Machine Learning -1/02/2025

Computer Seienca Hunas are intelu'gnt

(Leam nig apobibty)
Data S ence +Deision Mabing
+ Cognitie pro ess
Artifieial
(AI)Jaeiene
Artiicia
t Knoi lede representati on )
+NLP (NaAral'Lonuage froessin
compute Cemputer Vision
Opti
Saratition
y2
Macine Lear ning Ralationip between input Loutput
based on thad, eomputer can predit utput.

NL(Natnal
Conputer vsion
Langage froaising)
ML is a subpart ot A1 where machine will try to Learn tron the
enwir mment (data) and basel n that it will take deeision, elassify on
preiet
[TSP Trawelling olunan roblem
(ML

1959 Arthun Samuel

b ML (olf- Lear ning
IS a eomputer dedeat a deweleper in a

Instrueti on/Pregram
Imput Output
ML

Input Model Output

(learn fom
the input data)

ML Modet
’ Matthematieal
/statistieal Model or Eauatin
ML Algorithm
Unkne ML Algorithm
HNA’ Atihcial Neural Netork ’Bias (Baek fropogtion Agoritim)
Hyper- parameters farametens tat ae ae unknown to ud.

fr ANN, ty por parametrs

ML Madel Network
Bay sen
1 Linear Regresin Featureal
2. K-means clustering Gody shap)
3. SVM
ML Model
4. Deeision Tre face (ANN Male
female
Rigessim
6. ANN
(Eye
1. Bayes ian Network
Man Faetm tor ML implemetation:
1Data Dibferent input t Output eenbinati on)
2. ML Model
3. At NL
4.
Algerth
Cemputen
Tnstruetion Ys. HL
s DS, A1, ML

Lear ning 5 marks|

visd
Super ised Unupervised
Learning Learning Leamnig
’ ANN K-mens lusteri’ Autmomous vewele driving
’ Deeis ion Trea ’ vides gme playi
Input saw ail able bt [Supervis en is prent but it dort
output ore net ve er aet int mati n about utput
Tmputs k orespndi ng awail able Just tel whetem
botputs ame known and Me Supenvar wYn ALso give Ais
both are used lile Learnind positiv outeome rewards
for mat Data Conwentional Date format):

Input Dutput

4. Supervised Lear ningi- uper isor is avalatle (1nput cutput )

9. Unsupor vacd Learnng Ony 3nputs anail able, No eutpuk.
Seperotey data, (ereates elster).
3. Semi-super Vised leaning i Hybrid data. some dota
fes data contains input #qutput eombinam
only inpt.
 Machine Lear ning- 12/02/2025

Data ML Mal HL Algoitim Camputen

Dota Reppreserttim!
Tabulan tormat of data
SImpu/In dependent
Feture/Atribute/Variable t. output Depe ndemt
(Gist dor super vise Learnin)
2-3 G10A Tru
Ksw2G.2Bad 1 True

Vetr
Matrix
aregirement,
Gaenrally,Dataa Fotere
Featre 10

Minimum data requved( for Gteatures) 510 250

B oata No af fatures is mora than 100

ly Medical ImageSat ellita Tnage/ Giene Eprsion Data
(man) (m np
N. d speetrum

pro cesing Bie ata, CpU can fil. So, 6 GiPO is ed.
GpU Vertual Sraplies Proceacing syctem (on ine/ elod based
JervicQ

Pata Sorubbing (Dat ar elasmup ke preprocssin:

Feate ’ Inpul equtput
Featue Seleetin:
We shauld sebet relavent features and canal out tay feature
that is dupieate natue a ess usetul too predieti on.
Year, Ceuntry, Ceuntry Cok, Dollar value
dled.
Here, For euery cunty. a wminue eauntry code is provi
do the same job.
So, ony enunty
Rew Compresim:
Year Country Value Yean runtry Value
19 50 India In ia 40 5
50
1352
1951
952
Number et Raos cam be redued y delsting
mean vaue od features.
 One - Hot Fneodingi
body Hair Gundem
Shope (output
Male
Female (6)
Femal )

we eplaee the datures value (inguisti witth memeriead tn

value (/o) torr eomputational pur pose.

Biming Buaketing i
Flat si? Naluc
Kolkata 30L
936
Kolata 2/Small/2BH K
Kolkat a
4iS WAQd to eoTMent al value into muliple binay
features called binybuekets.
Nrmalization

True
True
false

Noriotion Noriation
Low

AU the values are braugt inte range lo.s]or [-1.].

Nami ation
cutput
norm ?0 Value- Xmin
mox m'n

It is te proseas aert, thorugh whieh Teseale all the valus

wi thin
Yang o,1] or C-1, 1), So that elheet of all the alues
ean bo cbsemved.
 Stondorigation' Fate vaue
nis
J4 is tha proeedure, to replaee el the data, eolan
with ts standard Deviati on.
LI4 distributi on s Gautsian

Hissing Nalue
tnsn it shaul be
1 vaue ot tes sam Aeature is
le meplaeed with maan/median (Mest
miss"g
Ommon valu),

Teolb
Eeapt Normalzatio) tter ting am be done wing
 /0/2025
Machine Lsarning
+Intro ducti on to ML
’ Data serubbing ond perfor mamce Meas urement

Linear Reges sicn

Logistie Rigpmessien Clasiticatin /Preietian
’K-NN
LK-mcans
elstorin Paten Roegriti
SVM
ANN
Deeision Tree
Bayesian Nehoork
Orer- iting. UnderAti
Data ML Model Neu
Data (Testing
Classifieation
Predietion
Training & Varianso
ML dutput is depend on Bias

Hiah ba Low vasian ce

Bias ’distanee trom rigro to predieted Bias is dasirabe )
JariancQ ’ Vary rom means.

vorian ee Law bias Lo varianee

 Naanee
bias
Varian ee
Predietion
Low Bias
Err
(Mean
Squart Test
Data

o Trairing Eyro small

hrain
Data
Underkitng Testing Erro tiah

Hedel Complevity
(1tevaton Number)

with MSE. to tuer teme

Rogularizabon ’To add a ponalty term
problem.

Penalty torm a Molel temnpnty

Suppont Veetm Machine
labelled Data
Supewvisd Learningsin Input+ utput
are nown
- elassii cotim pur pse
Marg

Optimal
Hypertane
Seattered ptet betucen YLX
Optimal Hyper plam' Th distanee ot the data points and hyperplanc dul
be minmum. Also mar gin sheuld be pinimmm
Tmanimum
Artieial Naral Netaor k (ANN)::
»Most Popu lar ML Aotel techique torr elassihieatn 4 predict
purpose

Soma
Axon
Glacbro ehemieal pulses. Dendron

Denites (trans potter)

(Rceiser)

Prn /quivalont Matlemateal Model

Arbheial Neuron APermceptran): -
| Dondron Soma
Synapse
Exute Not
Vo
(hceumulatr)
Thread (Vra)
Thnpuds
else

’ To consider ths importanca t each ter=m inputs, weigt term s

inbrodueed.
s a effset value, added to the aceeumulator. (Detult vale)
Bias
+ To fit with the
Bias traing data

Output od Aeeumylata, 7 bot X;wi

An Artiiial NeuTOn iS
called a ferugtron.
o called a
 Date- 12/ 03/2025
Artihcial Nearal Netoork (AN);
Accumulatn AQtivati on
funeti onm
y(Ve)
W
nput butput

Simple Artifieial Neuren

- Accumulated inpts

L=l

Dflerent activati on Funetion

1. Binary Threshald 1

9-0 net
0 else

2: Bipolar Thredhfe
Knet
y: ) - ele
3. Recewise inear func tion

net

Nast' popular 5

e
Mret
Detauttd K’Cortrd faramter 1 eontrl the slope)05y31
S. Hyperbslie tengendr lanh tuneton
et -Aet 2-5

et - Knet

- 2"5
thnet
6. Saussian Aetivoti sn
2ç

C> tanand Dewiation

clasiieation ot Neural Nehoork

fecd, Forwamd Neual NetworkN) Fead Back Neural Network

(eEN) (FBNN)
’Data Alow rom ingut towards ’Data travel t gm input to
output ana theme -is no fdbock Sukput amd vice verse, e
Path or loop in the network. there is a ecdbaek poth
lop in the. systuo
’Eg- A predieton,
Can een elassifieation. Eg- Weother der ecasti ng
Stook enelhange gredietin|
Tim Jerie
Predietn

Singe layem Fesd Muti layer Fed

For wand Neural
for ward Neural Network
Networ k
(SLFFNN) (ALFFNN)
Node

|biaslb)
Wa

Iiput Output
Layer layer
 Bias Matri

W
(Gaende)
1

eThere is no intermedinte lqyer in betwen input and autput layer, s,sFFNN

Hore, nput and autp ut are shown as nocos and eonneeted with edges
with proper weighta&
Dotatype
Non-linear Seperabe Data
Linear Sepurable data
(he relationship bekwen nput &autpue
-(u relationskip beheeen input k is mon- naar in notue. )
Butput is inear in nature.)
ine can't be drawn to
we ean draw a ine to sprate Segerat ot classify the data nto two
but th data in ' o'. Sets.

4 4

’ 7o solve k non incar seperade doto

o sowe Linear seperabe dota, SbfFNN MLFFNN eam be wed.
Cam e wsed.
lol

mol-prolem
Whot is xOf -problem in HL?
1 ’
 Yoe

SLFFNN