ANN Unit

Convolutiun Neura Nctwk

* Building blocks of cNNI

Input Iayer
blochs ct CN N Con volutien layey
Activatiun functien l4yr
Puol leyed
A Convolutico nutc ntwek has thc
ayes that
Agerithons.
Useful fos vius dep lrening folleing
2 Let uS Scc the weking af
these layes taking
an etumple of the imaqe havin4
dimnsien of 12 212x4.
Conncctcd
Irput Cenvolticn

Fcatuc extrecticn classificates

al Input layeri This layes will of

width I2, heiqht 12 and acccpt the image
dept 4.
47 convdlutien lcyes :IB CermRutes the vome ef the
inage by gctting the det prodoct bctwcen the
image filtees possibe and the inage patch.
pssiblt
then the volurne wil) bc
Cemputed as
Activatin function layer:
3 This lays applits alivation functicn to cach
clcmcnt in th output of the convolutionad lyer.
i Seme of th wcll aclepted activcticn functicns
aNc RcLu ,5igmoid, Tanh, Leaky. RelU ete,
i Thcs, fnctins will not changc the vlume
obtaincd at the
Convalyticnal laye and hcnce
H will equal to 122|O.

This functiern .mainly geduces the volume of

the interncdicte outputs, which enab)csfas t
Compvtetion of the netwozh medel cnd yeventes
it fom
CUeYfit ting.
* Convoluticn operatiun
Padding :
paddinq is the potess cf cddin
mere picls of 2eA0S call acUnd the bonderi
of an mAqc, in Cydey to inc7CGSe Hs cffective s:
2] 240 paddinq helps to makc Qutput outputdimcnsico%
cund kesnel Si2e indepcndent.
3 Con vol vtin operation cduces thc Si2c of the
Cqi" aye in comparison with the Si2e ct th
layea.
4 Ths type of
edtion in Si2c t cnds to
lese Seme infoticn along the
this ccan be
boderS of theins
Sclved by padding.
Valid convolvtin
Paddiny state gig Scune Convolbtin
full Convolution
a valid con volution: i Extcme CqSc in which
no 2eO- padding is uscdand conyoluticn Ksncl
is onls allowed.
a keynd of Si2c Kincany dimsiunthe
input shupe of m in the diection will bcCernc
m-ktL jn the output.

b] Scume con volution:

i Just encwgh
the size f the
2e70-pudding is added to kee
output Cqual to the sie ef the inpst
i3 rernd sire is K, the
input is padded by h
2405 in that dimnsien.

CFull Cun volutioni i] Entemc case where

enog
205 ac added to7 cVery piacl to be visited
K times , output imag- cf wjdth mtk-1
 SottMa RegrcsSicn
1 The Seftman functn
is nothing but a
acneralizatie ct the Sigreid
1 Thc Softman Tceresoisn
catcssicn that nemali an input valve in t
Uccte. cf valves that folics
whese
totedSums up to i prcbabilty dsote
3The output values aTe
Co,i between thesanye
we have to dced with
moc
than2 possiblelevels we Useclassificaticn
a
with
oi the
lcqistic Tceyessicn, calltd as Seftmeenesli2eticn
1 The functien
i5 USUelly uscd to Ygvessin
that .cen be cempuie lcS$es
erprcted awho trcinine adate Set.
The ckasitier
-\eued Functiun /nvelves Some hieh
notuticn.which
K

wherc Qw,hot
2 The w is the weight vccten,
Xis
thefcate Vcctey l tGining Smple
8 The term
is the bias thet nceds
te be added
cUey teatian.
kDep lremning Fomewevhs
JThc tem dcep usually cfegs te the numbe, t
hidden layers in the neuved nctcnk,
23 Deep leaming isa Subsct of mgchine leamiag
Cohjch. is preditaltd on theidea of loning
fsem cumple.
3] In machine lecrning instead of tcathing 4Gmguty
a Ma5sivC list of yules k i5 qiven to Solve the pcblc
43 A small sct of ins tvuctians corie qivn to
modify thc model 'whenit maks amisteke.
5] Deep leamninq climina tcs Some of data
pic-psoltssing thst is bypically involuc d with mL
Rclatiun betwcen AL,mL and Deeplceaininq

Dule
Sicre

of photos of diffesent pets and wc wentcd to

catcyoie by 'cat' and'dc'
s] Decp certning algorithm Ctun deteminc which
fca tunes ec mest impcxtant to
distinguis hcach
9 In deep Icaning a Cemputer medel
lcas to
pesterm classificatien s ks directly fromm image
tcxt Sound.
lo] Deep Leoning framewer ks_offee buil dinq
blcco
fos designing training and vali
ncuray ntwgks, dating decp
netwgks, thicugh high- le vd progtamming
They Ssuppents compote u'sion appieaticns
jnagc. Spceh an d naturcl languaqc Prclssin3
Tenser Flaw:
Tensen Flow i5 an cnd to end SuTcc
cpen
platfam fey deep
devcleped
lcortning
oy the reugle brein tom
ond cleascd Novembey 2o1S
Rthen oY Jeve, Scipt .to pyevide a
Convenint fTont- en API fery ouildinq applicatin.
4 Hes two majn actiens
a] Building the computatiencd qraph in the
Constet ion pheasc
b] Runining the Cemputehionod qraph in thc
cxccutien ph case.
S] Tensey ccun he cpTeSeotd as_ emultidimsined

ae Usually stuctued into

Constiuction phsSc and on cxceutien phase,
b] The
constructienphesc cSscmbleS a qraph that
has nedes cnd cdges
c Thc
erccution phase us<S a Sesson to CxoCute
opesuhns in the graph.
 is witttn in ythen cand-Syppnts
nultiplc b¡ck cnd neurd nctek ccmputticn
tshell
3 In c nut
Cnd has
Shell kerGS is.ligthweight essy- te-use cngines.
c minimalist approuch:
3] The pimcy Usage is in classificetir
qeneTaion, cand Summaiatien, Gte
translatiin alenq with spcech CCegnition
is a Complctly Pythcn bascd feapseo
which makes femewgL
it cusyto debug and plc.
Sem Smoothly on both
Cpuand apU
prouides Seven different detsets
which can bc loaded in csing Keras ditcetly
Kergs USeS Tenserflcw fer i ts backend,
is SceumlessIy inteqxates with
tunse flaw wey Klew.
 atz

,Bjes and Vaianie with misrmctched ata distributieni

Iyer
ne

orv

importent phencmenon Ccallcd the bias voyjnc dilun

1In Supsvised Icerniny .the class valuc assigncd A

by the lkarnine medel built b45cd on thc bairiny

duta mey diffe fom the actuad class value
1 This cO in lcersning Ccun be of two types
due tobias' and eTe duc to vcsiccc'
Lou
bis

trcde-e Dig

Lew vaviance High icnce

4 In order to Teduce the mneded
e e , the
designes Curn aim at cducing citheethe bieS c the
Varvjcnceas. the noisSe Congpannts js ycduciblc.
SBias vefers to the ersc intyoduced by appieni
-naing
6] Bies
a
Ycl-ueTldsablem with a Sirnplifid mcdel.
Captures the differnte bctwcen the cKprted
frecictions of the mdel and the truC velues.
?J meded miaht crhibit lcw bias on the
txcunjn
data but high b'as on the test date.
bJ Voriancc tefers to the vaiabili ty of meded pcditions
ferdiffeacnt ialning Scds.
hew Snsitive the nedel is to
luctuatins in the trcininq da te:
lo vaioncc Can be affcctcd when the model isccmplex
Cnd has high capacity.
 data dst ibtkn ,bett

i p ) t to strike c blcncc by etinn

cdels tchnivc to dept tc the nea distri butit
nd avoiding CxeSSive 5cn5itiuty te Smal
Vaiaticnsin the dete.

* Tansfe laoning LDomain adpteticn

3 Fonsfey lacyning gefer to Situations cuhec whct

has been lccancd in cne Setling is CAploited tu
impycye qeneraliztiun in cnother Se ting
23 It can bc achievcd ia vepaCStntetien
lceonin when there
there eist
exist featiy
fratuyes that cc
UScful fos diffent ste Settings oy tasks.

3] Two extem feym cf tsensf leasning

a] ane-shot lccgning:
il Cnly ene eenple ot tansfey tesk is.
givcen fes One -Shet
It is possible baausc leasning
the yepyeStatatien
kaxns
cltanly Separate the undelying' cdases, during fist S

b] 2es0 -Shot lcenig

iNo labeled exAmple ave qivcn at gll
i] It is only possible becausc addticgel
infemation his boen exploitedduzing taining
Rti ask aning is q Sub ficld of machinc
lcasnng that aims to sclue multiple diffent
tsk at the Scc time
:1It trakes adeentqe of the SimileLias bett
df{eseot tsks.
3 multitsk lcaninq /s a luey to impiev
genytiuticn
by poling the ecunplks yisinq out cf Seyeed tsls.
4 Oiffscnt Supe uiscd tsks shercd tht Scmc
inut and scme intesntdiate -levcl seresentatien.
$] Tnductive tensfe Cun help
improve amodel
by inoducing an inductive bjas,

hod pegamcte
pecmets

3It is qencially applitd by shaing the

hidden lcayers betwcen cl! tosks.
iJ Had peramety shering grstly
iSk of eveyfitting.
geduccs the

TesK
Specitic
lays

sherecd
lays
CNN modcls LeNct5
AlNt

Residucd Netwck.