1

Getting Started
CS771: Introduction to Machine Learning
Purushottam Kar

CS771: Intro to ML
 2
An overview of ML
 Study similarity of ML with a student preparing for an exam
 Look at a toy ML problem Warning: lots of
oversimplifications
 Learn what is training data, test data? ahead!
 Learn what is a model?

CS771: Intro to ML
A typical study cycle (e.g. in a course) 3
TEST TEST TEST

Preparation
Practice Tests/Study
Material
Actual Exam
EXAM

Subject
Pass / Fail Knowledge

CS771: Intro to ML
 4
Spam Filtering
 Suppose Mary has already tagged several old emails as
spam/non-spam, can we tag her new emails too?
 Trick: use the old tagged emails to try and understand what sort
of emails does Mary think of as spam and which as non-spam!
 E.g. may find that emails about shopping always tagged as spam
 E.g. may find that emails from Jill are never tagged as spam
 These insights/patterns are what are stored in the spam filter
 Our spam filter helps us make predictions on new emails
CS771: Intro to ML
A typical ML workflow 5
Ah! That is the fun (and artistic) bit
ML
about ML. We will learn tons of
Algorithm
ways on how ML algos store
patterns
Old Emails The spam filter stores
information about your personal
preferences about what looks
New Emails spam to you

How is this
information stored?
Spam
Spam/Non-spam Filter
CS771: Intro to ML
A typical ML workflow 6
ML
Algorithm
Training Data

Test Data

Model
Output
CS771: Intro to ML
ML as an “examination” 7

Our brain stores subject matter The model stores data patterns
Use subject matter to solve exam Use model to predict on test data
Critical to do well on exam-day Critical to do well on test data
Mock test results indicative Training accuracies indicative
No out-of syllabus questions Training/test data are similar
Should not leak exam paper Should not look at test data while
before exam training CS771: Intro to ML
A typical ML workflow 8

CS771: Intro to ML
ML can do lots of cool things with test data9
In this course, we will learn how to do
Test Data
most of these operations with test data

Regression Multi-classification

URGENT,
Clustering Tagging OFFICIAL,
TAX

Ranking
Binary Classification
¿ ¿ ¿ CS771: Intro to ML
A typical ML workflow 10

CS771: Intro to ML
 11
ML can take in lots of kinds of training data
We wont be able to cover all these training settings
in this course – there are entire courses devoted to
specific training settings e.g. CS773 (Online
Learning)
Training Data

Batch Learning Active Learning

Robust Learning
Semi-supervised
Learning
Supervised Learning
Online Learning
Unsupervised Learning Reinforcement Learning

CS771: Intro to ML
A typical ML workflow 12

CS771: Intro to ML
ML can store info in lots of innovative ways 13
ML Models and Algorithms
Linear/Opt Local
We can mix-n-match Neural/Deep
these methods too e.g.
Bayesian Deep We will learn how to use each of
Learning or Kernel these techniques in the course, but a
Nearest Neighbours
(Local) Probabilistic/Bayesian
bit briefly. As before, there are entire
Kernel courses devoted to each technique
e.g. CS772 (Prob ML), CS774 (Opt)
Correct! But we will not be able to
cover such advanced methods either

CS771: Intro to ML
 14
Fantastic Features
 … and how to find them
 What are vectors
 How are vectors used in ML
 Useful operations on vectors

CS771: Intro to ML
What Weare features
could have – but it does not carry
much information about spam/non-
Why did we not keep
the word “to” as a
feature?
15
Features are a way for us to give input to ML algorithms
spam since it is such a common
Most ML algorithmsword! use numerical vectors to represent features
Guys, something is wrong with our feature. The
For example, in spam classification,
word “do” meansevery old
different email
things (training
in two emails data)
as well as every new email (test data) must be converted into a
vector Do You Want Go Million Dollars Dinner Today

Do you want to go for dinner? 1 1 1 1 0 0 1 0

Do you want to win a million dollars? 1 1 1 0 1 1 0 0

I have a million things to do today! 1 0 0 0 1 0 0 1

Since I am basically a computer program, I need you

to convert your data into a nice set of numbers
Good catch – this may not be the best feature representation! CS771: Intro to ML
What are features 16

May represent images May instead encode May instead encode

as a vector/matrix of images using edges images using presence
pixel RGB values present in them of eyes/noses/ears
Makes
Raw/ sense. Features are also a way for us to store what
Low-level we knowHigh-
Derived/ about
data. If we throw away data, naturally doing anything becomes hard!
features level features
If your features are good, my job becomes twice as
easy. If not, then doing any ML may become
CS771: Intro to ML
Types of Features Discrete numerical features
often called ordinal features
17
Numerical Features
Dangal
Dhoom 3 B’LORE DELHI KANPUR

Categorical Features
A B AB O
BTECH MTECH MS PHD

Relational Features
¿ ¿ ¿
Ranking in Class
Social Media
CS771: Intro to ML
Derived Features
As we saw, the bag-of-words (also called
18
Bagged/binned features BoW) is not always a very good feature (it
confuses polysemous words). However, it
ESC101 (A), ESO207 (B), CS220 (B),
CS340 (C), MSO201 (A), CS771 (A)
is still extremely popular due to its
A B C D E F
simplicity
Pooled/aggregated features
ESC101 (A), ESO207 (B), CS220 (B), 10 (max), 8.67 (avg)
CS340 (C), MSO201 (A), CS771 (A)
The way we represented
emails in the spam
problem is called the
bag-of-words feature

CS771: Intro to ML
Feature Selection?? Useful for predicting
19
expenditure
DIET (VEG/NON-VEG) Not useful for predicting expenditure
EYE COLOR
TOTAL INCOME
EDUCATION LEVEL
FAMILY SIZE
HOUSE NO (ODD/EVEN)
In fact, one reason for the success
NO OF CHILDREN
RENTED/SELF of deep learning is that it learns
OWNED
(EXPENDITURE)
good features themselves!
SURNAME LENGTH

Can we perform automatic Yes indeed, but those methods are

feature selection? a bit advanced for now!
True, but more on deep learning
later! For now, back to basics. CS771: Intro to ML
A bit of caution with features 20

Tricks, mnemonics lessen Derived features make learning

cognitive load, increase speed easier, faster at test
Easy questions can be solved in What you are trying to predict is
one step with a mnemonic! However,just another
not to feature
worry. For most ofof
thisthe data!
Too many mnemonics can course, weToo
will give
many youuseless
pre-made features
feature can
vectors 
confuse you at time of exam confuse classifier
In fact, one of the main challenges in deep
learning is that it learns way too many features
CS771: Intro to ML