PAGE No

DATE

ASSTGNYENI-4
'Decoribe Convolutional Neural Netuooy k (CNN)
Acobvoutional Neural Network (CNND Ga
type of deep Leanning modal deiqned for
prDcsl Dq qxd-liko data such as sageTt
mathematcad cperation called conoluhon
to extract Lmpoztart featurca ike edqee,
texture,and patterna from the ngut.CNNs
compostd of several laycra

8ird

Cat

Dog
oput
O
Input Output
lay ers
layeyb
hidden lauyerrdu

fi¡: Convolutional Aeuzal Nehoosk

Convalutional layerik Theie apply hlkr (kernely)
do detc fentua ih anis hmmage

Poaling lascrei To tduca the apatal dinahdin.

 PAGE No.

DATE

0f the data, de creasing computation obile

ttainlh important tofoYmaion
Fully conneckcd layens AEkr feature extrachor
these laytr perfoo dassificahion.
ACNN s0kn ued in iageiccagni
honobicct dekachion,and dea aDalyck de du
to it bility tD automatially Itarn lpatial
hierarchies of featurtl
"2 Expain Deep Convotional Neural Nehoegk
Arcbitecturta
VGG 11s aa decp neural hchsork
knQuon fer te siopicity cnd architeeturt
Aenes ellectivencur It wee
small3x3 conuoluhopal tilkn
notautng depth by tackin layert. VaG -ls.
Land VGG -19 are ts
populasY versio n with ls
and (9 layert
onally Inttnie tspechivelyAlthough
VGG moded computa
transter leaenin due to Laldel
ed are

qeneralzahion apabilities their

2. PlaceNet :
PlacaNet
t cognition
kk
specificall tain ad for
for sCene
Iike Alex NettatkTt K bauecl on axchiechuzt
andVGGbut
clautityinq difkren Aocwe
beaches ferute) ratherQcene
than indiviuo
 PAGE N

cbËect. Placu Net eLtntial tar applicationa

guch ae autonornoe diving andbotie
ohereunaritanding the Cnuitonment a.
Crucfal

S. DenaeNeti
Dense Net I& a nne advancd architecture
that onntct cach laytr to every Lather
layer n a feed- foxnxd fathion Thie
approàchaalitater better featurea
ducee the nunaber parameters and
Impmveg qradient flas diing training -
makingDense Nat mo ctiient and
fox deep learning take ike
mage clastittoahan.
Qlai Explain the Corcept of Batch Normalrzaion
Ratch Normalizahion (&N) a technigue
in dep le dzning thatacceleratee and
stabilize the raining
nehuorkDuring braining.the Inputa o
Leachlauerichangei a theselghti art
updahed , ohich carn show douoh traininq
and lend to nutability.
BN tackles thieproblum hå noxmaizing
cach layer
he Inputa to cach layer so that they
Congistent mean of0 and
Variane of 1. This rtduCRA Tnterral covaziaBe
shift,allolng the modal to traln fasttr
oith hi¡her Jeazning ratea.
Additionally BN ntoduce two learn
 PAGE No.
DATE 7

able cale andchift to

the
parameter
normalizedinpuets, proutdin the
the neh unrt adjut
flexibility
It alo has
to find the optimal epzeaentat
has a requlaizinq eteot
normali zation adds sltaht naise ho thu
nehuork reducing the ikeihoodiof
BN
nehoorki
i parhaularly elfecthivein Vexy
ohereqradient flous
lnight
overdeephtihg
become? an ssue, a n d t Js widey uied
in modarncchite chures like i ResNetand
Inception.a
Qr:lhat is
Vkiopt Hypcrparam tkr Tuningn Compute
Hyperpaamtker tunin h computer vabn
refer to the Process af optimizinq non
learhablepaamettrs ina modelto nbanc
performancea
In the context of Convolutional Neuxal
Nehoork lcNNa)keyhypetpa ramckns îndde
learning xate,batchsizenumber of filteri
filter ize_, ands the nunbex oflayars Thec
have a siqnificant mpact on h
raoculs pertormana-,4encralizahion abilityan
com putahonalfficitncy
Hyperparam cter huning
ftnd thL best combihation that
iecintial n
mocal tn perfornm ptinally Oneeh ot allouo the

cohileavoiding
Techniauea offiting uhderfiting
and searchandom searb
Bay eaian opihni 2ation lars
commony
 PAGE No

DATE

ed for hypcrparamtr aptimizalionl, Grid

tarch exkauathely tie' all comhinahansof
Lhyperparmetra ohertag sandomlgearch.
Tandomy sebect pararmeten conmbinabth.
Baycsianeptimization k mort aduanced
method that modelsth performance ot
hyprparameee aentigurahbn to tati In
computor vialon, efleche hyper paxamtkr
lquxahbn and ntelligentychatca the.
hext combination to teet. In compLtkr viioD,
effective hyperparamekr tuninq can lead to
Subetanhial lmpr yemcnts Ib nage dascificahin
Lobiect detect'onand segmentation tauki

in Detail : VaG
VaGs a decp convotionanural
hetuaoxkarcbitechue develapedi by tha Vieual
Geonety roup at Dxfacd,and itte oidely
Khoon for ite Simplicity and effectivencat io
fmage ecogniion tauk. VaGk key nnovation
is the ue olsmalla3x3_convautional fllers,
ctackdn laytri to inucale' the dapth of
to Capkure
hetuork This fallour thenctusork tp
mereucopler and abrtrati patter n a th
data moYe4 depera through th layers
deperthraugh
WaG.-6_andVGG- 19 the nat
popular vertlort containing &.and alayers
TApectvely The arhitcchuecontlste of
mulbipla convolti onal lay ert folloed a by
max-poolnglayerswhlchdcky thu spatial
dimenaon Aftr fcatue exra chion, the data
 PAOIE NG
DATE

paartd throuah fully Conntcted layers o4

dascificahion. Despik Tt
VGG ts computationally mprtuive pcrtoinanr
largeamount otmemory Cxpcnalvecqiing
and poceing
DDoerdue to tdep gxchitccturtand
pwer
nunke
ot paramthox
HoatverVGr remand number
apopular
chotce tranafcrIearnlng becauac the
featur
estroctoCan_qencralizewcll
differentdataAct and taske a
Q's Alex Net Application
Alexnet k one of the moctnf luehtial
dep eaxning modal.
Cvolutontzed introductn2012.,
computr vision by
hat
the Image Net
LoargeScak Loibnin
Vicual Recognition.
ChallengeIts
clasication,where primayapplicatiaonI
it
thousande iof catcgarieaclassifieg
with
In tmae
fmaqesl into
AlexNet 4es high aceucacy.
utomathically
from maqcsch
denp'
extract conuolutional layexs to
hierarchiccl feahurca
IByond i ased, textur and ohiect
uied in fonagaclacittcation
abjeet ,AxNet ha beun
ndlo calizih detcchibn ,helpingtaimaqedentifying
objecta oitkin
bas aao Tt
keeohere
Hh and seenapplicatons
t aic in defeing in medical inaaging
turmorsabnocmalt
otherdireada
scanAlexNet has iketrom
contabutedsiqnificaly
MRI and cE
the deNelipmcntot autohafnos
ohiving systima by
enablinqvehiclez.
recognize obstaces,pedattany andadto
 hlet Res
VGG
ndmodelaike
paving learning-
edknced
deep
y the
or
hepotential
of dimohithatdt domainí
eAe Sh purpoSes fiCation Ver
and
secunty foItrc Qysterne tcognition
Fuxthermat
face Lutilhed
Io ls
DATE
No. PAGE