sígmold u the actatn funilian

, he
and
whue, x >0.5 1

- hy and h with AND , NOT, OR gä

h oait(-z,)) c(it)1,+)h
Yhith) o hith,4) 4,as

Cujsta fr makíng Bock Papag ation gor thm

Pefomone bete. :
Sequeneial Batch Updote
the saquantiad wode ot bak ropagalig learnig
tonpilston ally fustes than bath mocle
when tranv lage
MazimiLe intormaluon k coilent
Taining enamplechoosenpruentid t bouiparag
on the oauu that
Should
asgt paible tor
he tayk that seults
example
in lorqe hinung emY ard aadially dieent
ahd
$on retouy ed.
3 Activaion duntion! A muti layaa povefug
trained wt th
learn fala douit int
uokun
ohn siod ot'v tibn
model of he nelissk is aritiygmntic.
 dlv) a tan h (vb) non- lineor funtog oe
orm of hypu bolie orgent.
oh a l . 415
Conlant
conslarft
and b= a) 3
boli'? tangent funtlion dafehed
’7he hyr
he goltoutng
)o= -1 and )= -1
) At the origin the qain ot activation tunati
unity
=l.1424
)fhe Rcond derftie din attained at
V

Taiqet Vaues is anpoitant that tärgt

waluu wtll be choosen within the ange 4
veue
tiqned tuntin depnding
buteathe othvwie
bask algotthm hendu -to dsiye
hee oramlou o of the no to inginity
Linilng value
Vabes of -atE
 5. Aomalizing the inuil Fach inpt wate
Shauld be prproeued mean Vadee
on erttie cluse a
he ingut variable
Shoutd be n corelated
Coritin
thiu Can
he
be
tuinng
be done by
set

prinuul tonponent onalyu ( pca)

Mean ) decomelton
>gemua

) (ovainte
equali zátion

A. Tnitialiaion
6.
cnigned
the hAon
ita) values
to Saturatio
Consider a muti tayer peneption uing hypeutolic
tangnt funioo for the octivatio n untion

that p aid to each

ket it be csumed
heon vae
 ket it be puukon aMumed that inpau

oEji) =0, (i, i) pain

11

E
M

E[wjiJ- M wt
tov
whe M= Synanlic Corectos nesy

Why Bue pyag ation algo

it
When the ie fed to Ihe neuml nlo
fed
yany tonugb hidden laye ond eaches
Thiu

to

and com pataton ha t peat. hi

ndnar
cfr initod dalalu nd ay
JA

computatiso and eadjuwtrents oiga

tor be done b BA.
 It is a teur ctep octiity

backpopagalun wit fyradint dunt

3)
au iniLlied anc
Dyorard Pa set
caníed and moclal
pas
ito initial ppediction
) Cor functisn tho
how
trown ue value
with gndiut dewti
values e
Caleulil how mueh ue oft
each wciglt of mocl, o do
effeted by
his dlslance qoing
ft calatate s pveddicton
back fom enor funtlion o specifie nunn
weigku hat minimize eor tuntiy
Can belse udaled
) wegnt updale i
in Je Training ct tuken.
Sample
atar eacb
Thi ii not practial tyMcally boch of iqiu
a9gregat aputt bath cie and iterau
imyortant Aunnbag entir
ilir taini Set

t called epoch
 8f
Omied

ip ú medelled wing
andony seleted.
neuen
caluliled
t hiddes auu
Calulte the emoY in

feom ofp layaSuehto Tthat

widden lae to
inoy u denused .
ntil deired ofp
feiat Thi proon

sxuingbaed on human nevos

There retsovty ar

he rause
tNe Can enor bac ppqegaton f6to

) Aimard roqsgaton alyo

2) Barkond nfugalan algo sack
The Conentalos
 and 'at eaeh node erroys
andlyiod
the miimum qyadisn chooen
local niina and howe set of
and he rmin imum
mi mas
tocal
choosen tinally Mod
olich is caled
neght i
míc ma

Netok.
Back frorsghon
Rate
leaninq
the douk emor comection leorrng ago.
Joued on orert of corredion algorihy
es a eror
Consuti ct stes
in the Same bak
ntiook is oyofte diucton
that au) entae neual

) fowayd propagluon algo

9) Buckuand repoglon algo

’
tlantshing cpladling grodin,
Compaationad Comg lexi ty
’ tyreu parameley tnning
 ard ordeed derivatine
Hessfun Mai a scala
-) altehati BPA
C)
Cost tunoo Eay Co
Ih hesian Matix oe he eond desivatiie

Say Cw) uart vector 'wand is gren

H= oy Cw Heui an

point need to e conuiderod

imp otant
1. he yalues of
ot tesian mtix hoe und
Back
infuunea
leusnung
A. The inveue of Hei an mau ie vids en h

-laysa puaple.
3" Henion Ndie iu basi to tomulate The
ordey opimatos matod as an alteatie

qenali2atios:
bonoued
it y asumed Dat Tex t data u diLon
data

folouing
 Nonlinear
raping Jerealialoo not
Ta is
data
’da
oconsistent

Trut
(0 nonineav data iu
marming inconsitent

Generalialuog
Inut
() poorly fted dala
qnu alixd l prduea
een cken inut
kom tramyles. in
reual ofw endup
Tf a and cannot clasity the dila
many eramles
conetly
is called cu ovaftiny
majs faleu:
how xqoesentate
it y in envínmne nt of eyest

3, ben in hand
Cros- val datoo :
cDH- Vai daln st The a ilalsta
bandsnly fatitmed
dii ded into
dalaet,
|. Sstimluog sulyct- nd o select
dsjont
nodal
2. \Valid sulet- sec Nalid at
modlel .
Baysian Duikón
- o a wo las
) done C and (2
a) tott a coreetF clayiicatioo is o
3) (he cost
and

eiketihood test.
found y
C

iheli hood yatio

-

thyeshold Tat degined by

A() 2

|x-Mall)
standavd duatoo
moan
 decinan boundany
Equialoty

get

he

2
ond z

cile with cenlea d and 1aduiy r

circle fs
let n de<in s egün inside the
porti utaw pavmatau the capaim ent te cenlo
at e = (als| ond e 4
tet c den oles et ot toreet lani aitn outor
lated clunifi catti n outomy

ohee
Plelc)
and inilar ly flel<) = canditio d prbaiy
ot
 f, and prbabillbe of Ciand ,
ond

p(ele:) probabil:ti
oith P 12 he

pralilitu of comech
e-|-Pe Pe= prob of cyect
claiaia
ob of incomeet
Pe =
clanificatiN
15

Radial cfunction detioale (&BFN)

Ilplayer tiddenlayer tHddenlayer

ole layor

’RBEN netwoke invlve 3 layes with etiety

node
ditfut 4 of Suurce
| (Senusy nií) that Connect nleo to envimnment.
hdden laya ofiu mor-lineor tianfonaton
fom ilp snace to Aidden Space. hddn
ç of highr dinasianaity and he op
linear.
Ure Sirnilar to ENN modes theugb
the implemen ta tiwn ot both modlels u
 4 RBFN nlw pluces one K more RBE
Neurors sace desribed
n he sace dicribed by medude
Vavable o that Cal cutite eulidran
oint to he cenle
Aistance kom evauoted
o each neuon'
kernal funtun widh
krown as
Aata
also eaencd
ditone tnula
= KBE(distance)
RBAN Architec lue
7he tyncal avchitecuye ot RRE Contist of p
layer, felden kyer and suramution layer,
I t consisti ott one neuvon for
Trut layer ?
purtiuloy vavia ble , the input neuvon
eseuy neuon to he hicddlen
Pun the value of Pach
neuronS
tayer
values. and catgoris.
Con1ist of avidble
Hidden ayer
each neuon
mo-of neuons no -of
RBE cenlered on
poant. he
dimennons count ides with o.o padita
Vaviables.
hen an X Nelor t ile valtes u ted
calalales
Kom the ilp yer, hidde nuron
eucideen dutance blw the Testcase and
pout.Cuead then
sfundibo in vaues Reutting
 value Q¢ts kd oto

the vatue oblauhed

Qutput layer i-
hidden lauyen s mutipltd with weiqh t
elited t6 ond pcd to
Sumnal
layer , here
kre weight ualues
Sum reenled Cu meural nek
Outhut
RBFN XOR froblem -

he RGF Neoak to de întrotiqated consist ot

ot gausion fur ctis detred

uhen tenters t and te avee

te (0,o). foy he chanzlistog
outut we 8ume the Fltownq.
’ Output Wyes
by virtue of sypminety
’ the outut unít incucles bias b Cdata
indefendent vavabtes) .
-’ the desive ole vaues o XOR has ron- 2e
Aean.
RBrN the
XOR robln.
> the ilp .ole ot the
2
nlw
 To ft haining dota
whet J=l,2,3, 4
awhese and di u the

coropordig b2,3,4
she

Data pöurt j dË

(o,)
C,o)
4 (a,o)

ulhe G= 0.)353
0.3648 1

6 " I353

6. 3648 4

d (o o1)
w

t ’ additie invese

(64) with
Note hatt 6G s a square matix
Substi ute
uniqu invene
in
eq oblaun,
1-3192 -I.1509 0.64 24 --2501
6.62+ -12509 -1.25096)
-0.9202 |.4 20 2 -0.4202 1.4Q02
 Subst tutirg eqn's and
get w=-.50|8
-2.50| 8
t28409
which Comple ta the specification ct ReEN

Lías
trni ng daã
batiyatio tantiorn:
phe tosd
Comparsiorn of RBEN wtth FFNNN: Weun
’ RBEN hu locaU'ged FCNN ha
qauion unluoo oher s
basu fuiluog (gmod .
Can be solved iung inear Reqpes in
he sprad 's' and basis urtog 'n'
fzed. while FfNN hcoives non-lineay
optimad aly- o
mininine he emr.

cheaps to Solue hoe

’ Sihe RBE on
embedd cvos vali'dation in
the oula loo

->for unifosn data eti R8E may e betlen

Sine Chots val°datino khouldoite a
havealo
more alcurale gponcebutace
aulomaie deas
 Con' dey he xoR boolean funition hat has y
paramias in he 20 inut
Snace whure R-t.414 0onict a RB FN
the gien nlo
60 ’0

AA)
to, -0.
Hacn)

tt ()

Col -0

whee
-(-43 cohere C (o,)
e-a-
e
2
where
whe

Luctedian dilonee
fo ife aTlun
- (4-x)' +(ts-)
e (o,0)
dutane
(0-0 + (o-o =o
H 2)= -(a-o)
dist ot
po)'+ (0-1)
 H,(a) = e
ferm
(o-1)' + (o -o)
= 0.6

ty() -
Similanly O.4 and
2

-0.04

= (o,) o 3(,0)
H()= 0-6
H ) |.0 H0) = o4
3() = 0.4
Hy ( ) = 0.6 Hy( ) = g.6

for Cy (9))

Hl) o.6
Halx) = 6.6

tw;tf (*) = -0.04

 l,() H;(*) x)
(o,) 0.6 0. 4
(0,1) 0.6 |.0 0.3

(I9) .0.6 0.3 1

0.6 O.6 |.0 -6,04

hove acheied Vhe Oatulitsy