Machine Learning

Introduction
DATA SCIENCE
 Machine Learning
• Learning is any process by which a system
improves performance from past experiences.
Herbert Simon
~ Herbert Simon

• Machine Learning corresponds to computer

programs or algorithms which improve their
performance through experience without
being explicitly programmed.
Why?
• Develop systems which can automatically adapt and customize
themselves.
• Discover new knowledge from large datasets (Data mining)
• Ability to mimic human and replace monotonous tasks. Eg. OCR,
handwriting recognition.
• To develop systems that are too difficult and expensive to construct
manually.
• Speed up the innovation and analytics
• Make more sense of chaotic world around us
Why now?

• Surge of Big data in our lives : We

create new data in amounts of
petabytes everyday through our calls,
sms, chats, selfies, videos, emails.
• Increasing Computational power
(faster CPUs and GPU cores)
• Growing collaboration among
researchers in academia and industry.
• Profit margin for corporates
Challenges with Big Data
• Volume
The sheer size of data we are handling is increasing exponentially everyday
• Velocity
The rate at which new data is being gathered by our systems and sensors
• Variety
Diverse formats and sensor type result in different data point to represent
• Veracity
The data quality is really bad in most cases with little or no structure in them
Concept of learning in ML
We have a task T, performance measure P, experience E.

We say a system is learning when while performing task T, the performance

measure P improving as it goes through more experiences E.

Case in point: Task = To make better route decisions

Performance = Less travel time
Experience = Travelling through different routes and time

If the system is learning it should start making decisions about selecting routes
as it experiences more routes and time associated.

Depending on various features learning can go in either direction.

Take an example of object detection

The problem statement is to detect object present in a scene.

Now consider any natural image as one presented here. It can have
multiple features, brightness level and color difference through all
pixels it contains.

The way computer sees the data of just 4 x 4 pixel at base level:
x0011FD x0011FF x0011FD x0012FF
x0011FD x0011FF x0011ED x0011FF
x0011BD x0011FF x0011FE x0011FE
x0011ED x0011FF x0011FA x0011AB

The idea is to output the coordinates of pixels which form closest

rectangle around the object, label of the class and percentage.

If the system is learning ideally it should return correct classes with

closest coordinates around the object and with higher confidence
value.
What are the Challenges?
• To create systems which perform with accuracy and precision of
human intelligence and can leverage machine’s innate architecture to
accelerate and keep up with big data (Scales of terabytes or petabytes
of data)
• To allow for faster and better decision making through machines
wherever possible
• To convert large amounts of raw data into useful analytics
• To help forecast the future trends and correct our estimates based on
the analytic output
Terminology
Population : The population is any specific collection of objects of interest.

Sample : Sample is any subset of the population.

Measurement : A measurement is a number or attribute computed for each member of a

population or of a sample.

Parameter: Parameter or feature is a value that summarizes some aspects of the whole
population.

Inference: Any key knowledge about the data sample or population as a whole from its
attributes and properties
Features?

Color: Red
Type: Fruit
Weight: 100 gm
Price: 140/kg
Availability: Yes
Sweet: Yes
Organic: No
Types of Machine Learning
• Supervised Learning (We provide output labels/tags with input data)
• Regression/Prediction
• Classification

• Unsupervised Learning (No output labels are given)

• Reinforcement Learning (Works on reward policy feedback system)

Supervised
learning
(Labelled
Data)
We provide human annotated
data to the model as ground
truth
Unsupervis
ed learning
To infer a function that describes
the structure of "unlabeled" data
(i.e. data that has not been
classified or categorized)
Reinforcem
ent
learning
A bot/software ought to take
actions in an environment so as
to maximize some notion of
cumulative reward
Supervised vs Unsupervised
• The procedures in supervised learning are well comprehensible due
to their structure. It is possible to contrast different methods, to
parameterize and thereby find a solution that is optimal for the
application . The interpretation of the data is easier due to the given
traceability than with unsupervised learning methods.
• The disadvantage, however, is often a very high manual effort in the
preparation of the data.
• It require lot of man hours to prepare fully formatted data to work
with supervised learning.
Supervised vs Unsupervised
• The advantages of unsupervised learning are the partially fully
automated creation of models. These can produce a very good
prognosis about new data or even create new content. The model
learns with each new record and at the same time refines its
calculations and classifications. Manual intervention is no longer
necessary.
• The biggest disadvantages are there is no control over what model
learns. It can start to cluster wrong type as one group
• It can give bad results amiss the output labels and can lead to lot of
misclassifications.
Regression
• Regression refers to correlation between dependent and independent
variables.
Consider eqn of line: Y = mx + c
Constant/parameter

Independent variable

Dependent Variable Coefficient/parameter

In statistics regression is defined as a measure of the relation between the

mean value of one variable (e.g. output) and corresponding values of other
variables (e.g. time and cost).
• Regression is used to fit a
eqn through the given data
points.
• We try to approximate the
best function to fit through
the data
• The core idea it to predict
the points in future. We
assume the data
distribution will hold up and
new points will be near to
the function.
• It reduces uncertainty with
the data and helps in
correct estimation over a
period of time.
• Few Cases for Regression could be:
• Sales Prediction
• Product Pricing
• Expense overrun variables
• House pricing prediction
• Population estimates
• Pollution particulate estimates

Honestly regression finds its usage anywhere we want to find correlations:

Temperature vs. Number of cones sold at ice cream store

Inches of rain vs. new cars sold
Daily Snowfall vs. number of skier visits
Classification

• In machine learning and

statistics, classification is the
problem of identifying to
which of a set of categories
(sub-populations) a new
observation belongs, on the
basis of a training set of
data containing
observations (or instances)
whose category
membership is known.
• Use cases of Classification
• Spam Filtering
• Cancer diagnosis
• Text classification
• Sentiment analysis
• Object classification
• Face detection

Similar to Regression this technique can be generalized into any problem where
we know there are discrete classes.

Classification of people based on their eating habits:

Vegetarian, Vegan, Non-Vegetarian