Introduction to Machine Learning Completed

[Music]

[Music] hello River and welcome to this NPTEL  course on an introduction to machine learning in  

this course we will have a quick introduction to  machine learning and this will not be very deep in  

a mathematical sense but it will have some amount  of mathematical rigor and what we will be doing  

in this course is covering different paradigms  of machine learning and with special emphasis on  

classification and regression tasks and also will  introduce you to various other machine learning  

paradigms in this introductory lecture set of  lectures I will give a very quick overview of the  

different kinds of machine learning paradigms and  therefore I call this lectures machine learning a  

brief introduction with emphasis on brief right so  the rest of the course would be a more elongated  

introduction to machine learning right so what  is machine learning so I'll start off with a  

canonical definition put out by Tom Mitchell in 97  and so a machine or an agent I deliberately leave  

the beginning undefined because you could also  apply this to non machines like biological agents  

so an agent is said to learn from experience  with respect to some class of tasks right  

and the performance measure P if the learners  performance tasks in the class as measured by  

P improves with experience so what we get from  this first thing is we have to define learning  

with respect to a specific class of tasks right it  could be answering exams in a particular subject  

right or it could be diagnosing patients of a  specific illness right so but we have to be very  

careful about defining the set of tasks on which  we are going to define this learning right and the  

second thing we need is kind of a performance  measure P so in the absence of a performance  

measure P you would start to make vague statement  like oh I think something is happening right that  

seems to be a change and something learnt there  is some learning going on and stuff like that  

so if you want to be more clear about measuring  whether learning is happening or not you first  

need to define some kind of performance criteria  right so for example if you talk about answering  

questions in an exam your performance criterion  could very well be the number of marks that you  

get or if you talk about diagnosing illness then  there then your performance measure would be the  

number of patients that you save or the number of  patients who didn't have adverse reaction to the  

drugs you gave them there could be a variety of  ways of defining performance measures depending  

on what you are looking for right and the third  important component here is experience right so  

with experience the performance has to improve and  so what we mean by experience here in the case of  

writing exams it could be writing more exams right  so the more the number of exams you write the  

better you write it better you get it test taking  or it could be patience in the case of diagnosing  

illnesses like the more patients that you look at  the better you become at diagnosing illness right  

so so these are the three components so you need a  class of tasks you need a performance measure and  

you need some well-defined experience so this kind  of learning rate where you're learning to improve  

your performance based on experience it's known  as this kind of learning where you are trying  

to where you learn to improve your performance  with experience is known as inductive learning  

and and the basis of inductive learning goes back  several centuries people have been debating about  

inductive learning for hundreds of years now and  it's only more recently we have started to have  

more quantified mechanisms of learning right so  but one thing I always point out to people is  

that you should take this definition with a pinch  of salt so for example you could think about the  

task as fitting your foot comfortably right so you  could talk about whether a slipper fits your foot  

comfortably or little so I always say that you  should take this definition with a pinch of salt  

because let's take the example of a slipper you  know so the slipper is supposed to give protection  

to your foot right and a performance measure for  this slipper would be whether it is fitting the  

leg comfortably or not or whether it is you know  as people say there is biting your leg or is it  

Chaffin your feet right and with experience  you know as the slipper knows more and more  

about your foot as you keep wearing the slipper  for longer periods of time it becomes better at  

the task of fitting your footprint as measured by  whether it is shattering your foot or whether it  

is biting your foot or not right so would you say  that the slipper is learned to fit your foot well  

by this definition yes right so we have to take  this with a pinch of salt and so not every system  

that confirms to this definition of learning can  be set to learn usually okay so going on so there  

are different machine learning paradigms that we  will talk about and the first one is supervised  

learning where you learn an input to output map  right so you're given some kind of an input it  

could be a description of the patient who comes  to comes to the clinic and the output that have  

to produce is whether the patient has a certain  disease or not so this they had to learn this  

kind of an input to output map or the input could  be some kind of a question right and then output  

would be the answer to the question or it could be  a true or false question I give you a description  

of the question you have to give me true or  false is the output and in supervised learning  

what you essentially do is learn and mapping from  these inputs to the required output okay if the  

output that you are looking for happens to be a  categorical output like whether he has a disease  

or doesn't have a disease or whether the answer is  true or false then the supervised learning problem  

is called the classification problem right and  if the output happens to be a continuous value  

like so how long will this product last before  it fails right or what is the expected rainfall  

tomorrow right so those kinds of problems they  would be called as regression problems these are  

supervised learning problems where the output  is is a continuous value and these are called  

as regression problems so we look at in more  detailed classification and regression as we  

go on right so the second class of problems  are known as unsupervised learning problems  

right where the goal is not really to produce an  output in response to an input but given a set of  

in data right we have to discover patterns in the  data so that is more of the Tuscola unsupervised  

learning there is no real desired output that we  are looking for right we're more interested in  

finding patterns in the data so clustering is one  task one unsupervised learning task where you're  

interested in finding cohesive groups among  the input pattern rights for example I might  

be looking at customers who come to my shop right  and I want to figure out if there are categories  

of customers like so maybe college students could  be one category and sewing IT professionals could  

be another category and so on so forth and when  I'm looking at this kinds of grouping in my data  

so I would call that a clustering task right  so the other a popular unsupervised learning  

paradigm is known as Association rule mining or  frequent pattern mining where you are interested  

in finding a frequent co-occurrence of items  right in in in the data that is given to you  

so whenever a comes to my shop B also comes to  my shop right so those kinds of co-occurrence so  

I can always say that okay if I see a then that  is likely very likely that B is also in my shop  

somewhere you know so so I can learn these kinds  of associations between data right and again we  

look at this later in more detail and these are I  mean there many different variants on supervised  

and unsupervised learning but these are the  main ones that we look at so the third form  

of learning which is called reinforcement learning  it's neither supervised nor unsupervised in nature  

and typically these are problems where you are  learning to control the behavior of a system and  

I will give you more intuition into reinforcement  learning now in one of the later modules so like I  

said earlier so for every task right so you need  to have some kind of performance measure so if  

you're looking at classification the performance  measure is going to be classification error so  

typically right so we will talk about many many  different performance measures in the duration of  

this course but the typical performance measure  you would want to use is classification error  

right so how many of the items or how many  of the patients did I get incorrect so how  

many of them who are not having the disease did  I predict had the disease and how many of those  

that had the disease that I missed right so that  would be one of the measures that I would use and  

that would be the measure that we want to use but  we will see later that often that's not it's not  

possible to actually learn directly with respect  to this measure so we use other forms right and  

likewise for regression again so we have the  the prediction error suppose I say it's going  

to rain like 23 millimeters and then it ends  up raining like 49 centimeters I don't know  

so that's a huge huge prediction error right and  in terms of clustering so this is little becomes  

a little trickier to define performance measures  we don't know what is a good clustering algorithm  

because we don't know what how to measure the  quality of clusters so people come up with all  

different kinds of measures and so one of the more  popular ones is a scatter or spread of the cluster  

that essentially tells you how how spread out the  points are that belong to a single group if you  

remember we are supposed to find cohesive groups  so if the group is not that cohesive it's not all  

of them are not together then you would say the  clustering is of a poorer quality and if you have  

ways of measuring things like like was telling you  so if you know that people are college students  

right and then you can figure out that how many  what fraction of your cluster or college students  

so you can do these kinds of external evaluations  so one measure that people use popularly there  

is known as purity right and in Association rule  mining we use variety of measures called support  

and confidence it takes a little bit of work to  explain support in confidence so I'll defer it  

until I talk about Association rules in in detail  and in more in the reinforcement learning task so  

if we remember I told you it's learning to control  so you're going to have a cost for controlling the  

system and there so the measure here is cost and  you would like to minimize the cost that you're  

going to accrue while controlling the system so  these are the basic machine learning tasks so  

there are several challenges when you are trying  to build a build a machine learning solution right  

so a few of these I have listed on this slide  right the first one is you have to think about  

how good is a model that he have learned right  so I talked about a few measures on the previous  

slide but often those are not sufficient there  are other practical considerations that come  

into play and we look at some of these towards  the middle of the course somewhere right and the  

bulk of the time would be spent on answering  the second question which is how do I choose  

a model right so given some kind of data which  will be the experience that we are talking about  

so given this experience how would I choose how  would I choose a model right that somehow learns  

what I want to do right so how that improves  itself with experience and so on so how do I  

choose this model and how do I actually find  the parameters of the model that gives me the  

right answer right so this is what we will spend  most of our time on in this course and then there  

are a whole bunch of other things that you really  have to answer to be able to build useful machine  

useful data analytics or data mining solutions  questions like do I have enough data do I have  

enough experience to say that my model is good  right it's the data is is of sufficient quality  

could be errors in the data right suppose I have  medical data and a is recorded as 225 so what does  

that mean it could be 225 days in which case it's  a reasonable number it could be twenty two point  

five years again it's a reasonable number or  twenty two point five months is reasonable but  

it is 225 years it's not a reasonable number so  there's something wrong in the data right so how  

do you handle these things or noise in images  right or missing values so I'll talk briefly  

about handling missing values later in the course  but this is I say mentioned in the beginning is a  

machine learning course right and this is not it's  not primarily it's it's primarily concerned about  

the algorithms of machine learning and the and  the math and the intuition behind those and not  

necessarily about the questions of building  a practical systems based on this so I will  

be talking about many of these issues during the  course but just that I want to reiterate that'll  

not be the focus and so the next challenge  I have listed here is how confident can I be  

of the results and about that I certainly will  talk a little bit because the whole premise of  

reporting machine learning results depends on how  confident you can be of the results right and the  

last question am i describing the data correctly  okay so that's a very very domain dependent and  

and and the question that you can answer only  with your experience as a machine learning or  

a data data scientist professional with time  right so but there are typical questions that  

you would like to ask that are there on the  slides so from the neck in the next module  

we look at the different learning paradigms  in in slightly more detail [Music] [Applause]