Introduction to Machine Learning

Prof. Sudeshna Sarkar

Department of Computer Science and Engineering
Indian Institute of Technology, Kharagpur

Module - 1
Lecture - 02
Different Types of Learning

Good Morning. We start with second part of the first module that is introduction module.
We will introduce the Different Types of Machine Learning.

(Refer Slide Time: 00:32)

So, there are few broad types of machine learning.

(Refer Slide Time: 00:41)

There is Supervised learning.

(Refer Slide Time: 00:48)

In supervised learning what we have is. We said that in learning we use data and data we
can say can comprise of input and the corresponding output. For every data instance we
can have the input x and the corresponding output y. And, from this the machine learning
system will build a model so that given a new observation x will try to find out what is
the corresponding y. This called supervised learning because for every instance we tell
what is the output. So, this is called labeled data.
Then we have Unsupervised learning. In unsupervised learning you are only given x,
there is no label to the data. And given the different data points you may like to for
example plaster them or summarize them or find some patterns in them. The third type of
machine learning is Reinforcement learning. In reinforcement learning you have an agent
who is acting in a environment, and you want to figure out what actions the agent must
take at every step. The action that the agent takes is based on the rewards or penalty is
that the agent gets in different states.

Apart from these three, we have also Semi-supervised learning. In semi-supervised

learning it is a combination of supervised and unsupervised learning. That is you have
some labeled training data and you also have a larger amount of unlabeled training data,
and you can try to come up with some learning out of them that convert even when the
training data is limited.

(Refer Slide Time: 03:13)

So, let us look at this picture which gives us a schematic diagram of supervised learning.
In supervised learning you have a number of training instances. This contains the training
instances, for the training instances each instance comprises of input and output. This is
the first training instance, the second training instance, the third training instance, the nth
training instance, given all these training instances the learning algorithm will come up
with a model.
And this model can be used to classify or to find the output or corresponding y value for
a new observation x. So given a new input x you can use the model to find out the output
y. This is the schematic of supervised learning.

(Refer Slide Time: 04:12)

(Refer Slide Time: 04:17)

In unsupervised learning as we discussed, we only have x’s. We have different x’s; x 1, x

2, x n this is the data. And the learning algorithm will produce clusters will group this
data. Based on the similarity of the data items to each other we can find out certain
groups among the data. So that is called Unsupervised learning.
(Refer Slide Time: 04:47)

And as I said we also have semi-supervised learning. In semi-supervised learning, we

have a combination of labeled data and unlabeled data. This is labeled data, the data
which belong to two different classes so one class is circle the other class is triangle; in
semi-supervised learning, apart from having data from the two classes you also have
unlabeled data which is indicated by the small circles.

For example, for supervised learning based on the data. Supervised data you will come
up with some function and if you also have unlabeled data in addition to the labeled data
you might try to come up with the better function.
(Refer Slide Time: 05:35)

And we will very briefly talk about reinforcement learning; in fact we will not cover
reinforcement learning in this introductory course on machine learning. But in
reinforcement learning we have an agent which acts in the environment. The agent can
take action and this action can impact the environment. In a particular stage, the agent
takes an action and the environment goes to a new state and gives some reward to the
agent, that reward may be a positive reward can be a negative reward or penalty or can
be nothing at that particular time step. But the agent is continually acting in this world.

(Refer Slide Time: 06:25)

And the reinforcement learner what it will do is, it will learn a policy. That is, given a
state what action to take so that not only the short term reward is optimized, but the
overall utility of the agent over its entire time horizon is optimized. This is what
reinforcement learning is about we will not talk about details of reinforcement learning
in today’s class.

(Refer Slide Time: 06:53)

Most of this course we will talk about supervised learning and we will spend some time
on unsupervised learning. So let us look at more details about what supervised learning is
about. In supervised learning you have a set of input features.
(Refer Slide Time: 07:30)

So what you have is you have some features x 1, x 2, x n, these are the features with
respect of which you describe the instances. And you have a target fixture y, and you
have several instances each instance comprises of values. Suppose, this is the instance 1
it comprises of values of this feature x so you may have a 1, a 2, a n, and you have
instance 2 which is b 1, b 2, b n. Instance 3 which is c 1, c 2, c n. And the corresponding
you have the y values; y 1, y 2, y 3, etcetera.

So, your instances are described in terms of features. There are input features and there is
the output attribute. And you have a set of training examples this comprises the training
examples, and in each training example the values for the input features and the target
feature are given for each example. Each of these rows is one example and for each
example we have the value for instance 3 the value for x 1 is c 1, the value for x 2 is c 2,
the value for x n is c n and the value for y is y 3. This is used for training. During the
testing time you are only given, so use like this you have many training examples for test
you are given a test instance. In test instance you are only given the values of the
attributes; z 1, z 2, z n and you have to output the corresponding y.

Now the y or the target feature can be a discrete valued feature or a continuous valued
feature. If y is a discrete valued feature we call such problem classification problem. So y
can be discrete valued or continuous valued. For example, we can have a discrete valued
feature like whether it will rain or not tomorrow, such problems are called classification
problems. Or we can have a discrete valued feature which given the symptoms of the
patient, predicts whether a patient has a particular disease. Or you have continuous
valued feature then you call such problems regression.

Suppose given the values of certain features you want to predict the price of a house, so
given the location of the house, the floor area of the house, the number of bedrooms and
some such features, you want to predict the price of the house which is a continuous
valued or real valued number and such problems are called Regression problems.

(Refer Slide Time: 11:16)

So, let us look at some examples of classification problem. One example is Credit
scoring. Suppose, in this problem we are describing each input in terms of two features;
income and savings. We have a person, we look at the income and the savings of a
person and we want to predict whether this is a high risk person or a low risk person.
And you seeing this picture, the high risk persons are labeled with minus and the low risk
persons labeled with plus.

Now given this data you want to come up with a classifier which given the attributes of a
new person will predict whether that person is high risk or low risk. And this is an
example of a rule that you have found if income is greater than theta 1 and savings is
greater than theta 2 then the person is low risk else the person is high risk. This rule you
have come up with the rule based on the data. In this case you can visualize the data and
come up with the rule.
(Refer Slide Time: 12:31)

Now let us look at an example of a regression problem. You want to find out the price of
a used car and you use certain attributes of the car to predict its price. For example, let us
say that you have only looking at the mileage of the car and based on the mileage you
have predicting the price. And these different points they correspond to a particular car
show what is the mileage and what is the corresponding price.

So given the mileage price of several cars you have the data points and you can come up
with the function so that given a new car whose mileage is given you can predict the
price. In regression you come up with a function which takes the input instance and the
parameters of the model.
(Refer Slide Time: 13:28)

Now, we see that in order to describe an instance features play a very important role.
Observations that you see are analyzed into a set of quantifiable properties which are
called features. So if you want to predict the price of a car or the price of a house you
have to have the right features which will help you to come up with the price. If you
want to predict what disease a patient has you have to come up with the right tests and
parameters which will enable you to make that prediction.

Now feature are of various types. Features can be categorical. For example, blood group
is a feature which has four values “A”, “B”, “AB” and “O”, or gender is feature which
has certain values like, male, female etcetera. Or features can ordinal like, large,
medium, small. Features can be integer valued like, the number of words in a text or a
feature can be real valued like height.
(Refer Slide Time: 14:31)

Now, let us look at this table which shows a sample training examples and for these
training examples you have five features; action, author, thread, length, where. And the
different rows are the different instances. So instance e 1 says; action is skips, author is
known, thread is new, length is long, and the person is at home. These are the different
instances based on the training examples comprising of this labeled instances you
learned a model so that for a new instance you can predict the output feature.

Suppose, you have to predict the action of the user whether the user skips a link or reads
a link and given the value of Author, Thread, Length and Where.
(Refer Slide Time: 15:25)

So, if you come up with the schematic diagram in supervised learning you have the
training set, the learning algorithm uses the training set to come up with a model or
hypothesis, will introduce this term in more detail later on. And in the testing phase given
a new instance you use the hypothesis to predict the value of y.

(Refer Slide Time: 15:49)

This can also be shown like this, in the training phase you get the input and the label.
From the input you extract the features of the input and feed it to the machine learning
algorithm. Similarly in the testing phase given the input use a feature extractor to extract
the features and you feed it to the classifier model to get the label.

(Refer Slide Time: 16:12)

So in classification learning these are the components. You have a task T which has input
and output, there is a performance metric P, and there is experience E.

(Refer Slide Time: 16:23)

The task T comprises of input which is a set of instances d 1, d 2, d n, each instance has a
set of features we can represent an instance as a vector d equal to x 1 x 2 x n and you
have the output a set of predictions for the inputs.
(Refer Slide Time: 16:44)

Now, what we will do is that we will look at some sample classification tasks. Let us say
the task is medical diagnosis, instance are patient record, comprising of the features of
blood pressure, diastolic and systolic, age of the patient the sex of the patient, the BMI of
the patient, cholesterol of the patient. And the labels are whether the person has is low
risk of heart disease or high risk of heart disease, there are two classes we call them
positive class and negative class. One is the positive class; the other is a negative class.

Then there are some other examples finding company names in text. So given a
particular word you want to find out is it the name of a company. This is the entity
recognition task in text (Refer Time: 17:42). So, the instance is a word and the context of
the word. What words, which words come before the word, which words come after the
word, and may be certain features of the word.

For example, is the word capitalized, is the word following this word are i and c, is the
bigram or two words previous this word context acquired by and so on. And the output
labels are; first, later, outside. That is, if it is a company name is it the first word in that
name, is it inside a company name, is it an outside a company name.
(Refer Slide Time: 18:34)

For example, in a running text you may have Microsoft corporation acquired x y z. Let
us say Microsoft corporation together is the name of a company, and this is the first word
in the name, this is a later word in the name. And let us say xyz is also a company name
and this is also the first word in the name and the word acquired is not a company name
so it is outside. We may use this sort of convention in order to label company names in
text, and we can have a machine learning algorithm finding these labels.

Similarly, let us look at the fourth row which is image recognition, the input are image
which is 1920 by 1080 pixels and each pixel has a color. And your output is whether the
image is the image of a house or not.
(Refer Slide Time: 20:01)

So the performance metric in classification learning can be, what is the probability of
wrong prediction, what is the probability or it can be the performance metric accuracy or
the probability of wrong prediction of the error on examples from the distribution from
which the instances are drawn. In classification learning the task comprises as we said a
set of input instances d 1, d 2 d n and the output is a set of predictions y 1, y 2 y n.

The performance metric is what is the probability of wrong prediction and the experience
is a set of labeled examples where y are the true labels for x for those training examples
we have the ground truth data. And these examples come from some fixed distribution.
During the training this examples are drawn from a distribution and what is expected
when we get text instances, those examples also come from the same distribution.
(Refer Slide Time: 21:16)

Now, how do you get data for the learning problem? For example, in medical diagnosis
problem which we have seen, in order to get data when you have a patient you find out
the different parameters of the patient and we have to wait for the next several years to
find out did the patient get heart disease or not. So, we have to wait to look further heart
disease.

Finding company name in text, how do you get the data? You get some text and you get
the manually annotated for company names this forms the training set. For image
recognition the data is a set of images for which some human or some people have
labeled those images. So, this is how you get data.
(Refer Slide Time: 22:07)

As we said that when we want to look at the representation of the function there are two
things; one is the features, the other is the function class. And we will talk about this in
greater detail later. The function class can be a linear function in terms of the attributes;
the function can be a decision tree which we will talk about in week 2. It can be a linear
function; it can be a decision tree.

(Refer Slide Time: 22:39)

Or, it can be other functions like multivariate linear function as shown in the picture or
single layer perceptron which is a unit of a neural network which we will talk about later
in this class.

(Refer Slide Time: 22:52)

Or it can be multilayer neural network which we will also cover later.

(Refer Slide Time: 22:57)

And given the type of function, the hypothesis space is the set of candidate outputs that
you can get, candidate functions that you get. So supervised learning you can think of,
there is a set of functions which comprise the hypothesis space and you want to find out
that function from the hypothesis space which is most probable given you a training
example. Again we will talk about this in detail later.

(Refer Slide Time: 23:36)

So, if you just look at the basic terminology we have some features which are distinct
traits or attributes that are used to describe each instance. And the set of features they
comprise the feature vector. The instance space is a set of possible objects that you can
describe by the feature, and each example has the value of the input and the value of the
output.

(Refer Slide Time: 24:04)

Suppose you are try to find out is this the object at image of a phase. So, the concept is a
subset of objects and the instance space, among all possible images some images a subset
of the images and images of phase. The target function is a function that you are trying to
learn it will map each instance to its label. That is it a phase or is it not a phase. We have
already talked about what is example and what is training data.

So with this we end today’s module and we will star with the next module in the next
class.

Thank you very much.