Thanks to visit codestin.com
Credit goes to www.scribd.com

0% found this document useful (0 votes)
8 views5 pages

Deep Learning - CNN

The document discusses the architecture and functionality of Convolutional Neural Networks (CNNs), highlighting their ability to capture spatial relationships in data. It covers concepts such as convolution, pooling, and the significance of kernel weights in feature extraction. Additionally, it addresses challenges like translation invariance and the impact of padding and dilation in CNNs.

Uploaded by

Archit Tiwari
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
8 views5 pages

Deep Learning - CNN

The document discusses the architecture and functionality of Convolutional Neural Networks (CNNs), highlighting their ability to capture spatial relationships in data. It covers concepts such as convolution, pooling, and the significance of kernel weights in feature extraction. Additionally, it addresses challenges like translation invariance and the impact of padding and dilation in CNNs.

Uploaded by

Archit Tiwari
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Aprly seala cud shile.

katekhas, And sring zsting we


enpeical
de hochvalyation Sn oBtained by
smemcnts chich ae 1 ae0.98
t144 -1.06
-CWeaginq out he boitch and o
B.44 6. 08
* CNN hext clas...
9) Preblen on batch omalyatien 2I Narah, 90241
oth 2 featues. sanplos
- Considera hin batch ef 3 Convolutional Neunal Netaank
Rt, 2 featnes
’God at oavuivg spalal xalation.
2 6
-Visual porceptien.
- Covelithons
-5
- Poalig avaton
F2 3], p-1 2],6 - 10 - Convolulional netaJark
th horanalgafssped output ndor the hod.popling
Compat hasmaligatie
.afler bath Gompuinq His lany anad
4 advcn
batoh
botteh
Visual forcepbin
- 1-63
Signal, ohile thaoing imagos
’ Hebel and Wiesel Cot Eypeumevct
’ Theelecheodes wete
edge was 4movin,
slide'sodae
19%2..
aclvatng th
’ Thinner linets Wee oulh
jnital Jnewens cnd lvgelot neuens,
sutne wete att vaing thu
cenbinethese
+B otthe lator layers we
+8

Tnduive biases
Brngun Aasumptiens of tha Oate, do the
’dnuwial netwak,
-1-22 -1.02 - kscalily tzanalatien
- Inuauance to
|122 1s6| - Hiezanahgicol ompoiionlt
houndles
’ Localy w spalallallonship
’ Hiecv the same ineov onfoumalhon
but
oyed
-Asmall BindonosQata/suctau.
thnough outthe Maßix,
- Now the tondous js the weight
data
WILL.
boindsois calloo
- n CNN, the
’leCunn 1993,built a tonvalalionNN konel
And got geod reualls, W- 4941

CGnvolalion. spatial ceW lanl)=lnle)


’ CNN s MsSo cohouthae is
relalionahip, eg images.
’ Con albo ke uued faaudis signal. Maing keualtrough tha baa
t>
Inage aio spatialrelkionship.
’Each pixel ’Also the weights aruyeßuced
heovilyalmost g5. reclucbhon,
botwec
imaq hnoef 8ata.
0-255
fe 8-hits.
fax
each sucos)Ve newern, we
pou pitel, ’ As fe dmovirgthz bame
caloukaithznkyphavieus layen,
’Conkinationof RyGi,Bzhannel gives kenal thUctu
the same kenel.

loyor to.clases

’ simple hewnal netook con'tAccunt


jut keaiusgone keunel odl
no bramslation fnyaManDe, >But,
Je3tinbias Hence we ta,to
leorn uliple keunels fon the
3 3 3ane hidelen ayor.

We can brinn tiamslation invaiance


Uing convoltions.
deh cuen.
1D tonvelution ’ Fon a, singa kenel, we Ipage 34|
’far to tasar ,given ingat vectan
wcights:
u is a veclo Size of wW- kot1. We babe nulliple kn thannels un
tta bidden layos ie we learu multipla
keunols.
Let a hicolon laor has C'hanrels
lechuieall,.nsts Sross -CONeladien . we naad to lacou
in CV, Bt (orhxC11)c wight.
lUsually, ahannels
’ Jmentionesl, ane had explicity
but it is oloays thoe in
+Gnvalutien c0n implment difeantiat the convalfien
konsls is nor. o feätwe haps
’No. oe
4233 34
we ae Jeauning.
keuestoes naspbinfeld
in convetion are the Same.

Anal outgut oosize


to betact CH-st1)x(W-0+1)
’ Corvaluthon tan be sed change
eges s eoges -bave C-1
in antansig.
hD cenvoltien
teR
CXHXW Theene ralabion bho chanaels,
isi ta t a babe
ueR inguffentuie ’ lohon thasbide
sides aeresthe
inpuft featuehap. thz output.
hxeis tha Size of thz W
ateespbive ficled a keruel.
Jbotoisefrst, HO
’ The keunel odll move
than invementcolunn Dunelcgon ineuose sbucle, itl taoucng
’ We oukput size n6 yucky.
paducthe-featuie
loction, tha elament-oisecind
’ At eack the to prLsWe he foaties, We kale
kctseu kenel.cnd input, ’ So
stude>1, Key 2, We noesse
ho. of
Ahisklon
channels n thesubseopent
layes.
pavametov. pages5
DL
vetsIOn of the feotie.
Zetoed rame acoeo COonoth
input bself ohich covos vaieusScenaios
-Podding is added-to input to mako of scale cnoJolalion.
oukput'iforont size 'acarong
Lke instead 24 Masch 2025]
seelced we wout same siZe
oheh ’ fgjoct Ealuotien 18Apsil 2025
- The soudk spcfies a sop size
moaulates th Convdlations,as matuix
- The ial Oilalion
epansionof tha filter o aöling Awalliplcolien

WxH
input hx9

Yabax.

’Convolution 8Apcialcose o MLP


Paddng Deheeweigyhte ae sianed.

Dialation ds wsed ehen we oalo


’ lsth amall filiee,
6ome applcalion, ohen
in
Ttis sed ralalion jsapased,
tha spakiol spakalralation os SX3
’ Dialation loses considaed. A4
Lomcpivelsae
inol Eg. XeR; ue R
4581| L41]
Euisvariance XU 1888|143
3664 38
331J
l6578
4101430 33 1 0 00 00
1
Eqyivauan is Noelal tohen we knas
- Some lacal unstion is ueul eveuyuhea
Inaiancaín Poling
Afuuctin fis invasjat to

invziana
’ Poaling ayes pyovitinade
tocuypomlothion Onetall.
Itxouts in (hseude)-inveiance
Poollohsningtinputvaluneislovse,poolnq -Jn nest of thz cases ydecal
tanslatien oll net clange the
poclng tusult.
cane mareabout
Jonscalng epoaben. ’ This b helplul i we
but
’ Infemobion los till hagpsn,
we
nen-ovetogpingaspresewe
’Wedt take inputsize bnd
to eouca
antglobal
Consider, poelng oan size hx the stuctie .
Cnd input xeR
Convolukienal Neunal Netuank(CNN)
a torton OeR hetaak is genwally
Polig prudueas ’Aconvclufienal
uconyesition ofcónvshtierol
bolinedas layn( PooL),
poling
Jayor(coN), (Rel ) a oiny hon-linew
m1ax Xoim,sism linen ectiien
funclions, no Ballydomast>
nch,kto oetivakion
Sagrs(F). end os alere
FCs Ane nedad atthenwe tsll latten
Oc,i,i22jtn,st+m PooL koyes
hu h=0 m=0 CoNVmd
no.

channelsis rayed afto.


No.of
tz poling
Tiainng bimetperds en boonany
’ lanoble ponemelers Atethoe size,.>EC RELUK-’ Fc
shelol dividte
sizeaoes ih inpet
the
so-that it snet overlop.

You might also like