
tive
Lin
ear
Co
mb
iner
All
tha
t is
nec
essa
ry t
o co
mpl
ete
the
disc
ussi
on i
s to
det
erm
ine
the
valu
e of
at
each
suc
cess
ive
iter
atio
n st
ep.
The
val
ue o
f w
as d
eter
min
ed a
naly
tica
lly
in t
he p
revi
ous
sect
ion.
Equ
atio
n (2
.6)
or E
q. (
2.9)
cou
ld b
e us
ed h
ere
to d
eter
min
e b
ut w
ew
ould
hav
e th
e sa
me
prob
lem
tha
t w
e ha
d w
ith
the
anal
ytic
al d
eter
min
atio
nof
W
e w
ould
nee
d to
kno
w b
oth
R a
nd p
in
adva
nce.
T
his
know
ledg
eis
equ
ival
ent
to k
now
ing
wha
t th
e w
eigh
t su
rfac
e lo
oks
like
in
adva
nce.
To
circ
umve
nt t
his
diff
icul
ty,
we
use
an a
ppro
xim
atio
n fo
r th
e gr
adie
nt t
hat
can
bede
term
ined
fro
m i
nfor
mat
ion
that
is
know
n ex
plic
itly
at
each
ite
ratio
n.Fo
r ea
ch s
tep
in t
he i
tera
tion
pro
cess
, w
e pe
rfor
m t
he f
ollo
win
g:
1. A
pply
an
inpu
t ve
ctor
, t
o th
e A
dalin
e in
puts
.
2.
Det
erm
ine
the
valu
e of
the
err
or s
quar
ed,
usi
ng t
he c
urre
nt v
alue
of
the
wei
ght
vect
or
(2.1
2)
3.
Cal
cula
te a
n ap
prox
imat
ion
to b
y us
ing
as
an a
ppro
xim
atio
nfo
r
(2
.13)
=
(2.1
4)
whe
re w
e ha
ve u
sed
Eq.
to
calc
ulat
e th
e gr
adie
nt e
xpli
citl
y.
4. U
pdat
e th
e w
eigh
t ve
ctor
acc
ordi
ng t
o E
q. u
sing
Eq.
as
the
appr
oxim
atio
n fo
r th
e gr
adie
nt:
+ 1
) =
+
(2.1
5)
5. R
epea
t st
eps
1 th
roug
h 4
wit
h th
e ne
xt i
nput
vec
tor,
unti
l th
e er
ror
has
been
redu
ced
to a
n ac
cept
able
val
ue.
Equ
atio
n (2
.15)
is
an e
xpre
ssio
n of
the
LM
S a
lgor
ithm
. T
he p
aram
eter
dete
rmin
es t
he s
tabi
lity
and
spee
d of
con
verg
ence
of
the
wei
ght
vect
or t
owar
dth
e m
inim
um-e
rror
val
ue.
Bec
ause
an
appr
oxim
atio
n of
the
gra
dien
t ha
s be
en u
sed
in E
q. t
hepa
th t
hat
the
wei
ght
vect
or t
akes
as
it m
oves
dow
n th
e w
eigh
t su
rfac
e to
war
dth
e m
inim
um w
ill
not b
e as
sm
ooth
as
that
ind
icat
ed i
n Fi
gure
Fig
ure
show
s an
exa
mpl
e of
how
a s
earc
h pa
th m
ight
loo
k w
ith
the
LM
S al
gori
thm
of
Eq.
(2.
15).
Cha
nges
in
the
wei
ght
vect
or m
ust
be k
ept
rela
tive
ly s
mal
l on
eac
hite
ratio
n.
If c
hang
es a
re t
oo l
arge
, th
e w
eigh
t ve
ctor
cou
ld w
ande
r ab
out
the
surf
ace,
nev
er f
indi
ng t
he m
inim
um,
or f
indi
ng i
t on
ly b
y ac
cide
nt r
athe
r th
anas
a r
esul
t of
a s
tead
y co
nver
genc
e to
war
d it
. T
he f
unct
ion
of t
he p
aram
eter
is t
o pr
even
t th
is a
imle
ss s
earc
hing
. In
the
nex
t se
ctio
n, w
e sh
all
disc
uss
the
para
met
er,
and
oth
er p
ract
ical
con
side
rati
ons.
64
Ad
alin
e a
nd
Fig
ure
T
he h
ypo
the
tica
l pat
h t
ake
n b
y a
we
igh
t ve
cto
r a
s it
sea
rch
es
for
the
min
imu
m
err
or
usi
ng t
he
alg
ori
thm
is
no
t a
smo
oth
cu
rve b
eca
use
the
gra
die
nt
is b
ein
g a
pp
roxi
ma
ted
at e
ach p
oint
. N
ote a
lso t
hat
step
siz
es
get
smal
ler
as t
hem
inim
um
-err
or
solu
tion is
ap
pro
ach
ed
.
2.2.
2 P
ract
ical
Con
side
ratio
ns
The
re a
re s
ever
al q
uest
ions
to
cons
ider
whe
n w
e ar
e at
tem
ptin
g to
use
the
AL
Cto
sol
ve a
par
ticu
lar
prob
lem
:
• H
ow m
any
trai
ning
vec
tors
are
req
uire
d to
sol
ve a
par
ticu
lar
prob
lem
?
• H
ow i
s th
e ex
pect
ed o
utpu
t ge
nera
ted
for
each
tra
inin
g ve
ctor
?
• W
hat
is t
he a
ppro
pria
te d
imen
sion
of
the
wei
ght
vect
or?
• W
hat
shou
ld b
e th
e in
itia
l va
lues
for
the
wei
ghts
?
• Is
a b
ias
wei
ght
requ
ired
?
• W
hat
happ
ens
if t
he s
igna
l st
atis
tics
var
y w
ith
tim
e?
2.2
an
d t
he
Ad
apti
ve L
inea
r C
om
bin
er65
• W
hat
is t
he a
ppro
pria
te v
alue
for
• H
ow d
o w
e de
term
ine
whe
n to
sto
p tr
aini
ng?
The
ans
wer
s to
the
se q
uest
ions
dep
end
on t
he s
peci
fic
prob
lem
bei
ng a
ddre
ssed
,so
it
is d
iffi
cult
to
give
wel
l-de
fine
d re
spon
ses
that
app
ly i
n al
l ca
ses.
Mor
eove
r,fo
r a
spec
ific
cas
e, t
he a
nsw
ers
are
not
nece
ssar
ily
inde
pend
ent.
Con
side
r th
e di
men
sion
of
the
wei
ght
vect
or.
If t
here
are
a w
ell-
defi
ned
num
ber
of f
rom
mul
tipl
e t
here
wou
ld b
e on
e w
eigh
tfo
r eac
h in
put.
The
ques
tion
wou
ld b
e w
heth
er to
add
a b
ias
wei
ght.
Figu
rede
pict
s th
is c
ase,
wit
h th
e bi
as t
erm
add
ed,
in a
som
ewha
t st
anda
rd f
orm
tha
tsh
ows
the
vari
abil
ity
of t
he w
eigh
ts,
the
erro
r te
rm,
and
the
feed
back
fro
mth
e ou
tput
to
the
wei
ghts
. A
s fo
r th
e bi
as t
erm
its
elf,
inc
ludi
ng i
t so
met
imes
help
s co
nver
genc
e of
the
wei
ghts
to
an a
ccep
tabl
e so
luti
on.
It i
s pe
rhap
s be
stth
ough
t of
as
an e
xtra
deg
ree
of fr
eedo
m,
and
its
use
is l
arge
ly a
mat
ter
ofex
peri
men
tatio
n w
ith t
he s
peci
fic
appl
icat
ion.
A s
itua
tion
dif
fere
nt f
rom
the
pre
viou
s pa
ragr
aph
aris
es i
f th
ere
is o
nly
asi
ngle
inp
ut s
igna
l, sa
y fr
om a
sin
gle
elec
troc
ardi
ogra
ph (
EK
G)
sens
or.
For
= 1
(bi
as in
put)
des
ired
outp
ut
Fig
ure
2.1
3
This
figure
show
s a s
tandard
dia
gra
m o
f the A
LC
with
multi
ple
inpu
ts a
nd
a bi
as t
erm
. W
eigh
ts a
re i
nd
ica
ted
as v
ari
ab
lere
sist
ors
to
em
phas
ize
the
adap
tive
natu
re
of
the
devi
ce.
Calc
ula
tion o
f the e
rror,
is s
how
n e
xp
licitly
as the
additi
on
of
a n
egativ
e o
f th
e o
utp
ut
sig
na
l to
th
e d
esir
ed o
utp
ut
valu
e.
66A
dal
ine
and
exam
ple,
an
AL
C c
an b
e us
ed t
o re
mov
e no
ise
from
the
inp
ut s
igna
l in
ord
er t
ogi
ve a
cle
aner
sig
nal a
t the
out
put.
In a
cas
e su
ch a
s th
is o
ne, t
he A
LC
is
arra
nged
in a
con
figu
rati
on k
now
n as
a t
rans
vers
e fi
lter
. In
thi
s co
nfig
urat
ion,
the
inp
utsi
gnal
is
sam
pled
at
seve
ral
poin
ts i
n ti
me,
rat
her
than
fro
m s
ever
al s
enso
rs a
ta
sing
le t
ime.
Fig
ure
sho
ws
the
AL
C a
rran
ged
as a
tra
nsve
rse
filt
er.
For
the
tran
sver
se f
ilte
r, e
ach
addi
tion
al s
ampl
e in
tim
e re
pres
ents
ano
ther
degr
ee o
f fr
eedo
m t
hat
can
be u
sed
to f
it t
he i
nput
sig
nal
to t
he d
esir
ed o
utpu
tsi
gnal
. T
hus,
if
you
cann
ot g
et a
goo
d fi
t w
ith
a sm
all
num
ber
of s
ampl
es,
try
a fe
w m
ore.
O
n th
e ot
her
hand
, if
you
get
good
con
verg
ence
wit
h yo
ur f
irst
Fig
ure
2.1
4 In
an
A
LC
ar
rang
ed
as
a tr
an
sve
rse
filte
r,
the
sam
ple
s a
re p
rovi
de
d b
y n—
1,
pre
sum
ab
ly e
qual
, tim
e d
ela
ys,
T
he
AL
C s
ees
the
sig
na
l at
the
cu
rre
nt
time,
as
we
ll as
its v
alu
e at
th
e p
revi
ou
s n
- 1
sam
ple
tim
es.
W
hen
data
is
initi
ally
applie
d,
rem
em
ber
to w
ait
at
least
for
data
to b
epr
esen
t at
all
of t
he A
LC
's i
nput
s.
2.2
A
dal
ine a
nd
th
e A
dap
tive L
inea
r C
om
bin
er
67
choi
ce,
try
one
wit
h fe
wer
sam
ples
to
see
whe
ther
you
get
a s
igni
fica
nt s
peed
upin
con
verg
ence
and
stil
l ha
ve s
atis
fact
ory
resu
lts (
you
may
be
surp
rise
d to
fin
dth
at t
he r
esul
ts a
re b
ette
r in
som
e ca
ses)
. M
oreo
ver,
the
bia
s w
eigh
t is
pro
babl
ysu
perf
luou
s in
thi
s ca
se.
Ear
lier,
we
allu
ded
to a
rela
tions
hip
betw
een
trai
ning
tim
e an
d th
e di
men
sion
of t
he w
eigh
t ve
ctor
, es
peci
ally
for
the
sof
twar
e si
mul
atio
ns t
hat
we
cons
ider
in t
his
text
: M
ore
wei
ghts
gen
eral
ly m
ean
long
er t
rain
ing
tim
es.
Thi
s eq
uati
onm
ust
be c
onst
antly
bal
ance
d ag
ains
t ot
her
fact
ors,
suc
h as
the
acc
epta
bilit
y of
the
solu
tion.
As
stat
ed i
n th
e pr
evio
us p
arag
raph
, us
ing
mor
e w
eigh
ts d
oes
not
alw
ays
resu
lt in
a b
ette
r so
lutio
n. F
urth
erm
ore,
the
re a
re o
ther
fac
tors
tha
t af
fect
both
the
tra
inin
g ti
me
and
the
acce
ptab
ility
of
the
solu
tion.
The
para
met
er i
s on
e fa
ctor
tha
t ha
s a
sign
ific
ant
effe
ct o
n tr
aini
ng.
If i
s to
o la
rge,
con
verg
ence
wil
l ne
ver
take
pla
ce,
no m
atte
r ho
w l
ong
is t
hetr
aini
ng p
erio
d.
If t
he s
tatis
tics
of t
he i
nput
sig
nal
are
know
n, i
t is
pos
sibl
e to
show
tha
t th
e va
lue
of i
s re
stri
cted
to
the
rang
e
0
whe
re i
s th
e la
rges
t eig
enva
lue
of th
e m
atri
x R
, the
inp
ut c
orre
latio
n m
atri
xdi
scus
sed
in S
ectio
n A
lthou
gh i
t is
not
alw
ays
reas
onab
le t
o ex
pect
thes
e st
atis
tics
to b
e kn
own,
the
re a
re c
ases
whe
re t
hey
can
be e
stim
ated
. T
hete
xt b
y W
idro
w a
nd S
tear
ns c
onta
ins
man
y ex
ampl
es.
In t
his
text
, w
e pr
opos
e a
mor
e he
uris
tic a
ppro
ach:
Pic
k a
valu
e fo
r s
uch
that
a w
eigh
t do
es n
ot c
hang
eby
mor
e th
an a
sm
all
frac
tion
of it
s cu
rren
t val
ue.
Thi
s ru
le i
s ad
mitt
edly
vag
ue,
but
expe
rien
ce a
ppea
rs t
o be
the
bes
t te
ache
r fo
r se
lect
ing
an a
ppro
pria
te v
alue
for
As
trai
ning
pro
ceed
s, t
he e
rror
val
ue w
ill
dim
inis
h (h
opef
ully
), r
esul
ting
in s
mal
ler
and
smal
ler
wei
ght
chan
ges,
and
, he
nce,
in
a sl
ower
con
verg
ence
tow
ard
the
min
imum
of t
he w
eigh
t sur
face
. It
is s
omet
imes
use
ful t
o in
crea
se t
heva
lue
of d
urin
g th
ese
peri
ods
to s
peed
con
verg
ence
. B
ear
in m
ind,
how
ever
,th
at a
lar
ger
may
mea
n th
at t
he w
eigh
ts m
ight
bou
nce
arou
nd t
he b
otto
m o
fth
e w
eigh
t su
rfac
e, g
ivin
g an
ove
rall
erro
r th
at i
s un
acce
ptab
le.
Her
e ag
ain,
expe
rien
ce i
s ne
cess
ary
to e
nabl
e us
to
judg
e ef
fect
ivel
y.O
ne m
etho
d of
com
pens
atin
g fo
r di
ffer
ence
s in
pro
blem
s is
to
use
norm
al-
ized
inp
ut v
ecto
rs.
Inst
ead
of u
se •
A
noth
er t
acti
c is
to
scal
e th
ede
sire
d ou
tput
val
ue.
Thes
e m
etho
ds h
elp
part
icul
arly
whe
n w
e ar
e se
lect
ing
init
ial
wei
ght
valu
es o
r a
valu
e fo
r I
n m
ost
case
s, w
eigh
ts c
an b
e in
itia
lize
d to
rand
om v
alue
s of
sm
all
real
bet
wee
n -1
.0 a
nd T
he v
alue
of
is
usua
lly
best
kep
t si
gnif
ican
tly
less
tha
n 1;
a v
alue
of
or
even
0.0
5 m
aybe
rea
sona
ble
for
som
e b
ut v
alue
s co
nsid
erab
ly le
ss m
ay b
e re
quir
ed.
The
que
stio
n of
whe
n to
sto
p tr
aini
ng i
s la
rgel
y a
mat
ter
of th
e re
quir
emen
tson
the
out
put
of t
he s
yste
m.
You
det
erm
ine
the
amou
nt o
f er
ror
that
you
can
on
the
outp
ut s
igna
l, an
d tr
ain
unti
l th
e ob
serv
ed e
rror
is
cons
iste
ntly
less
tha
n th
e re
quir
ed v
alue
. Si
nce
the
mea
n sq
uare
d er
ror
is t
he v
alue
use
d to
deri
ve t
he t
rain
ing
algo
rith
m,
that
is
the
quan
tity
tha
t us
uall
y de
term
ines
whe
n
68A
dal
ine
and
Mad
alin
e
a sy
stem
has
con
verg
ed t
o it
s m
inim
um e
rror
sol
utio
n. A
lter
nati
vely
, ob
serv
ing
indi
vidu
al e
rror
s is
oft
en n
eces
sary
, si
nce
the
syst
em p
erfo
rman
ce m
ay h
ave
a re
quir
emen
t th
at n
o er
ror
exce
ed a
cer
tain
am
ount
. N
ever
thel
ess,
a m
ean
squa
red
erro
r th
at f
alls
as
the
iter
atio
n nu
mbe
r in
crea
ses
is p
roba
bly
your
bes
tin
dica
tion
tha
t th
e sy
stem
is
conv
ergi
ng t
owar
d a
solu
tion
.W
e us
uall
y as
sum
e th
at t
he i
nput
sig
nals
are
sta
tist
ical
ly s
tati
onar
y, a
nd,
ther
efor
e,
is e
ssen
tially
a
cons
tant
afte
r th
e op
timum
val
ues
have
been
det
erm
ined
. D
urin
g tr
aini
ng,
wil
l ho
pefu
lly
decr
ease
tow
ard
a st
able
solu
tion.
Su
ppos
e, h
owev
er,
that
the
inp
ut s
igna
l st
atis
tics
chan
ge s
omew
hat
over
tim
e, o
r un
derg
o so
me
disc
onti
nuit
y: A
dditi
onal
tra
inin
g w
ould
be
requ
ired
to c
ompe
nsat
e.O
ne
way
to
deal
w
ith
this
sit
uati
on
is t
o ce
ase
or r
esum
e tr
aini
ng c
on-
diti
onal
ly,
base
d on
the
cur
rent
val
ue o
f
If t
he s
igna
l st
atis
tics
cha
nge,
trai
ning
can
be
rein
itiat
ed u
ntil
is
aga
in r
educ
ed t
o an
acc
epta
ble
valu
e.T
his
met
hod
pres
umes
tha
t a
met
hod
of e
rror
mea
sure
men
t is
ava
ilabl
e.Pr
ovid
ed t
hat t
he i
nput
sig
nals
are
sta
tist
ical
ly s
tati
onar
y, c
hoos
ing
the
num
-be
r of
inp
ut v
ecto
rs t
o us
e du
ring
tra
inin
g m
ay b
e re
lativ
ely
sim
ple.
Y
ou c
anus
e re
al,
inp
uts
as t
rain
ing
vect
ors,
pro
vide
d th
at y
ou k
now
the
desi
red
outp
ut f
or e
ach
inpu
t ve
ctor
. If
it
is p
ossi
ble
to i
dent
ify
a sa
mpl
e of
inpu
t ve
ctor
s th
at a
dequ
atel
y re
prod
uces
the
sta
tistic
al d
istr
ibut
ion
of t
he a
ctua
lin
puts
, it
may
be
poss
ible
to
trai
n on
thi
s se
t in
a s
hort
er t
ime.
Th
e ac
cura
cyof
the
tra
inin
g de
pend
s on
how
wel
l th
e se
lect
ed s
et o
f tr
aini
ng v
ecto
rs m
odel
sth
e di
stri
buti
on o
f th
e en
tire
inpu
t si
gnal
spa
ce.
The
oth
er,
rela
ted
ques
tion
is h
ow t
o go
abo
ut d
eter
min
ing
the
desi
red
outp
ut
for
a gi
ven
inpu
t ve
ctor
. A
s w
ith
man
y qu
esti
ons
disc
usse
d in
thi
sse
ctio
n, t
his
depe
nds
on t
he s
peci
fic
deta
ils o
f th
e pr
oble
m.
Fort
unat
ely,
for
som
e pr
oble
ms,
kn
owin
g th
e de
sire
d re
sult
is e
asy
com
pare
d to
fin
ding
an
algo
rith
m f
or t
rans
form
ing
the
inpu
ts i
nto
the
desi
red
resu
lt.
The
AL
C w
ill
ofte
n so
lve
the
diff
icul
t pa
rt.
The
"eas
y" p
art
is l
eft
to t
he e
ngin
eer.
Exe
rcis
e 2.
4:
A l
owpa
ss f
ilte
r ca
n be
con
stru
cted
wit
h an
Ada
line
havi
ng t
wo
wei
ghts
. C
onsi
der
a si
mpl
e ca
se o
f th
e re
mov
al o
f a
rand
om n
oise
fro
m a
cons
tant
sig
nal.
The
con
stan
t si
gnal
lev
el i
s C
=
3,
an
d th
e ra
ndom
noi
sesi
gnal
has
a c
onst
ant
pow
er,
— —
0.0
25.
Ass
ume
that
the
ran
dom
noi
seis
com
plet
ely
wit
h th
e co
nsta
nt in
put s
igna
l. C
alcu
late
the
optim
umw
eigh
t ve
ctor
and
the
mea
n sq
uare
d er
ror
in t
he o
utpu
t af
ter
the
opti
mum
wei
ght
vect
or h
as b
een
foun
d.
By
find
ing
the
eige
nval
ues
of t
he m
atri
x, R
, de
term
ine
the
max
imum
val
ue o
f th
e co
nsta
nt f
or u
se i
n th
e L
MS
algo
rith
m.
2.3
AP
PL
ICA
TIO
NS
O
F
AD
AP
TIV
ES
IGN
AL
PR
OC
ES
SIN
G
Up
to n
ow,
we
have
bee
n co
ncer
ned
wit
h th
e A
dali
ne m
inus
the
thr
esho
ldco
ndit
ion
on t
he o
utpu
t. In
Sec
tion
2.4,
on
the
Mad
alin
e, w
e w
ill
repl
ace
the
thre
shol
d co
ndit
ion
and
exam
ine
netw
orks
of
Ada
lines
. In
thi
s se
ctio
n, w
e w
ill
L
2.3
Ap
plic
atio
ns
of
Ad
apti
ve S
igna
l P
roce
ssin
g
look
at
a fe
w e
xam
ples
of
adap
tive
sig
nal
proc
essi
ng u
sing
onl
y th
e A
LC
por
tion
of t
he A
dali
ne.
2.3.
1 E
cho
Can
cella
tion
in
Tel
epho
ne C
ircu
its
You
may
hav
e ex
peri
ence
d th
e ph
enom
enon
of
echo
in
tele
phon
e co
nver
sati
ons:
you
hear
the
wor
ds y
ou s
peak
int
o th
e m
outh
piec
e a
frac
tion
of
a se
cond
lat
erin
the
ear
phon
e of
the
tele
phon
e. T
he e
cho
tend
s to
be
mos
t no
ticea
ble
on l
ong-
dist
ance
cal
ls,
espe
cial
ly t
hose
ove
r sa
tell
ite
link
s w
here
tra
nsm
issi
on d
elay
s ca
nbe
a s
igni
fica
nt f
ract
ion
of a
sec
ond.
Tele
phon
e ci
rcui
ts
cont
ain
devi
ces
calle
d hy
brid
s th
at
are
inte
nded
to
isol
ate
inco
min
g si
gnal
s fr
om o
utgo
ing
sign
als,
thu
s av
oidi
ng t
he e
cho
effe
ct.
Unf
ortu
nate
ly,
thes
e ci
rcui
ts d
o no
t alw
ays
perf
orm
per
fect
ly,
due
to c
ause
s su
chas
im
peda
nce
mis
mat
ches
, re
sult
ing
in s
ome
echo
bac
k to
the
spe
aker
. E
ven
whe
n th
e ec
ho s
igna
l ha
s be
en a
ttenu
ated
by
a su
bsta
ntia
l am
ount
, it
still
may
be a
udib
le,
and
henc
e an
ann
oyan
ce t
o th
e sp
eake
r.C
erta
in e
cho-
supp
ress
ion
devi
ces
rely
on
rela
ys t
hat
open
and
clo
se c
ircu
its
in t
he o
utgo
ing
lines
so
that
inc
omin
g vo
ice
sign
als
are
not
sent
bac
k to
the
spea
ker.
Whe
n tr
ansm
issi
on d
elay
s ar
e lo
ng,
as w
ith
sate
llit
e co
mm
unic
atio
ns,
thes
e ec
ho s
uppr
esso
rs c
an r
esul
t in
a l
oss
of p
arts
of
wor
ds.
Thi
s ch
oppy
-sp
eech
eff
ect
is p
erha
ps m
ore
fam
ilia
r th
an t
he e
cho
effe
ct.
An
adap
tive
fil
ter
can
be u
sed
to r
emov
e th
e ec
ho e
ffec
t w
itho
ut t
he c
hopp
ines
s of
the
rela
ys u
sed
in o
ther
ech
o su
ppre
ssio
n ci
rcui
ts [
9, 7
J.Fi
gure
is
a b
lock
dia
gram
of
a te
leph
one
circ
uit
wit
h an
ad
aptiv
efi
lter
use
d as
an
echo
-sup
pres
sion
dev
ice.
T
he e
cho
is c
ause
d by
a l
eaka
ge o
fth
e in
com
ing
voic
e si
gnal
to
the
outp
ut l
ine
thro
ugh
the
hybr
id c
ircu
it.
Thi
sle
akag
e ad
ds t
o th
e ou
tput
sig
nal
com
ing
from
the
mic
roph
one.
Th
e ou
tput
of
the
adap
tive
fil
ter,
is
subt
ract
ed f
rom
the
out
goin
g si
gnal
, s
+ w
here
s i
sth
e ou
tgoi
ng p
ure
voic
e si
gnal
and
is
the
nois
e, o
r ec
ho c
ause
d by
lea
kage
of
the
inco
min
g vo
ice
sign
al t
hrou
gh t
he h
ybri
d ci
rcui
t. T
he s
ucce
ss o
f th
e ec
hoca
ncel
latio
n de
pend
s on
how
wel
l th
e ad
aptiv
e fi
lter
can
mim
ic t
he l
eaka
geth
roug
h th
e hy
brid
cir
cuit
.N
otic
e th
at t
he i
nput
to
the
filt
er i
s a
copy
of
the
inco
min
g si
gnal
, n,
and
that
the
err
or i
s a
copy
of
the
outg
oing
sig
nal,
E s
+ -
y
(2.1
6)
We
assu
me
that
y i
s co
rrel
ated
wit
h th
e no
ise,
but
not
wit
h th
e pu
re v
oice
sign
al,
s. I
f th
e qu
anti
ty,
y,
is n
onze
ro,
som
e ec
ho s
till
rem
ains
in
the
out-
goin
g si
gnal
. Sq
uari
ng a
nd t
akin
g ex
pect
atio
n va
lues
of
both
sid
es o
f E
q.gi
ves
=
+ +
-
(2.1
7)
= +
-
(2.1
8)
Equ
atio
n (2
.18)
fol
low
s, s
ince
s i
s no
t co
rrel
ated
wit
h ei
ther
y o
r r
esul
ting
in t
he l
ast
term
in
Eq.
bei
ng e
qual
to
zero
.
70
Ad
alin
e an
d
Voi
ce s nois
e,
Hyb
ridci
rcui
tA
dapt
ive
filte
r
To earp
hone
Figu
re 2
.15
1
Ada
ptiv
efil
ter
r
Hyb
ridci
rcui
t
eT
oea
rpho
ne
Voi
cesi
gnal
Th
is f
igu
re
is a
sc
he
ma
tic o
f a
tele
ph
on
e ci
rcu
it us
ing
anad
aptiv
e fil
ter
to c
ance
l ec
ho.
The
ada
ptiv
e fil
ter
is d
epic
ted
as a
box
; th
e sl
ante
d ar
row
rep
rese
nts
the
adju
stab
le w
eigh
ts.
The
sign
al p
ower
, {.
s2 }, i
s de
term
ined
by
the
sour
ce o
f th
e vo
ice
say,
som
e am
plif
ier
at t
he t
elep
hone
sw
itch
ing
stat
ion
loca
l to
the
sen
der.
Thu
s,
is n
ot d
irec
tly a
ffec
ted
by c
hang
es i
n
The
adap
tive
filte
r at
tem
pts
to m
inim
ize
and
, in
doi
ng s
o,
min
imiz
es
((n'
- t
he p
ower
of
the
unca
ncel
ed n
oise
on
the
outg
oing
lin
e.Si
nce
ther
e is
onl
y on
e in
put
to t
he a
dapt
ive
filt
er,
the
devi
ce w
ould
be
conf
igur
ed a
s a
tran
sver
se f
ilte
r. W
idro
w a
nd S
tear
ns [
9] s
ugge
st s
ampl
ing
the
inco
min
g si
gnal
at
a ra
te o
f 8
KH
z an
d us
ing
128
wei
ght
valu
es.
2.3.
2 O
ther
Ap
plic
atio
ns
Rat
her
than
go
into
the
det
ails
of t
he m
any
appl
icat
ions
tha
t can
be
addr
esse
d by
thes
e ad
aptiv
e fi
lter
s, w
e re
fer
you
once
aga
in t
o th
e ex
cell
ent
text
by
Wid
row
and
Stea
rns.
In
thi
s se
ctio
n, w
e sh
all
sim
ply
sugg
est
a fe
w b
road
are
as w
here
adap
tive
filte
rs c
an b
e us
ed i
n ad
ditio
n to
the
ech
o-ca
ncel
latio
n ap
plic
atio
n w
eha
ve d
iscu
ssed
.Fi
gure
sho
ws
an a
dapt
ive
filt
er th
at i
s us
ed p
redi
ct th
e fu
ture
val
ue o
fa
sign
al b
ased
on
its
pres
ent
valu
e. A
sec
ond
exam
ple
is s
how
n in
Fig
ure
In t
his
exam
ple,
the
ada
ptiv
e fi
lter
lea
rns
to r
epro
duce
the
out
put
from
som
epl
ant
base
d on
inp
uts
to t
he s
yste
m.
Thi
s co
nfig
urat
ion
has
man
y us
es a
s an
adap
tive
con
trol
sys
tem
. T
he p
lant
cou
ld r
epre
sent
man
y th
ings
, in
clud
ing
ahu
man
ope
rato
r. I
n th
at c
ase,
the
ada
ptiv
e fi
lter
cou
ld l
earn
how
to
resp
ond
toch
angi
ng c
ondi
tion
s by
wat
chin
g th
e hu
man
ope
rato
r. E
vent
uall
y, s
uch
a de
vice
mig
ht r
esul
t in
an
auto
mat
ed c
ontr
ol s
yste
m,
leav
ing
the
hum
an f
ree
for
mor
eim
port
ant
Ano
ther
use
ful
appl
icat
ion
of t
hese
dev
ices
is
in
adap
tive
bea
m-f
orm
ing
ante
nna
arra
ys.
Alt
houg
h th
e te
rm a
nten
na i
s us
uall
y as
soci
ated
wit
h el
ectr
o-
as
trai
ning
ano
ther
ada
ptiv
e fi
lter
wit
h th
e St
anda
rd &
Poo
rs 5
00.
2.3
Ap
plic
atio
ns o
f A
dap
tive S
igna
l P
roce
ssin
g71
Cur
rent
sig
nal
Pre
dict
ion o
f
curr
ent
sign
al
Pas
t sig
nal
Fig
ure
2.1
6
Th
is s
chem
atic
show
s an
adaptiv
e f
ilter
used
to p
redic
t si
gnal
valu
es.
The
input s
ignal u
sed to
tra
in the n
etw
ork
is a
dela
yed
valu
e of
the
actu
al s
igna
l; th
at i
s, i
t is
the
sig
nal
at s
ome
past
time.
The
exp
ecte
d ou
tput
is
the
curr
ent
valu
e of
the
sig
nal.
The
ada
ptiv
e fil
ter
atte
mpt
s to
min
imiz
e th
e e
rro
r be
twee
n its
outp
ut a
nd t
he c
urre
nt s
igna
l, ba
sed
on a
n in
put
of th
e si
gnal
valu
e fr
om s
ome
time
in t
he p
ast.
Onc
e th
e fil
ter
is c
orr
ect
lypr
edic
ting
the
curr
ent
sign
al
base
d on
the
pa
st s
igna
l, th
ecu
rrent
sign
al c
an b
e u
sed d
ire
ctly
as
an i
nput
with
out
the
dela
y.
The f
ilter
will
then m
ake
a p
redic
tion o
f th
e f
utu
resi
gnal
valu
e.
Inpu
t si
gnal
s
Pre
dict
ion
of
plan
t ou
tput
Figu
reT
his
e
xam
ple
sh
ow
s an
a
da
ptiv
e fil
ter
used
to
m
odel
th
eo
utp
ut
fro
m a
sys
tem
, ca
lled t
he
pla
nt.
Inputs
to t
he f
ilter
are
the
sam
e as
tho
se t
o th
e pl
ant.
The
filt
er
ad
just
s its
wei
ghts
base
d o
n t
he d
iffere
nce
bet
wee
n i
ts o
utp
ut
and t
he
outp
ut
ofth
e p
lant.
72A
dal
ine
and
mag
neti
c ra
diat
ion,
we
broa
den
the
defi
niti
on h
ere
to i
nclu
de a
ny s
pati
al a
rray
of s
enso
rs.
The
bas
ic t
ask
here
is
to l
earn
to
stee
r th
e ar
ray.
At
any
give
n ti
me,
a si
gnal
may
be
arri
ving
fro
m a
ny g
iven
dir
ectio
n, b
ut a
nten
nae
usua
lly
are
dire
ctio
nal
in t
heir
rec
epti
on c
hara
cter
isti
cs:
The
y re
spon
d to
sig
nals
in
som
edi
rect
ions
, bu
t no
t in
oth
ers.
T
he a
nten
na a
rray
wit
h ad
aptiv
e fi
lter
s le
arns
to
adju
st i
ts d
irec
tion
al c
hara
cter
isti
cs i
n or
der
to r
espo
nd t
o th
e in
com
ing
sign
alno
mat
ter
wha
t th
e di
rect
ion
is,
whi
le r
educ
ing
its
resp
onse
to
unw
ante
d no
ise
sign
als
com
ing
in f
rom
oth
er d
irec
tion
s.O
f co
urse
, w
e ha
ve o
nly
touc
hed
on t
he n
umbe
r of
app
licat
ions
for
the
sede
vice
s.
Unl
ike
man
y ot
her
neur
al-n
etw
ork
arch
itec
ture
s, t
his
is a
rel
ativ
ely
mat
ure
devi
ce w
ith
a lo
ng h
isto
ry o
f su
cces
s.
In t
he n
ext
sect
ion,
we
repl
ace
the
bina
ry o
utpu
t co
ndit
ion
on t
he A
LC
cir
cuit
so
that
the
lat
ter
beco
mes
, on
ceag
ain,
the
com
plet
e A
dali
ne.
2.4
TH
E M
AD
AL
INE
As
you
can
see
from
the
dis
cuss
ion
in C
hapt
er 1
, th
e A
dalin
e re
sem
bles
the
perc
eptr
on c
lose
ly;
it a
lso
has
som
e of
the
sam
e li
mit
atio
ns a
s th
e pe
rcep
tron
.Fo
r ex
ampl
e, a
tw
o-in
put
Ada
line
can
not
com
pute
the
XO
R f
unct
ion.
C
om-
bini
ng A
dalin
es i
n a
laye
red
stru
ctur
e ca
n ov
erco
me
this
dif
ficu
lty,
as
we
did
inC
hapt
er 1
wit
h th
e pe
rcep
tron
. Su
ch a
str
uctu
re i
s il
lust
rate
d in
Fig
ure
2.18
.
Exe
rcis
e 2.
5:
Wha
t lo
gic
func
tion
is
bein
g co
mpu
ted
by t
he s
ingl
e A
dali
ne i
nth
e ou
tput
lay
er o
f Fi
gure
Con
stru
ct a
thr
ee-i
nput
Ada
line
that
com
pute
sth
e m
ajor
ity
func
tion
.
2.4.
1 M
adal
ine
Arc
hit
ectu
re
Mad
alin
e is
the
acr
onym
for
Man
y A
dali
nes.
Arr
ange
d in
a m
ulti
laye
red
arch
i-te
ctur
e as
ill
ustr
ated
in
Figu
re 2
.19,
the
Mad
alin
e re
sem
bles
the
gen
eral
neu
ral-
netw
ork
stru
ctur
e sh
own
in C
hapt
er I
n th
is c
onfi
gura
tion
, th
e M
adal
ine
coul
dbe
pre
sent
ed w
ith
a la
rge-
dim
ensi
onal
inp
ut t
he p
ixel
val
ues
from
a ra
ster
sca
n.
Wit
h su
itab
le t
rain
ing,
the
net
wor
k co
uld
be t
augh
t to
res
pond
wit
h a
bina
ry o
n on
e of
sev
eral
out
put
node
s, e
ach
of w
hich
cor
resp
onds
to
a di
ffer
ent
cate
gory
of
inpu
t im
age.
E
xam
ples
of
such
cat
egor
izat
ion
are
dog,
arm
adil
lo,
jave
lina
} an
d {F
logg
er,
Tom
Cat
, E
agle
,
In s
uch
ane
twor
k, e
ach
of f
our
node
s in
the
out
put
laye
r co
rres
pond
s to
a s
ingl
e cl
ass.
For
a gi
ven
inpu
t pa
ttern
, a
node
wou
ld h
ave
a o
utpu
t if
the
inp
ut p
atte
rnco
rres
pond
ed t
o th
e cl
ass
repr
esen
ted
by t
hat
part
icul
ar n
ode.
T
he o
ther
thr
eeno
des
wou
ld h
ave
a o
utpu
t. If
the
inp
ut p
atte
rn w
ere
not
a m
embe
r of
any
know
n cl
ass,
the
res
ults
fro
m t
he n
etw
ork
coul
d be
am
bigu
ous.
To t
rain
su
ch
a ne
twor
k,
we
mig
ht b
e te
mpt
ed t
o be
gin
wit
h th
e L
MS
algo
rith
m a
t th
e ou
tput
lay
er.
Sinc
e th
e ne
twor
k is
pre
sum
ably
tra
ined
wit
hpr
evio
usly
ide
ntif
ied
inpu
t pa
tter
ns,
the
desi
red
outp
ut v
ecto
r is
kno
wn.
W
hat
2.4
T
he M
adal
ine
73
Fig
ure
2.1
8 M
an
y A
da
line
s (t
he
Ma
da
line
) ca
n
com
pu
te
the
X
OR
fun
ctio
n o
f tw
o i
np
uts
. N
ote t
he a
dd
itio
n o
f th
e b
ias
term
s to
ea
ch A
da
line
. A
pos
itive
an
alo
g ou
tput
fro
m a
n A
LC
re
sults
in a
+1 o
utpu
t fro
m t
he a
sso
cia
ted A
da
line
; a n
eg
ativ
e a
na
log
outp
ut r
esu
lts i
n a
Lik
ew
ise
, a
ny
inpu
ts t
o th
e d
evi
ce t
ha
ta
re b
ina
ry i
n na
ture
mus
t us
e ±1
ra
the
r th
an
1 an
d 0.
we
do n
ot k
now
is
the
desi
red
outp
ut f
or a
giv
en n
ode
on o
ne o
f th
e hi
dden
laye
rs.
Furt
herm
ore,
the
LM
S al
gori
thm
wou
ld o
pera
te o
n th
e an
alog
out
puts
of t
he A
LC
, no
t on
the
bip
olar
out
put
valu
es o
f th
e A
dali
ne.
For
thes
e re
ason
s,a
diff
eren
t tr
aini
ng s
trat
egy
has
been
dev
elop
ed f
or t
he M
adal
ine.
2.4
.2
Th
e T
rain
ing
Alg
ori
thm
It i
s po
ssib
le t
o de
vise
a m
etho
d of
tra
inin
g a
str
uctu
re b
ased
on
the
LM
S al
gori
thm
; ho
wev
er,
the
met
hod
reli
es o
n re
plac
ing
the
line
ar t
hres
hold
outp
ut f
unct
ion
wit
h a
cont
inuo
usly
dif
fere
ntia
ble
func
tion
(th
e th
resh
old
func
-tio
n is
dis
cont
inuo
us a
t 0;
hen
ce,
it is
not
dif
fere
ntia
ble
ther
e).
We
tak
e up
the
stud
y of
thi
s m
etho
d in
the
nex
t ch
apte
r.
For
now
, w
e co
nsid
er a
met
hod
know
n as
Mad
alin
e ru
le I
I (M
RII
). T
he o
rigi
nal
Mad
alin
e ru
le w
as a
n ea
rlie
r
74
Adal
ine a
nd
Mad
alin
e
Out
put l
ayer
of
Hid
den la
yer
of M
adal
ines
Figu
re
Man
y A
da
line
s ca
n b
e jo
ine
d in
a l
ayer
ed n
eura
l ne
twor
ksu
ch a
s th
is o
ne.
met
hod
that
we
shal
l no
t di
scus
s he
re.
Det
ails
can
be
foun
d in
ref
eren
ces
give
nat
the
end
of
this
cha
pter
. r
esem
bles
a p
roce
dure
wit
h ad
ded
inte
llige
nce
in t
hefo
rm o
f a
min
imum
dis
turb
ance
pri
ncip
le.
Sinc
e th
e ou
tput
of
the
netw
ork
is a
ser
ies
of b
ipol
ar u
nits
, tr
aini
ng a
mou
nts
to r
educ
ing
the
num
ber
of i
ncor
-re
ct o
utpu
t no
des
for
each
tra
inin
g in
put
patte
rn.
The
min
imum
dis
turb
ance
prin
cipl
e en
forc
es t
he n
otio
n th
at t
hose
nod
es t
hat
can
affe
ct t
he o
utpu
t er
ror
whi
le i
ncur
ring
the
lea
st c
hang
e in
the
ir w
eigh
ts s
houl
d ha
ve p
rece
denc
e in
the
lear
ning
pro
cedu
re.
Thi
s pr
inci
ple
is e
mbo
died
in
the
foll
owin
g al
gori
thm
:
1.
App
ly a
tra
inin
g ve
ctor
to
the
inpu
ts o
f th
e M
adal
ine
and
prop
agat
e it
thro
ugh
to t
he o
utpu
t un
its.
2. C
ount
the
num
ber
of i
ncor
rect
val
ues
in t
he o
utpu
t la
yer;
cal
l th
is n
umbe
rth
e er
ror.
3.
For
all
units
on
the
outp
ut l
ayer
,
a.
Sele
ct t
he f
irst
pre
viou
sly
unse
lect
ed n
ode
who
se a
nalo
g ou
tput
is
clos
-es
t to
zer
o.
(Thi
s no
de i
s th
e no
de t
hat
can
reve
rse
its
bipo
lar
outp
ut
2.4
T
he M
adal
ine
75
wit
h th
e le
ast
chan
ge i
n it
s t
he t
erm
min
imum
dis
tur-
banc
e.)
b. C
hang
e th
e w
eigh
ts o
n th
e se
lect
ed u
nit
such
tha
t th
e bi
pola
r ou
tput
of
the
unit
cha
nges
.c.
Pr
opag
ate
the
inpu
t ve
ctor
for
war
d fr
om t
he i
nput
s to
the
out
puts
onc
eag
ain.
d. I
f th
e w
eigh
t ch
ange
res
ults
in
a re
duct
ion
in t
he n
umbe
r of
err
ors,
acce
pt t
he w
eigh
t ch
ange
; ot
herw
ise,
res
tore
the
ori
gina
l
4. R
epea
t st
ep 3
for
all
laye
rs e
xcep
t th
e in
put
laye
r.
5.
For
all
unit
s on
the
out
put
laye
r,a.
Sel
ect
the
prev
ious
ly u
nsel
ecte
d pa
ir o
f un
its
who
se a
nalo
g ou
tput
s ar
ecl
oses
t to
zer
o.b.
App
ly a
wei
ght
corr
ectio
n to
bot
h un
its,
in
orde
r to
cha
nge
the
bipo
lar
outp
ut o
f ea
ch.
c.
Prop
agat
e th
e in
put
vect
or f
orw
ard
from
the
inp
uts
to t
he o
utpu
ts.
d. I
f th
e w
eigh
t ch
ange
res
ults
in
a re
duct
ion
in t
he n
umbe
r of
err
ors,
acce
pt t
he w
eigh
t ch
ange
; ot
herw
ise,
res
tore
the
ori
gina
l w
eigh
ts.
6. R
epea
t st
ep 5
for
all
laye
rs e
xcep
t th
e in
put
laye
r.
If n
eces
sary
, th
e se
quen
ce i
n st
eps
5 an
d 6
can
be r
epea
ted
wit
h tr
iple
tsof
uni
ts,
or q
uadr
uple
ts o
f un
its,
or
even
lar
ger
com
bina
tions
, un
til
satis
fact
ory
resu
lts a
re o
btai
ned.
Pr
elim
inar
y in
dica
tions
are
tha
t pa
irs
are
adeq
uate
for
mod
est-
size
d ne
twor
ks w
ith
up t
o 25
uni
ts p
er l
ayer
At
the
time
of t
his
wri
ting,
the
was
stil
l un
derg
oing
exp
erim
enta
tion
to d
eter
min
e its
con
verg
ence
cha
ract
eris
tics
and
oth
er p
rope
rtie
s.
Mor
eove
r, a
new
lea
rnin
g al
gori
thm
, h
as b
een
deve
lope
d. i
s si
mil
ar t
o M
RII
,bu
t the
ind
ivid
ual u
nits
hav
e a
cont
inuo
us o
utpu
t fun
ctio
n, r
athe
r th
an th
e bi
pola
rth
resh
old
func
tion
In
the
next
sec
tion,
we
shal
l us
e a
Mad
alin
e ar
chite
ctur
eto
exa
min
e a
spec
ific
pro
blem
in
patte
rn r
ecog
niti
on.
2.4.
3 A
Mad
alin
e fo
r T
ran
slat
ion
-In
vari
ant
Pat
tern
Rec
og
niti
on
Var
ious
Mad
alin
e st
ruct
ures
hav
e be
en u
sed
rece
ntly
to
dem
onst
rate
the
app
li-
cabi
lity
of
this
arc
hite
ctur
e to
ada
ptiv
e pa
ttern
rec
ogni
tion
havi
ng t
he p
rope
rtie
sof
tra
nsla
tion
inv
aria
nce,
rot
atio
n in
vari
ance
, an
d sc
ale
inva
rian
ce.
The
se t
hree
prop
ertie
s ar
e es
sent
ial
to a
ny r
obus
t sy
stem
tha
t w
ould
be
calle
d on
to
rec-
ogni
ze o
bjec
ts i
n th
e fi
eld
of v
iew
of
opti
cal
or i
nfra
red
sens
ors,
for
exa
mpl
e.R
emem
ber,
how
ever
, th
at e
ven
hum
ans
do n
ot a
lway
s in
stan
tly
reco
gniz
e ob
-je
cts
that
hav
e be
en r
otat
ed t
o un
fam
ilia
r or
ient
atio
ns,
or t
hat
have
bee
n sc
aled
sign
ific
antl
y sm
alle
r or
lar
ger
than
the
ir e
very
day
size
. T
he p
oint
is
that
the
rem
ay b
e al
tern
ativ
es t
o tr
aini
ng i
n in
stan
tane
ous
reco
gniti
on a
t al
l an
gles
and
scal
e fa
ctor
s.
Be
that
as
it m
ay,
it i
s po
ssib
le t
o bu
ild
neur
al-n
etw
ork
devi
ces
that
exh
ibit
the
se c
hara
cter
isti
cs t
o so
me
degr
ee.
76
Ad
alin
e a
nd
Figu
re 2
.20
show
s a
port
ion
of a
net
wor
k th
at i
s us
ed t
o im
plem
ent
tran
sla-
tion
-inv
aria
nt r
ecog
niti
on o
f a
patte
rn T
he r
etin
a is
a 5
-by-
5-pi
xel
arra
y on
whi
ch b
it-m
appe
d re
pres
enta
tion
of p
atte
rns,
suc
h as
the
let
ters
of
the
alph
abet
,ca
n be
pla
ced.
T
he p
orti
on o
f th
e ne
twor
k sh
own
is c
alle
d a
slab
. U
nlik
e a
laye
r, a
sla
b do
es n
ot c
omm
unic
ate
wit
h ot
her
slab
s in
the
net
wor
k, a
s w
ill
bese
en s
hort
ly.
Eac
h A
dali
ne i
n th
e sl
ab r
ecei
ves
the
iden
tical
25
inpu
ts f
rom
the
retin
a, a
nd c
ompu
tes
a bi
pola
r ou
tput
in
the
usua
l fa
shio
n; h
owev
er,
the
wei
ghts
on t
he 2
5 A
dali
nes
shar
e a
uniq
ue r
elat
ions
hip.
Con
side
r th
e w
eigh
ts o
n th
e to
p-le
ft A
dalin
e as
bei
ng a
rran
ged
in a
squ
are
mat
rix
dupl
icat
ing
the
pixe
l ar
ray
on t
he r
etin
a.
The
Ada
line
to
the
imm
edia
te
Mad
alin
e sl
ab
Ret
ina
Figu
re 2
.20
T
his
sin
gle
sla
b of
Ad
alin
es
will
giv
e th
e s
ame
outp
ut
(eith
er+
1
or -1
) fo
r a p
art
icula
r patte
rn o
n t
he r
etin
a,
regard
less
of
the
ho
rizo
nta
l or
vert
ica
l a
lign
me
nt
of th
at
pattern
on
the
retin
a.
All
25
ind
ivid
ua
l A
da
line
s a
re c
on
ne
cte
d t
o a
single
Adalin
e t
hat
com
pute
s th
e m
ajo
rity
fu
nct
ion
: If
mos
tof
the i
nputs
are
+1,
the m
ajo
rity
ele
ment
resp
onds
with
a+
1
outp
ut.
The n
etw
ork
de
rive
s its
tra
nsl
atio
n-i
nva
ria
nce
pro
pe
rtie
s fr
om
th
e
pa
rtic
ula
r co
nfig
ura
tion
of
the
weig
hts
.S
ee the t
ext
for
deta
ils.
2.4
T
he M
adal
ine
77
righ
t of
the
top
-lef
t pi
xel
has
the
iden
tica
l se
t of
wei
ght
valu
es,
but
tran
slat
edon
e pi
xel
to t
he r
ight
: T
he r
ight
mos
t co
lum
n of
wei
ghts
on
the
firs
t un
it w
raps
arou
nd t
o th
e le
ft t
o be
com
e th
e le
ftm
ost
colu
mn
on t
he s
econ
d un
it.
Sim
ilar
ly,
the
unit
bel
ow t
he t
op-l
eft
unit
als
o ha
s th
e id
enti
cal
wei
ghts
, bu
t tr
ansl
ated
one
pixe
l do
wn.
T
he b
otto
m r
ow o
f w
eigh
ts o
n th
e fi
rst
unit
bec
omes
the
top
row
of
the
unit
und
er i
t. T
his
tran
slat
ion
cont
inue
s ac
ross
eac
h ro
w a
nd d
own
each
col
umn
in a
sim
ilar
man
ner.
Fig
ure
2.21
ill
ustr
ates
som
e of
thes
e w
eigh
tm
atri
ces.
B
ecau
se o
f th
is r
elat
ions
hip
amon
g th
e w
eigh
t m
atri
ces,
a
sing
lepa
tter
n on
the
ret
ina
wil
l el
icit
iden
tical
res
pons
es f
rom
the
sla
b, i
ndep
ende
nt
Key
wei
ght m
atrix
: top
row
, le
ft co
lum
n W
eigh
t m
atrix
: top
row
, 2n
d co
lum
nw
w
w
w
12
13
14
15
Wei
ght m
atrix
: 2n
d ro
w,
left
colu
mn
W
W
W
W51
52
53
45
W
W
W
W22
23
24
25
w12
W13
35 4544
W32
33
34
35
W
W42
43
44
45
Wei
ght m
atrix
: 5th
row
, 5t
h co
lum
n
Fig
ure
2.2
1
The w
eig
ht
matr
ix i
n t
he u
pper
left i
s th
e k
ey w
eig
ht
matr
ix.
All
othe
r w
eigh
t m
atric
es o
n th
e sl
ab a
re d
eriv
ed f
rom
thi
sm
atr
ix.
The
matr
ix t
o th
e
right
of
the
key
weig
ht
matr
ixre
pre
sents
the m
atr
ix o
n the
direct
ly to
the r
ight o
f the
one w
ith t
he k
ey w
eigh
t m
atr
ix.
Not
ice that
the f
ifth c
olum
nof
the k
ey w
eigh
t m
atrix
has
wra
pped
aro
und t
o b
ecom
e th
efir
st c
olu
mn,
with
the o
ther
colu
mns
shift
ing o
ne s
pace
to
the r
ight.
The m
atr
ix b
elo
w t
he k
ey
weig
ht
matr
ix is
the
one o
n t
he A
da
line d
irect
ly b
elo
w t
he A
da
line w
ith t
he k
ey
we
igh
t m
atr
ix.
The
ma
trix
dia
gonal
to t
he k
ey
weig
ht
matr
ixre
pre
sents
th
e m
atr
ix o
n t
he A
da
line a
t th
e l
ower
rig
ht o
f th
esl
ab.
78
Adal
ine a
nd
of th
e pa
ttern
's t
rans
lati
onal
pos
itio
n on
the
ret
ina.
We
enco
urag
e yo
u to
ref
lect
on t
his
resu
lt fo
r a
mom
ent
(per
haps
sev
eral
mom
ents
), t
o co
nvin
ce y
ours
elf
ofits
val
idit
y.T
he m
ajor
ity
node
is
a si
ngle
Ada
line
tha
t co
mpu
tes
a bi
nary
out
put
base
don
the
out
puts
of
the
maj
orit
y of
the
Ada
lines
con
nect
ing
to i
t. B
ecau
se o
f th
etr
ansl
atio
nal
rela
tion
ship
am
ong
the
wei
ght
vect
ors,
the
pla
cem
ent
of a
par
ticu
lar
patte
rn a
t an
y lo
catio
n on
the
ret
ina
wil
l re
sult
in t
he i
dent
ical
out
put
from
the
maj
ority
ele
men
t (w
e im
pose
the
res
tric
tion
that
pat
tern
s th
at e
xten
d be
yond
the
reti
na b
ound
arie
s w
ill
wra
p ar
ound
to
the
oppo
site
sid
e, j
ust
as t
he v
ario
usw
eigh
t m
atri
ces
are
deri
ved
from
the
key
wei
ght
Of
cour
se,
a pa
ttern
diff
eren
t fr
om t
he f
irst
may
elic
it a
diff
eren
t re
spon
sefr
om t
he m
ajor
ity e
lem
ent.
Bec
ause
onl
y tw
o re
spon
ses
are
poss
ible
, the
sla
b ca
n di
ffer
entia
te tw
o cl
asse
s on
inpu
t pa
tter
ns.
In t
erm
s of
hyp
ersp
ace,
a s
lab
is c
apab
le o
f di
vidi
ngin
to t
wo
regi
ons.
To o
verc
ome
the
limita
tion
of o
nly
two
poss
ible
cla
sses
, th
e re
tina
can
beco
nnec
ted
to m
ultip
le s
labs
, eac
h ha
ving
dif
fere
nt k
ey w
eigh
t mat
rice
s (W
idro
wan
d W
inte
r's t
erm
for
the
wei
ght
mat
rix
on t
he t
op-l
eft
elem
ent
of e
ach
slab
).G
iven
the
bin
ary
natu
re o
f th
e ou
tput
of
each
sla
b, a
sys
tem
of
n sl
abs
coul
ddi
ffer
enti
ate
2"
diff
eren
t pa
ttern
cla
sses
. Fi
gure
2.2
2 sh
ows
four
suc
h sl
abs
prod
ucin
g a
four
-dim
ensi
onal
out
put c
apab
le o
f dis
tingu
ishi
ng d
iffe
rent
inp
ut-
patte
rn c
lass
es w
ith t
rans
latio
nal
inva
rian
ce.
Let
's r
evie
w t
he b
asic
ope
ratio
n of
the
tra
nsla
tion
inv
aria
nce
netw
ork
inte
rms
of a
spe
cifi
c ex
ampl
e. C
onsi
der t
he le
tters
A P
, as
the
inp
ut p
atte
rns
we
wou
ld l
ike
to i
dent
ify
rega
rdle
ss o
f th
eir
or
left
-rig
ht t
rans
lati
onon
the
5-b
y-5-
pixe
l re
tina.
The
se t
rans
late
d re
tina
patte
rns
are
the
inpu
ts t
o th
esl
abs
of t
he n
etw
ork.
E
ach
retin
a pa
ttern
res
ults
in
an o
utpu
t pa
ttern
fro
m t
hein
vari
ance
net
wor
k th
at m
aps
to o
ne o
f th
e 16
inp
ut c
lass
es (
in t
his
case
, ea
chcl
ass
repr
esen
ts a
let
ter)
. B
y us
ing
a lo
okup
tab
le,
or o
ther
met
hod,
we
can
asso
ciat
e th
e 16
pos
sibl
e ou
tput
s fr
om t
he i
nvar
ianc
e ne
twor
k w
ith o
ne o
f th
e16
pos
sibl
e le
tters
tha
t ca
n be
ide
ntif
ied
by t
he n
etw
ork.
So f
ar,
noth
ing
has
been
sai
d co
ncer
ning
the
val
ues
of t
he w
eigh
ts o
n th
eA
dali
nes
of t
he v
ario
us s
labs
in
the
syst
em.
Tha
t is
bec
ause
it
is n
ot a
ctua
lly
nece
ssar
y to
tra
in t
hose
nod
es i
n th
e us
ual
sens
e.
In f
act,
each
key
wei
ght
mat
rix
can
be c
hose
n at
ran
dom
, pro
vide
d th
at e
ach
inpu
t-pa
ttern
clas
s re
sult
ina
uniq
ue o
utpu
t ve
ctor
fro
m t
he i
nvar
ianc
e ne
twor
k.
Usi
ng t
he e
xam
ple
of t
hepr
evio
us p
arag
raph
, an
y tr
ansl
atio
n of
one
of t
he l
ette
rs s
houl
d re
sult
in t
he s
ame
outp
ut f
rom
the
inv
aria
nce
netw
ork.
Fu
rthe
rmor
e, a
ny p
atte
rn f
rom
a d
iffe
rent
clas
s (i
.e.,
a di
ffer
ent
lette
r) m
ust
resu
lt in
a d
iffe
rent
out
put
vect
or f
rom
the
netw
ork.
Thi
s re
quir
emen
t m
eans
tha
t, if
you
pic
k a
rand
om k
ey w
eigh
t m
atri
xfo
r a
part
icul
ar s
lab
and
find
tha
t tw
o le
tters
giv
e th
e sa
me
outp
ut p
atte
rn,
you
can
sim
ply
pick
a d
iffe
rent
wei
ght
mat
rix.
As
an a
lter
nati
ve t
o ra
ndom
sel
ectio
n of
key
wei
ght
mat
rice
s, i
t m
ay b
epo
ssib
le t
o op
tim
ize
sele
ctio
n by
em
ploy
ing
a tr
aini
ng p
roce
dure
bas
ed o
n th
e I
nves
tiga
tion
s in
thi
s ar
ea a
re o
ngoi
ng a
t th
e ti
me
of t
his
wri
ting
2.5
S
imula
ting
th
e A
dal
ine
79
Ret
ina
Figu
re 2
.22
Eac
h of t
he four
slab
s in
the
sys
tem
depic
ted
her
e w
ill p
roduce
a +
1 o
r a —
1 o
utpu
t va
lue f
or
eve
ry p
atte
rn t
hat
appe
ars
onth
e r
etin
a.
The
outp
ut
vect
or
is a
fo
ur-
dig
it b
ina
ry n
umbe
r,so
the
sys
tem
ca
n p
ote
ntia
lly d
iffe
ren
tiate
up
to 1
6 di
ffere
ntcl
ass
es
of i
nput
pat
tern
s.
L
2.5
SIM
UL
AT
ING
TH
E A
DA
LIN
E
As
we
shal
l fo
r th
e im
plem
enta
tion
of a
ll ot
her
netw
ork
sim
ulat
ors
we
wil
lpr
esen
t, w
e sh
all
begi
n th
is s
ectio
n by
des
crib
ing
how
the
gen
eral
dat
a st
ruc-
ture
s ar
e us
ed t
o m
odel
the
Ada
line
uni
t an
d M
adal
ine
netw
ork.
Onc
e th
e ba
sic
arch
itec
ture
has
bee
n pr
esen
ted,
we
wil
l des
crib
e th
e al
gori
thm
ic p
roce
ss n
eede
dto
pro
paga
te s
igna
ls t
hrou
gh t
he A
dali
ne.
The
sec
tion
conc
lude
s w
ith
a di
scus
-si
on o
f th
e al
gori
thm
s ne
eded
to
caus
e th
e A
dalin
e to
sel
f-ad
apt
acco
rdin
g to
the
lear
ning
law
s de
scri
bed
prev
ious
ly.
2.5.
1 A
dal
ine
Dat
a S
tru
ctu
res
It i
s ap
prop
riat
e th
at t
he A
dali
ne i
s th
e fi
rst
test
of
the
sim
ulat
or d
ata
stru
ctur
esw
e pr
esen
ted
in C
hapt
er 1
for
tw
o re
ason
s:
1.
Sinc
e th
e fo
rwar
d pr
opag
atio
n of
sig
nals
thr
ough
the
sin
gle
Ada
line
is
vir-
tual
ly i
dent
ical
to
the
forw
ard
prop
agat
ion
proc
ess
in m
ost
of t
he o
ther
netw
orks
we
wil
l st
udy,
it
is b
enef
icia
l fo
r us
to
obse
rve
the
Ada
line
to
80
Ad
alin
e a
nd
gain
a b
ette
r un
ders
tand
ing
of w
hat
is h
appe
ning
in
each
uni
t of
a l
arge
rne
twor
k.
2. B
ecau
se t
he A
dali
ne i
s no
t a
netw
ork,
its
im
plem
enta
tion
exe
rcis
es t
heve
rsat
ility
of
the
netw
ork
stru
ctur
es w
e ha
ve d
efin
ed.
As
we
have
alr
eady
se
en,
the
Ada
line
is
only
a
sing
le p
roce
ssin
g un
it.
The
refo
re,
som
e of
the
gen
eral
ity w
e bu
ilt
into
our
net
wor
k st
ruct
ures
wil
l no
tbe
req
uire
d. S
peci
fica
lly,
the
re w
ill
be n
o re
al n
eed
to h
andl
e m
ulti
ple
unit
s an
dla
yers
of
unit
s fo
r th
e A
dalin
e.
Nev
erth
eles
s, w
e w
ill
incl
ude
the
use
of t
hose
stru
ctur
es,
beca
use
we
wou
ld l
ike
to b
e ab
le t
o ex
tend
the
Ada
line
easi
ly i
nto
the
Mad
alin
e.W
e be
gin
by d
efin
ing
our
netw
ork
reco
rd a
s a
stru
ctur
e th
at w
ill c
onta
inal
l th
e pa
ram
eter
s th
at w
ill b
e us
ed g
loba
lly,
as w
ell
as p
oint
ers
to l
ocat
e th
edy
nam
ic a
rray
s th
at w
ill
cont
ain
the
netw
ork
data
. In
the
cas
e of
the
Ada
line,
a go
od c
andi
date
str
uctu
re f
or t
his
reco
rd w
ill t
ake
the
form
record Adaline =
: float;
"layer;
output :
end record
for stability
to input
to output
Not
e th
at,
even
tho
ugh
ther
e is
onl
y on
e un
it in
the
Ada
line,
we
wil
l us
etw
o la
yers
to
mod
el t
he n
etw
ork.
Thu
s, t
he i
np
ut
and
ou
tpu
t po
inte
rs w
illpo
int
to d
iffe
rent
lay
er r
ecor
ds.
We
do t
his
beca
use
we
will
use
the
in
pu
tla
yer
as s
tora
ge f
or h
oldi
ng th
e in
put
sign
al v
ecto
r to
the
Ada
line.
The
re w
ill b
eno
con
nect
ions
ass
ocia
ted
wit
h th
is l
ayer
, as
the
inp
ut w
ill
be p
rovi
ded
by s
ome
othe
r pro
cess
in
the
syst
em (
e.g.
, a
time-
mul
tiple
xed
con
vert
er,
or a
n ar
ray
of s
enso
rs).
Con
vers
ely,
the
ou
tpu
t la
yer
will
con
tain
one
wei
ght
arra
y to
mod
el t
heco
nnec
tions
bet
wee
n th
e in
pu
t an
d th
e o
utp
ut
(rec
all t
hat o
ur d
ata
stru
ctur
espr
esum
e th
at P
Es p
roce
ss i
nput
con
nect
ions
pri
mar
ily).
K
eepi
ng i
n m
ind
that
we
wou
ld l
ike
to e
xten
d th
is s
truc
ture
eas
ily t
o ha
ndle
the
Mad
alin
e ne
twor
k,w
e w
ill
reta
in t
he i
ndir
ectio
n to
the
con
nect
ion
wei
ght
arra
y pr
ovid
ed b
y th
e a
rray
des
crib
ed i
n C
hapt
er
Not
ice
that
, in
the
cas
e of
the
Ada
line,
how
ever
, th
e a
rray
will
con
tain
onl
y on
e va
lue,
the
poin
ter
to t
he i
nput
con
nect
ion
arra
y.T
here
is
one
othe
r th
ing
to c
onsi
der
that
may
var
y be
twee
n A
dalin
e un
its.
As
we
have
see
n pr
evio
usly
, th
ere
are
two
part
s to
the
Ada
line
str
uctu
re:
the
linea
r A
LC
and
the
bip
olar
Ada
line
unit
s.
To
dist
ingu
ish
betw
een
them
, w
ede
fine
an
enum
erat
ed t
ype
to c
lass
ify
each
Ada
line
neu
ron:
type NODE_TYPE :
We
now
hav
e ev
eryt
hing
we
need
to
defi
ne t
he l
ay
er
reco
rd s
truc
ture
for
the
Ada
line
. A
pro
toty
pe s
truc
ture
for
thi
s re
cord
is
as f
ollo
ws.
2.5
Sim
ula
ting
th
e
record layer =
activation : NODE_TYPE
of Adaline
to unit output
weights :
access to weight
end record
Fin
ally
, th
ree
dyna
mic
ally
allo
cate
d ar
rays
are
nee
ded
to c
onta
in t
he o
utpu
tof
the
Ada
line
uni
t, th
e a
nd t
he c
onne
ctio
n w
eig
hts
val
ues.
We
wil
l no
t sp
ecif
y th
e st
ruct
ure
of t
hese
arr
ays,
oth
er t
han
to i
ndic
ate
that
the
ou
ts a
nd w
eig
hts
arr
ays
wil
l bo
th c
onta
in f
loat
ing-
poin
t va
lues
, w
here
as t
he a
rray
will
sto
re m
emor
y ad
dres
ses
and
mus
t th
eref
ore
cont
ain
mem
ory
poin
ter
type
s.
The
entir
e da
ta s
truc
ture
for
the
Ada
line
sim
ulat
or i
sde
pict
ed i
n Fi
gure
2.2
3.
2.5.
2 S
igna
l P
ropa
gatio
n T
hrou
gh t
he A
dalin
e
If s
igna
ls a
re t
o be
pro
paga
ted
thro
ugh
the
Ada
line
suc
cess
full
y, t
wo
acti
viti
esm
ust
occu
r:
We
mus
t ob
tain
the
inp
ut s
igna
l ve
ctor
to
stim
ulat
e th
e A
dali
ne,
and
the
Ada
line
mus
t pe
rfor
m
its
inpu
t-su
mm
atio
n an
d ou
tput
-tra
nsfo
rmat
ion
func
tion
s.
Sinc
e th
e or
igin
of
the
inpu
t si
gnal
vec
tor
is s
omew
hat
appl
icat
ion
spec
ific
, w
e w
ill
pres
ume
that
the
use
r w
ill
prov
ide
the
code
nec
essa
ry t
o ke
epth
e da
ta l
ocat
ed i
n th
e a
rray
in
the
in
pu
ts l
ayer
cur
rent
.W
e sh
all
now
con
cent
rate
on
the
mat
ter
of c
ompu
ting
the
inp
ut s
timul
atio
nva
lue
and
tran
sfor
min
g it
to t
he a
ppro
pria
te o
utpu
t. W
e ca
n ac
com
plis
h th
ista
sk t
hrou
gh t
he a
ppli
cati
on o
f tw
o al
gori
thm
ic f
unct
ions
, w
hich
we
wil
l na
me
and
The
alg
orit
hms
for
thes
e fu
ncti
ons
are
as f
ollo
ws:
weig
hts
Fig
ure
2.2
3 T
he A
da
line s
imu
lato
r d
ata
str
uct
ure
is
sho
wn
.
82
Adal
ine a
nd
Madalin
e
function
(INPUTS
WEIGHTS
return float
var sum
float;
temp : float;
ins :
wts :
i integer;
begin
sum = 0;
ins = INPUTS;
wts = WEIGHTS'
input
connection
for i = 1 to
do
all weights in
temp = ins[i] *
sum = sum + temp;
modulated
end do
end function;
the modulated
function compute_output (INPUT : float;
ACT : NODE TYPE) return float
begin
if (ACT = linear)
then return (INPUT)
else if
(INPUT >= 0.0)
then return
else return (-1.0)
end function;
the Adaline is a linear
just return the
the input is
return a binary
return a binary
2.5.
3 A
dapt
ing
the
Ada
line
Now
tha
t our
sim
ulat
or c
an f
orw
ard
prop
agat
e si
gnal
inf
orm
atio
n, w
e tu
rn o
ur a
t-te
ntio
n to
the
impl
emen
tatio
n of
the
lear
ning
alg
orith
ms.
Her
e ag
ain
we
assu
me
that
the
inp
ut s
igna
l pa
ttern
is
plac
ed i
n th
e ap
prop
riat
e ar
ray
by a
n ap
plic
atio
n-sp
ecif
ic p
roce
ss.
Dur
ing
trai
ning
, ho
wev
er,
we
wil
l ne
ed t
o kn
ow w
hat
the
targ
et o
utpu
t i
s fo
r ev
ery
inpu
t ve
ctor
, so
tha
t w
e ca
n co
mpu
te t
he e
rror
term
for
the
Ada
line
.R
ecal
l th
at,
duri
ng t
rain
ing,
the
alg
orit
hm r
equi
res
that
the
Ada
line
upda
te
its
wei
ghts
af
ter
ever
y fo
rwar
d pr
opag
atio
n fo
r a
new
in
put
patte
rn.
We
mus
t al
so c
onsi
der
that
the
Ada
line
appl
icat
ion
may
nee
d to
ada
pt t
he
2.5
Sim
ulat
ing
the
S3
Ada
line
whi
le i
t is
run
ning
. B
ased
on
thes
e ob
serv
atio
ns,
ther
e is
no
need
to s
tore
or
accu
mul
ate
erro
rs a
cros
s al
l pa
tter
ns w
ithi
n th
e tr
aini
ng a
lgor
ithm
.T
hus,
we
can
desi
gn t
he t
rain
ing
algo
rith
m m
erel
y to
ada
pt t
he w
eigh
ts f
or a
sing
le p
atte
rn.
How
ever
, th
is d
esig
n de
cisi
on p
lace
s on
the
app
lica
tion
pro
-gr
am
the
resp
onsi
bilit
y fo
r de
term
inin
g w
hen
the
Ada
line
ha
s tr
aine
d su
ffi-
cien
tly. Thi
s ap
proa
ch i
s us
uall
y ac
cept
able
bec
ause
of
the
adva
ntag
es i
t of
fers
ove
rth
e im
plem
enta
tion
of
a se
lf-c
onta
ined
tra
inin
g lo
op.
Spec
ific
ally
, it
mea
ns t
hat
we
can
use
the
sam
e tr
aini
ng f
unct
ion
to a
dapt
the
Ada
line
init
iall
y or
whi
leit
is o
n-li
ne.
The
gen
eral
ity
of t
he a
lgor
ithm
is
a pa
rtic
ular
ly u
sefu
l fe
atur
e,in
tha
t th
e ap
plic
atio
n pr
ogra
m m
erel
y ne
eds
to d
etec
t a
cond
itio
n re
quir
ing
adap
tatio
n.
It c
an t
hen
sam
ple
the
inpu
t th
at c
ause
d th
e er
ror
and
gene
rate
the
corr
ect
resp
onse
"on
the
fly
," p
rovi
ded
we
have
som
e w
ay o
f kn
owin
g th
atth
e er
ror
is i
ncre
asin
g an
d ca
n ge
nera
te t
he c
orre
ct d
esir
ed v
alue
s to
acc
om-
mod
ate
retr
aini
ng.
Thes
e va
lues
, in
tur
n, c
an t
hen
be i
nput
to
the
Ada
line
trai
ning
alg
orit
hm,
thus
ada
ptat
ion
at r
un t
ime.
F
inal
ly,
it al
so r
e-du
ces
the
hous
ekee
ping
cho
res
that
mus
t be
per
form
ed b
y th
e si
mul
ator
, si
nce
we
wil
l no
t ne
ed t
o m
aint
ain
a lis
t of
exp
ecte
d ou
tput
s fo
r al
l tr
aini
ng p
at-
tern
s. We
mus
t no
w d
efin
e al
gori
thm
s to
com
pute
the
squ
ared
err
or t
erm
the
appr
oxim
atio
n of
the
gra
dien
t of
the
err
or s
urfa
ce,
and
to u
pdat
e th
e co
n-ne
ctio
n w
eigh
ts t
o th
e A
dalin
e.
We
can
agai
n si
mpl
ify
mat
ters
by
com
bin-
ing
the
com
puta
tion
of t
he e
rror
and
the
upd
ate
of t
he c
onne
ctio
n w
eigh
tsin
to
one
func
tion
, as
th
ere
is
no
need
to
co
mpu
te
the
form
er
wit
hout
perf
orm
ing
the
latt
er.
We
now
pr
esen
t th
e al
gori
thm
s to
ac
com
plis
h th
ese
func
tions
:
function
(A : Adaline; TARGET : float)
return float
var tempi : float;
temp2 : float;
err : float;
term for
begin
tempi =
temp2 =
(tempi,
err
absolute (TARGET -
return
end function;
function
(A : Adaline; ERR : float)
return void
var grad : float;
gradient of the
ins :
to inputs
wts :
to weights
i : integer;
84
Ad
alin
e a
nd
begin
ins =
start of input
= A.
start of
for i = 1 to
do
all connections,
grad = -2 * err *
=
- grad *
end
end
2.5.
4 C
om
ple
ting
the
Ad
alin
e S
imul
ator
The
algo
rith
ms
we
have
jus
t de
fine
d ar
e su
ffic
ient
to
impl
emen
t an
Ada
line
sim
ulat
or i
n bo
th l
earn
ing
and
oper
atio
nal
mod
es.
To o
ffer
a c
lean
int
erfa
ceto
any
ext
erna
l pr
ogra
m t
hat
mus
t ca
ll ou
r si
mul
ator
to
perf
orm
an
Ada
line
func
tion
, w
e ca
n co
mbi
ne t
he m
odul
es w
e ha
ve d
escr
ibed
int
o tw
o hi
gher
-lev
elfu
ncti
ons.
The
se f
unct
ions
will
per
form
the
tw
o ty
pes
of a
ctiv
ities
the
Ada
line
mus
t per
form
: a
nd
function
var tempi : float;
(A : Adaline) return void
begin
tempi =
A.
=
end function;
function adapt_Adaline
return float
var err : float;
(A : Adaline; TARGET : float)
until
begin
input
err =
(A,
(A,
end
2.5.
5
Mad
alin
e S
imu
lato
r Im
ple
men
tatio
n
As
we
have
dis
cuss
ed e
arlie
r, t
he M
adal
ine
netw
ork
is s
impl
y a
colle
ctio
n of
bina
ry A
dali
ne u
nits
, co
nnec
ted
toge
ther
in
a la
yere
d st
ruct
ure.
How
ever
, ev
enth
ough
the
y sh
are
the
sam
e ty
pe o
f pr
oces
sing
uni
t, t
he l
earn
ing
stra
tegi
es
2.5
S
imula
ting
th
e A
dalin
e85
men
ted
for
the
Mad
alin
e ar
e si
gnif
ican
tly
diff
eren
t, a
s de
scri
bed
in S
ectio
n 2.
5.2.
Prov
idin
g th
at a
s a
guid
e, a
long
wit
h th
e di
scus
sion
of t
he d
ata
stru
ctur
es n
eede
d,w
e le
ave
the
algo
rith
m d
evel
opm
ent
for
the
Mad
alin
e ne
twor
k to
you
as
an e
x-er
cise
. In t
his
rega
rd,
you
shou
ld n
ote
that
the
lay
ered
str
uctu
re o
f th
e M
adal
ine
lend
s its
elf d
irec
tly
to o
ur s
imul
ator
dat
a st
ruct
ures
. A
s il
lust
rate
d in
Fig
ure
2.24
,w
e ca
n im
plem
ent
a la
yer
of A
dali
ne u
nits
as
easi
ly a
s w
e cr
eate
d a
sing
leA
dali
ne.
The
maj
or d
iffe
renc
es h
ere
wil
l be
the
len
gth
of t
he a
rray
s in
the
lay
er
reco
rds
(sin
ce t
here
will
be
mor
e th
an o
ne A
dalin
e ou
tput
per
laye
r),
and
the
leng
th a
nd n
umbe
r of
con
nect
ion
arra
ys (
ther
e w
ill b
e on
e w
eig
hts
arra
y fo
r ea
ch A
dalin
e in
the
lay
er,
and
the
arr
ay w
ill b
eex
tend
ed b
y on
e sl
ot f
or e
ach
new
weig
hts
arr
ay).
Sim
ilarl
y, t
here
will
be
mor
e la
yer
reco
rds
as t
he d
epth
of
the
Mad
alin
ein
crea
ses,
an
d, f
or e
ach
laye
r, th
ere
wil
l be
a c
orre
spon
ding
inc
reas
e in
the
num
ber
of w
eig
hts
, an
d a
rray
s.
Bas
ed o
n th
ese
ob-
serv
atio
ns,
one
fact
tha
t be
com
es i
mm
edia
tely
per
cept
ible
is
the
com
bina
tori
algr
owth
of
both
mem
ory
cons
umed
and
com
pute
r ti
me
requ
ired
to
supp
ort
a li
n-ea
r gr
owth
in
netw
ork
size
. T
his
rela
tion
ship
bet
wee
n co
mpu
ter
reso
urce
s an
dm
odel
siz
ing
is t
rue
not
only
for
the
Mad
alin
e, b
ut f
or a
ll A
NS
mod
els
we
wil
lst
udy.
It i
s fo
r the
se r
easo
ns th
at w
e ha
ve s
tress
ed o
ptim
izat
ion
in d
ata
stru
ctur
es.
ou
tpu
ts
Ma
da
lin
e
activation
outs
weights
We
°3
ight p
ou
tpu
ts
we
igh
ts
Fig
ure
2.2
4
Ma
da
line
data
str
uct
ure
s a
re s
ho
wn
.
86
Ada
line
and
Pro
gram
min
g E
xerc
ises
2.1.
Ext
end
the
Ada
line
sim
ulat
or to
inc
lude
the
bia
s un
it,
0, a
s de
scri
bed
in t
hete
xt.
2.2.
Ext
end
the
sim
ulat
or t
o im
plem
ent
a th
ree-
laye
r M
adal
ine
usin
g th
e al
go-
rith
ms
disc
usse
d in
Sec
tion
2.3
.2.
Be
sure
to
use
the
bina
ry A
dali
ne t
ype.
Test
the
oper
atio
n of
you
r si
mul
ator
by
trai
ning
it to
sol
ve th
e X
OR
pro
blem
desc
ribe
d in
the
tex
t.
2.3.
We
have
ind
icat
ed t
hat
the
netw
ork
stab
ility
ter
m,
can
gre
atly
aff
ect
the
abili
ty o
f the
Ada
line
to c
onve
rge
on a
sol
utio
n. U
sing
fou
r di
ffer
ent
valu
esfo
r o
f yo
ur o
wn
choo
sing
, tr
ain
an A
dalin
e to
elim
inat
e no
ise
from
an
inpu
t si
nuso
id r
angi
ng f
rom
0 t
o (
one
way
to
do t
his
is t
o us
e a
scal
edra
ndom
-num
ber
gene
rato
r to
prov
ide
the
nois
e).
Gra
ph th
e cu
rve
of tr
aini
ngite
ratio
ns v
ersu
s
Sugg
este
d R
eadi
ngs
The
auth
orita
tive
text
by
Wid
row
and
Ste
arns
is
the
stan
dard
ref
eren
ce t
o th
em
ater
ial
cont
aine
d in
thi
s ch
apte
r
The
ori
gina
l de
lta-
rule
der
ivat
ion
isco
ntai
ned
in a
196
0 pa
per
by W
idro
w a
nd H
off
[6],
whi
ch i
s al
so r
epri
nted
in
the
colle
ctio
n ed
ited
by A
nder
son
and
Ros
enfe
ld
Bib
liogr
aphy
Jam
es A
. And
erso
n an
d E
dwar
d R
osen
feld
, edi
tors
. F
oun-
datio
ns o
f Res
earc
h. M
IT P
ress
, C
ambr
idge
, M
A,
1988
.
[2]
Dav
id A
ndes
, B
erna
rd W
idro
w,
Mic
hael
and
Eri
c W
an.
Aro
bust
alg
orith
m f
or t
rain
ing
anal
og n
eura
l ne
twor
ks.
In P
roce
edin
gs o
fth
e In
tern
atio
nal
Join
t C
onfe
renc
e on
Neu
ral
Net
wor
ks,
page
s I-
533-
I-53
6, J
anua
ry
1990
.
[3]
Ric
hard
W.
Ham
min
g.
Dig
ital
Filt
ers.
Pr
entic
e-H
all,
Engl
ewoo
d C
liffs
,N
J, 1
983.
[4]
Wilf
red
Kap
lan.
Adv
ance
d C
alcu
lus,
3rd
edi
tion.
Add
ison
-Wes
ley,
Rea
ding
,M
A,
1984
.
[5]
Ala
n V
. O
ppen
heim
ari
d R
onal
d W
. Sc
hafe
r. S
igna
l P
roce
ssin
g.Pr
entic
e-H
all,
Eng
lew
ood
Cli
ffs,
NJ,
19
75.
[6]
Ber
nard
Wid
row
and
Mar
cian
E.
Hof
f. A
dapt
ive
swit
chin
g ci
rcui
ts.
In 7
960
WE
SCO
N C
onve
ntio
n R
ecor
d, N
ew Y
ork,
pag
es 1
960.
IR
E.
[7]
Ber
nard
Wid
row
and
Rod
ney
Win
ter.
Neu
ral
nets
for
ada
ptiv
e fi
lter
ing
and
adap
tive
patte
rn r
ecog
nitio
n. C
ompu
ter,
Mar
ch 1
988.
Bib
liog
rap
hy
[8]
Win
ter
and
Ber
nard
Wid
row
. M
AD
AL
INE
RU
LE
II:
A t
rain
ing
algo
rith
m f
or n
eura
l ne
twor
ks.
In P
roce
edin
gs o
f th
e IE
EE
Sec
ond
In-
tern
atio
nal
Con
fere
nce
on N
etw
orks
, S
an D
iego
, C
A,
July
19
88.
[9]
Ber
nard
Wid
row
and
Sam
uel
D.
Stea
rns.
Ada
ptiv
e Si
gnal
Pro
cess
ing.
Sig
nal
Proc
essi
ng S
erie
s. P
rent
ice-
Hal
l, E
ngle
woo
d C
liff
s, N
J, 1
985.