ASSTGNNENT-3

QExplan Batth
qchcralizaton
noxmatiz chianctvfilbihq nd
Let wndustand Berm Normal/zntion before
maving tobatch norhm allzahonA dato
prep eAcinqfohoiqLe called hehnàliaahon
dowrnunerital data
without dhanqiog he ctucture.
qcncral tend to madfy the numkea
to abalarncad)
Balanced toLScaleshen
the data ko a machlncaxoinq ar dep
learhln9 alqottbm,we Dormalize ih paxt
make that Qunmadel qneralizec
ottly

fig Bat dh Nocmalizahion

and Genualinatiorn
Overfilling
frior to furthran gencralealon
going
cthtal te cemprchend
Oveahtho
learning., chls
only while Icorning. under
 PAGE NO
DATE

averfiting
One
a conccYh

for 1 machine leatning madel

to learh and comprehend data s through
supervisedleazning,uhupt xvised and zeinfur cement
leanibq are tuo typea o learning hut thoze
Lare toplce for anoherda and another
blog post

trainin
earning
data
The
uhenwslng
nomt training data thi
upervin td
model
has acclas to the better It qeti at
prnducinq
to da s fom thi
prtdichion Sinc the modelleazn
tralpng data utth
raininq data h prdetrmined

Undertiting Balan ced

Overitting
Oueritling. tduc
Statyi
4 hecan the complaxity
(that vtducing the nuoher oh hiddon
laytre).
lecáuld augmnt the data.
 PAGE NO.

DATE

Ae:xplain tht bretlotd Boltznmarn machins wt

onallprofecte
The Reabicktd boltzmann machinedixdaped
hìnton in 1925s a netuot
Symmetically connecte wnit that funchons
like heumn and make ato chashic judgementa
Tt I a type oof unALpervited learning
helptul Aor callaboxatuc Alkia
Peatura Learningdimenaionally duchon
qrtssion, dassittcahon and Peatuzt eazning

Viçible unit
Hrdden unit
Resbicted boltzmann machinu sto chahic.
two Laåertd ntuTal nehusark that can
automatically td enhity wnderlying paltrn n
data by tconttuctinginput They ase' a
Subret herqs_-batrd haodali. Thay have tua
layers ohich s hidin
The hiddin layer i madup of_podee
that create the vistble tater and collect
featut iotermation fom the dataTh: ouLtp
 PAGE No

DATE

nade,ihtch might apptar sid,and they

dont have the yplcal binany aukput that
allowr for thL learnln of paterns
They Vany
place "withaut
because tedrning
thu capadty ale don't cancarned
cannot take

about hidden node; we only cane about

inpt ncdes . R&MS automahicalli Capture all
the pattetn,parametr and elahionship.
Among th data onca the
th nput
nput rqivth
Adintageyi
( Training c chalenginq because it ie
chal nging to calaulate the enerqy gradient
Aunchion
) The hack pzopog ahion algsribm j mtre
well knoon than th co-k algonthm which
Ie utilised InRBM<

Qe DifEcr encd betueth AI, decp learning and

machine earnihq
Artificial Intelliqenca Machin learninDeep learning.
1) Making the conput Acompudtr ayton The exhruchon of
yrtem o think that uer algonthn mort sophasticaltd
Land tak decioionand stahiehcal feature from
hke human beingmode to anal ze data is accom
to Derfarm data pattrr ardl plishd by wing
Cpedfied or reaùtd draw Intrcnct rtficial neua
task hom them.. nehwtyk that
 PAGE No.

DATE

leagnlo4 with conaishin af

pastdata uldipla ayout
ofprcLLing
2)ith ML and The subsct of The subset of
DL a tc A7 that MI ML
Conshihuent partr,
AI the
langtz family
tupe oh A machinu learnig DL ie aimachind
cammpuker algcritha algonthm . ML, uea leaoin al¡onith
thot cued divisinn data to tach a that analyze
mmking o demons icyetern data cinq
trate ntellgna thing (multiLitr

()Crating ntelliger Algonthm enablng Aeuzal netlaork

mochinemlrnicing compattr to leazn toith multipl
humaninkelligere fom data layera and o
hierar chlca
praentation
leantng

Qa:Explain outo encocare with trms consthuche

cdmplinq,deip
dlveqtnegibbusdampling ,deip belief Nehuak
AnabD encoder fealtype of
Larthcial heual networ k ed to leazn
eticient Yeprentatibk of data typicaly tor
 PAGE Na.

DATE

dimendichality zeduchon featurt leazzning

maih campontnt
Encoderi
Tranform
comptsed
the input
stprcsentation.
data into

Decoderi
Reconttud the riginal Input data som
he comprtucd prtserbation
Autocncodex are traéned an conupervkcd
mahhlr to minimize the econitruting
PrKor betwcen and Inputcconutruced
outpit eaninq
cfitctvely lansning a looer-dimenaib
hal atur rtprtatntation.
Conatructive Diverqenca Cco):
Constructive Diycrgeaca (ca) iean alaont
uicd to train enerqs -baued modl uscd
to ikL Reatncta Boltzman machineRRMa)l.

The cD alaszthon oczke 2s follot?

41. Chate ith an iniial nput vector
2. Perfom a qibha Samplinq stp to
qenerat sample(conutuchon) trom
the nput
g. Compare the ctatchic& of theoiginal
data and the generatesd sample
Llae the diftrent io ctatistc to updak the
model
 DATE

Gibba ianping mark 0V chain mo

Casle McMc) alaoithm uscad to
(ompledhi¡h -dimenaional prohbilihy ditabuk

Deep Belief Netuwcyk DB N)

The De cp Belief Nehwork BA) J a
qnerattd model compasted typeat
acoeratike multipe aytr ot otochahic, lalnt
vartablu, ollen constuchd uin muliple
reticttd boltzm ann mahine (RDM)
Training Bacac'
DBNs traintd aytx hy- laytr a
qtecy wnsupenis edfasbion ain R&M.
tratned to
to mode the data dishibuton
in it layer , abich helps to prt- trah th
netuk in a bierarchfcal mann er

Ql s Explain dfarenhal achvation ubchan in

tabulor foxm.
Activationl Deacription LEajuation.
Eunchnn

The met tundarnental

acivation tunction t the
lincar one becade t doe
Dot change tha neupo
ouctput ata
 PAGE No

DATE

(2)Step
205
hatshald tu for value that a Øl0:1 u,oherwye
aboveepeolfied hrtshall
(3) Slamotdall IIt k aa popular optiorn fu
loglctica feed- foruard NN's that )E | 4 e
must output only peibve
nambera.
(aHypabole T s alao a popular
Tangent achvatiorn functon fax NN&d) tan hla)
whtchmust Output valuea
Tn the Yange between
-1and1

6) Relected I I K a near, non-satu-

- d)max(o, z)
Iineax gahn funchanRulu dou
unit not gdturate

Q'e what I h diflereoc behuen

based CANN R-CNN) and a fully cohuallhonal
nethwock (eN)?
AapectR-cNN kegion -Based
cNN) CN(lly.
convo luthoete
Aunpouc Cbjedt dekechan and Schematic
localizaton seg mntahon.
Architecture Combina selectiesearch Convert sta
standad
oith CNN for tgion CNN oto a filly
propouale Lconvolutiopal
torm
 PAGE No.

DATE

3Tnput Extrad rtgion pmpaal ProcAiea the

prDatinq andpastes thenn entie Tmage a
through a CaN nput in an
ndiutually