Introduction to Machine Learning

Sanjay Talbar
Professor, Department of CSE-AIML,
CSE PCCOE Pune
(Ex-Professor,
Professor, SGGS Institute of Engineering & Technology, Nanded (MS) India)
nowledgements
chine Learning, Deep Learning and Computer Vision researchers who have m
ir resources publicly available on the internet.
internet
line coarse “Machine Learning” and “Deep Learning Specialization” by Andrew Ng
ursera.

TEL Video Courses:

ntroduction to Machine Learning by Prof. Sudeshana Sarkar (IITKGP)
ntroduction to Machine Learning by Prof. B. Ravindran (IITM)
Machine Learning and Deep Learning for Engineers and Scientists By Prof Balaji (IITM
Machine Learning by Carl Gustaf Jansson, KTH The Royal Institute of Technol
tockholm, Sweden
Machine Learning Course by Stanford University, CA
Machine Learning Foundation by Google/CML/IBM…..
Google/CML/IBM
oduction to AI, ML,DL:

N: LSTM, GRU Linear Regression,

Regression
N: AlexNet, Logistic Regression, Robot Examples like-
GNet, Decision Tree, Sophia(Hanson)/Asino(Ho
ogleNet, ResNEt, SVM, Naive Bayes.
Bayes Rolling Bolts(LG)
gNet, Unet, etc kNN, K-Means
Means, Spot(Google)Spider/Snake
Random Forest, etc Kangaroo/Butterfly Robot
 ML, ANN, DL …….Machine Learning

Any method that tries to replicate the results of some aspect of

man recognition
: Program/Algorithm that performs better with experience
N: ML Algorithm based on biological neurons
: Subset of ANN (Classification, Recognition, Detection,
gmentation)
achine Learning is a branch of Artificial Intelligence and Computer
ence which focuses on the use of data and algorithms to imitate the
y that human learns, gradually improving its accuracy.
chine Learning:
hine Learning definition by Tom Mitchell
rmally, A computer program is said to learn from experience (E )
spect to some class of task (T) and performance measure (P) i
rformance at tasks in T as measured by P, improves with experience.

ample: Recognizing emails as spam or not?

Recognizing spam mail
Reading mails and label them as spam or not spam
How many mails are you label correctly as a spam?
Accuracy or % classification

formance of doing particular task is improved

h experience then it is called Machine Learning
chine Learning Applications
chine Learning Applications
 ications
mputer Vision: Classification, detection, Recognition
ical Character Recognition (OCR)/Handwritten Digit/
t recognition
ech Recognition, Identification, Speech to Text, Text
Speech, Language translators,
medical Diagnosis: X-rays, MRI, CT, FMRI, ECG, EEG...
metric
Fingerprint Recognition/Identification
Iris recognition
Face recognition
onomous Cars/Vehicles/Underwater Vehicles/Military
ks/Space vehicles
omated Target detection
ustrial Inspection
ecasting
Financial Forecasting/ Rain Forecasting
Housing Price forecasting/ Stack Market forecasting
Electricity forecasting
L Approach
Feature
Classifier
Representation

ure Extraction Methods Classical Classifiers

atures based on first-order/ • Baye’s Classifier
econd-order statistics (Mean, • Decision Trees Random Forest,
ariance, Energy, Entropy, GLCM, • K-Nearest
w’s Energy Masks) • SVM, Adaboost
BP, Ternary patterns, directional • Neural Network (MLP-
atterns Backpropagation Algorithm)
ansform based (DWT, Gabor, • Clustering
urvelet, etc)
arning Paradigm or ML Techniques:
vised Learning: Supervised machine learning, is defined by its use of labeled datas
algorithms to classify data or predict outcomes accurately.

Data Supervised
Classical Answers Rules
Machine
Programming
Learning
Approach Answer
Approach

put data is fed into the model, the model adjusts its weights until it has been fitted
priately. This occurs as part of the validation process to ensure that the model
s overfitting or underfitting.
vised learning helps organizations solve a variety of real-world
real problems at scale, s
fying spam in a separate folder from your inbox.
vised learning:- Neural networks, Naïve Bayes,
Bayes Linear Regression, Logistic regressio
om forest, and Support Vector Machine (SVM).
 arning Paradigm or ML Techniques:
pervised Learning: It uses ML algorithms to analyze and cluster unlabeled datasets.

over hidden pattern or data groupings

out the need of human intervention. Unsupervised Clusters , G
Data
Cluster similar documents based on text Machine
Learning
Approach

method has ability to discover similarities and differences in information which mak
for exploratory data analysis, cross-selling
selling strategies, customer segmentation, and
attern recognition.

so used to reduce the number of features in a model: Dimension reduction-PCA

reduction an

method includes:---- NN, K-means

means clustering, and probabilistic clustering methods
arning Paradigm or ML Techniques:
orced Learning:

forcement machine learning is a machine learning model that is similar to supervis

ng, but the algorithm isn’t trained using sample data.

s model learns as it goes by using trial and error.

equence of successful outcomes will be reinforced to develop the best recommend

icy for a given problem.

mple: Chess game playing, car driving, Bicycle learning

arning Paradigm or ML Techniques:
forcement learning is a machine learning training method based on rewarding desi
viors and punishing undesired ones.
neral, a reinforcement learning agent -- the entity being trained -- is able to percei
nterpret its environment, take actions and learn through trial and error.
 ven Steps in ML Paradigm:
Gathering data: Deciding what data required is? means is a part of a problem

reparing/Preprocessing Data: Ensuring that there is no missing data, removing out

ltering, no bias, no class imbalance, normalization etc.

hoosing a model/Algorithm/Hypothesis/Function: Random Forest, SVM, ANN…

raining: Using data to determine model parameters (possibly gradient descent)

valuation: How well did a model perform? (Validation/testing)

Hyper-parameter
parameter Tuning: Adjusting the parameter values which was fixed before tra

rediction: Deployment
attern recognition
Bag of Algorithms that can be used to provide some intelligence to a machine
These algorithms have a solid probabilistic framework

The assignment of physical object or event to one of the several pre-specified

pre categ
…..Duda & Heart (Pattern Recognition)

Pattern is an object, process or event that can given a name

Pattern Class (or category) is a set of patterns sharing common attributes and
ginating from the same source
eature represents a measurable piece of data that can be used for analysis e.g. Nam
e, Sex, Fare, and so on. Features are also referred to as “variables” or “attributes.”
epending on what you're trying to analyze, the features in your dataset can vary wi
eatures across a class need to be discriminative for better classification performanc
uring recognition (or classification) given objects are assigned to prescribed classes
classifier is a machine which performs classification
ata Mining
Data mining makes use of various methods:-
methods: Classification, clustering, &recognition
exploit the insights which are present in a large dataset
t helps to predict outcome based on history of events taken place.
Ex:
• The amount person spends on a monthly basis based on his previous transactio
• The frequent items which are bought by customers like bread, butter, milk, jam
always together
• Trend in the market can be analyzed like demand for umbrella during rainy seas
• Demand of ice-cream
cream in summer/Demand of sweaters in winters, etc..
The main objective is to analyze the pattern present in the dataset and obtain usefu
nformation based on target requirement
Ex: What would be yield of crop in present year?
• What are chances of a person having particular disease when all symptoms are given?
• What is expected sale of groceries in a particular month?
• What is expected loss/profit percentage expected in th coming year?
ecommendation System
System which predicts ratings of user give to specific item (IMDB-ratings
(IMDB to a movie
t deals with “likes” and “dislikes” of the user.

Objectives: to recommend an item to a user which has higher chances of liking or is

need of a particular user based on his previous purchases

The aim of the recommender system is to supply simple, accessible, high quality,
ecommendation for user community
Ex:
• Amazon uses such RS to suggest products to users base on data they look for
• Spotify recommend songs similar to one you repeatedly listened

Similarly RS is used by Google, YouTube, Netflix, Instagram, Spotify, Ameron, retail m

Banking sector, credit cards
urse Syllabus: Th:3credits, Pr:2hrs (1credit) (10+10+30+
Unit-1: Basics of ML, Applications, Machine Learning Techniques: Supervise
earning, Unsupervised Learning, Reinforcement Learning, Data Format
Underfitting and Overfitting, Creating training and test datasets, Feature selectio
nd filtering, PCA

UNIT-2: Regression Techniques: Linear Regression and higher dimensionality, Ridg

asso and ElasticNet Regression, Polynomial Regression, Robust Regression, wit
andom sample consensus, Stochastic gradient descent algorithm, Finding optim
arameters through grid search

UNIT-3: Supervised Learning: Classification:

Classification K-Nearest Neighbour, Decision Tre
lassification, Random Forest, Support Vector Machine, Evaluation Measures: SS
MME, R2, Confusion Matrix, Precision, Recall, F-Score, ROC-Curve, Cross-validation
urse Syllabus:
Unit-IV: Unsupervised Learning: Introduction to clustering, Types of clusterin
Hierarchical, Agglomerative Clustering and Devise clustering, Partitional clustering
-means clustering, Finding optimal number of clusters using Elbow metho
ssociation Rule Learning
UNIT-V: Artificial Neural Network: Introduction to ANN, Perceptron, Cost Functio
Gradient Checking, MLP and Backpropagation Algorithm used to learn parameters o
NN, Random Initialization
UNIT-VI: Reinforcement Learning & Ensemble Method: Reinforcement Learnin
hrough feedback network, Function Approximation, Ensemble Method: Bagging an
oosting, Stackig and learning with ensembles, Random Forest
eferences:
• Guiseppe Bonaccorso, Machine learning Algorithms, Packt Pub
• John Patterson, Adom Gibson, “Deep Learning:
Learning A Practitioner's Approach, O’REILLY
• Han, Jiawei Kamber et al., “data Mining: Concepts and Techniques”, Elsevier
urse Objectives and Outcomes:
ourse Objectives:
1. To understand the basics of Machine Learning
2. To apply different machine learning models to various datasets
3. To design and implement various machine leaning algorithms and improve m
performance
4. To interpret the model results and analyze the accuracy of the model
ourse Outcomes:
After learning the course, student will able to do:
1. Distinguish basic applications and issues of Machine Learning
2. Use different data processing techniques to prepare training and test data set.
3. Solve various problems using regression algorithm and improve model performance
4. Analyze various ML techniques and algorithms
5. Evaluate the performance of the model using different performance metrics
6. Interpret and Analyze results with reasoning using different ML techniques
age Classification,, Localization and Detection

assification:: Just classify the main object in the image.

(Classify
Classify an image based on the dominant object inside it)
it

assification + Localization: Want

ant to know the bounding box of the main object.

bject Detection:: There are multiple objects in the image, and we want to know the class an
e bounding box of each object for all known classes.

stance Segmentation: Label each pixel of an image by the object class and object instance that
elongs to.
ssification vs Segmentation
mage segmentation is the process of assigning a label to every pixel
n an image in such way that pixels with the label share certain
characteristics. It is mainly used to locate objects and boundaries like
ines and curves in the images.

e difference between segmentation and classification is clear at some extend. And there is a one
erence between both of them.
he classification process is easier than segmentation, in classification all objects in a single image is
grouped or categorized into a single class.
While in segmentation each object of a single class in an image is highlighted with different shades t
make them recognizable to computer vision.
p Learning Pipeline:

Classifier
gredients for Deep Learning

Data

Computation

Algorithms
Email: [email protected]
[email protected]
[email protected]
Cell phone: 9850978050