0:02 Hello, am I audible? Hello? Hello, hello, am I audible? Yeah, you are. Yes, ma'am. Okay, thank you so much. Actually, the Gmeet don't allow 100 participants, so I got disconnected and was not able to connect. Well, first of all, welcome you all. Very good afternoon for this class for introduction to machine learning instructed by professor Balaraman sir from IIT Madras. So myself Shreya Bansal, I am a PhD scholar from IIT Ropar. Currently I am in my fourth year of my PhD and I am working on multi-modal data analysis as my thesis work. So this is all about
0:52 me and I'll be with you in this semester for these lectures i think someone's mic is already on please try to mute yourself okay fine so uh so this meeting is open to you all and you can whenever open your mic and ask questions in between also it's not an issue uh but i would rather uh you know suggest you to ask at the last so that everyone can have the full session but if If some question is related to a particular slide, you can obviously on your mic and ask your questions. But it's a request that please put your mics off because it will
1:38 be a disturbance to this meeting. All right. So if you have any question, then we can start with your questions. Then otherwise, I'll start. Hello, I have a doubt actually. So please raise your hand. Yeah, yeah. Talking. Yes, yes, yes, yes. Yeah, yeah, yeah, yeah. So actually I have a doubt regarding this session like we will get a recording of this session or we have to attend live this session. Okay, so let me get these things clear to you first. So you all have joined the NPTEL portal for this session. So there you will find the course content. in the form of videos as
2:24 well as some assignments are there on that portal. Okay. So you need to complete those assignments there. You will be marked according to that assignments. Secondly, these sessions are for your help. Okay. So the major content of the course is already given to you in the form of videos by Balaraman sir. Those are your actual videos. And here these sessions I have created in such a way so that Some of the contents which may be high level for you, we can discuss it here. Okay. So I generally cover majority of the topics that are already covered by Baleraman sir. Just that in somehow in a very naive fashion here in this class. Secondly, these lectures, these video lectures will be there on the YouTube.
3:14 One sheet will be shared to you on your portal itself in the live session. Maybe the heading is called live sessions. There one Excel sheet is floated. will be floated actually in that every week you will have a YouTube link and one drive link is shared to you for my PPT or whatever content I am going to use I will be putting in that drive so one drive will be shared to you and every week YouTube link will be shared to you okay alright so these are weekly yeah it's alright alright Yes, this will be the weekly session, Saturday 3 to 5. Okay, so you might. Yes. Will you be
4:03 sending the recurring invite? This is the recurring. This link is recurring. So same invite you will get on your emails as a reminder every week. So what if I am 101st person to join in? This is actually the... problem of the platform. So I would suggest please join early. Madam, same link will apply every week? Yes, same link will be applied with every week. For this meeting. For this meeting, yeah. Same time. Same time, 3 to 5 Saturday. Okay, so I'll share my email. Yes, I'll share my email ID also. So
4:52 generally, because I have taken this subject two times in the last year, means I have taken this subject two times. And what we generally do is I share my email IDs. And when you have any issues regarding any problem or any question, you can email me. Okay, but generally it happens that if I have not replied you in a week, please let me know in this meeting. Sometimes your email are in this in my spam folder. So I'm going to share my email ID as well. All right.
5:42 So can we start our session now? Any more doubts if you are having? Okay, alright. So this session is the initial or introduction to machine learning. So first let me know about yourself that can I know from which background you all are? So it will be helpful for me to give a start of the course according to the audience that I am having. So which courses you are taking or, you know, if I can get an idea in the chat box that from which field you are working. So I may use those examples that can connect you well with this machine learning course. I am a
6:28 business analyst. So if possible, please write in the chat so that I can have an idea and everyone can have it. Okay, CSE department, CSE, AIML, great. Pharma strategy, data science, CSE, civil engineering, MCA, CSE, CSE, banking. So why I'm reading this to all, because you should also know that our audience is having not only CSE members, but they are having different backgrounds, including banking, one from civil engineering also, this computer science, ISE, MCA, IT, AIML. Majority you are having from a technical background, but maybe some
7:17 of you are having some business analysis. Okay, political science, good. Okay, did BSD. Okay, all right, fine. So I'm trying to keep it in a way so that it can be understood by everyone. But still, if my examples or any term is, if it is not good, clear to you then you can ask me anytime. Alright. So let me start with my PPT. Alright, Kashish, yes. Good evening, mom. Good evening. Ma'am, actually I have a doubt. Ma'am, in the final paper, how it
8:06 will come? Like MCQs are written? Generally, what I know is we are not, you know, part of the team for your final papers. But what I know from the past year is that you will going to have MCQ type questions. And content should be covered by your lectures given by Balaramans. Okay. And how much marks will it contain? So I think it's on your platform. I don't know much about that because we as a TA have just the duty to solve your queries. So we don't participate in those examinations or so. So we don't have any idea that how much weightage is in your marks. That is on your portal. You can check there. But yes, what I
8:54 know is that some of the marks are included from your assignments as well. So do complete all the assignments. Thank you, ma'am. Excuse me, ma'am. Ma'am, whatever the problems you are solving now, it's same type of, I mean, that will be come for the final exam. we don't have final exam papers I practice I mean week zero zero assign assignment there are 10 minute questions were posted month I mean that have been questions but all are different I mean it is purely related to the mathematics basics only like probability and linear algebra I need are you going to all man that one only now okay so look Let me
9:45 clear this thing also. So till now, I think you have an access to week zero. Week zero is not about machine learning, but it's the prerequisite to this course. Prerequisite in the sense, I would say that whatever you are learning as a CSE subject, it is quietly related to mathematics, right? And this machine learning course is also related to mathematical mathematics. terms so to involve you with those terms sir has in uploaded some videos regarding week zero in which generally I think he has covered two or three topics first is linear algebra second is probability and maybe optimization have been covered I don't know about week zero so these are your prerequisites these are not from for the ML course so
10:33 generally what you will get in your exam is for ML Okay, please switch off your mic because maybe it will be not cleared. My voice will not be clear. So, week 0 is for your practice to know that these terms and these definitions are important. You can have an idea from other books as well for those content. But we are starting from week 1. Week 0 is just a prerequisite. I have not gone through what Questions you have in your session right now because those are your, you know, assignments that you need to solve. So I'm not here for solving those questions which on which your marks are given. We can solve the last year problems, right? If you can get that. So those are
11:22 the questions for your practice. If those questions are for you to evaluate, you need to solve them. Yes, I can help you. I can give you a hint. But those questions are for you people. I am not going to solve this year questions. I have some questions before this year. So you can have an idea that how to solve those questions. Got it? I am a scientist from Big One. So whatever we are going to learn from your lecture, so same type of pattern we can expect for the exam related to the problem solving. Most probably because week one week one week two also it's already posted assignments. So week one and week two whatever they posted. So that is repeated questions from
12:08 week zero. Okay. I have one. It's okay if question is repeated, you need to just solve them. I have not gone through the questions what is given to you this year. Okay. So if you have some repeated question, maybe week zero is not on the evaluation basis. It will be just for your practice. I don't know about that. Okay. But if you have any question, you can obviously just pin me that this question is having you, you are having some problem. I can go through that question as well. All right. Okay, ma'am. Thank you. Okay. So, let's start with this week one, which is itself about the introduction to machine learning. Okay. So, here in this lecture, we will be seeing that what… is machine learning.
12:53 So there are some types of machine learning which many of you may have heard about at least supervised, unsupervised and reinforcement. These are the three basic types. And finally, the important one among these three which we generally start with is supervised. Okay. And in the supervised study, there are two types regression and classification. We will go through this. And finally, bias and variance. So these terms may be Like some of you may be familiar with it. Some of you maybe just have heard about that. So don't worry, we'll go through all these things. All right. So first of all, let's start with what is machine learning. So as the name suggests, machine and learning. So before this, I generally start with this history that...
13:41 What is artificial intelligence? Right. So we have this mathematics as a course from centuries back. Right. We started CSE from decades back. But what is this term AI? Why this is very boom in this time? Because just think about this. When mathematics started, it said that this is all about reasoning. Right. What is happening around us? It gives you the reason why. Now, when we switch towards machines, especially these high computed computers, we started to learn about computer science or you can say also include this information technology also. So, what is the major, you know, the
14:28 push towards these type of machine is that when we as a human was not able to solve big problems, Then we start looking out the solutions so that it can help the human. Right. So that's why big computers, supercomputers are there around. But the problem was that at that time the main issue is to solve the problem or accurately or in less time. Right. After that we saw that we are completely depending on the machines. Machines are doing too much of our works of our daily life. But do you think that machine and human are same or can we say that machine can act like as a human? That's why this study comes artificial intelligence, right? So this artificial
15:15 intelligence as you may know from the term itself, artificial means machine and intelligence means the capability to understand and take decision. This intelligence capability is given by God to us and God, what God has made cannot be replicated obviously. So what we are doing is we are just simulating what the understanding of learning or intelligence into the machine, right? So, once a scientist asks this question that what is intelligence to a human? Why we are saying that human is so much intelligent? So, one pattern come into the mind that It is because the human is capable to learn, right? So just think about this. When we were a kid, we were a small kid, our parents, our teachers, let us
16:02 too see and observe many things. So at first when you make a child see a chair, a table, okay? So those type of furniture you make the child see or you make the child see a cat, a dog. So he has some capability in his mind to capture the patterns, right? Somehow slowly, slowly that child is able to distinguish between different animals, different sounds, different mechanisms, different machines, different furnitures, right? So, this is how he or she is learning day by day. So, this is this quality of learning was very important and that's why they started this course that is called machine learning. So, to make a machine intelligent, let's have an idea how it can learn. Because learning is
16:53 quite a dynamic thing, right? So for example, let's say when you are coming in this class too, you are also trying to learn what the subject is, right? And it's not about like that you are coming it, you are hearing some of the terms, you say, okay, I learned this term, what is the definition of machine learning and that's it. No, you want to apply this definition in your subjects. Right. So this is the ultimate goal of learning. What we learn from our teachers or parents as a human that book can give you the knowledge, but you need to apply it into day to day life. So this is how this concept we are including it in the machine learning too. OK. So whenever whatever you are going to learn here, it's just relate how you learn as well. So, according to the definition of
17:42 Tom Michel in 1997, they say that an agent is said to learn from experience with respect to some class of task and performance measure. If it is performance, if its performance on task in T is measured by P and improves with experience E. So, this definition looks very, you know, robotic. But let me tell you one thing that whoever are here from, especially from these CSE and... Hello? Anything? Any doubt? Okay. So, whoever is here for... have the background of some technical student or a professor, I would suggest that take this subject
18:30 to learn how to write also. Okay. So sometimes what we see here as our day-to-day life, we know many things, but we don't know how to be presented. Right. So in mathematics or in computer science, if you are writing something, it's need to be crisp and clear. So that's why this definition may look very robotic to you, why this these complicated terms are there in the definition we can know that what is machine learning it's just that learning from some experience it's very easy to say but no mathematics say you had to write that what is assumed what are your constraints what is the input how the output is going to be so these three four things are there to be in the definition so just
19:15 think about this here They have written in a very systematic way. This is called systematic writing that an agent is set to learn. So that's why this term is agent is there. This agent can be a machine, can be a robot, can be a system, can be a CPU. So that agent is set to learn. It means now what the learning definition is there is after this. from an experience. So you always learn from an experience. You had some experience, now you get some insights out of it. So let's say if you have seen a lot of animals around you, okay, and you are able to distinguish, okay, this is cat, this is cow, this is peacock. This is what you learn by seeing that
20:03 picture or what is in the environment, right? So that comes from the experience. Okay. And then for experience with respect to some class of tasks. So let's say when I give you this example that let's say you have seen some a bunch of animals. Then you are able to distinguish between different types of animals. This is first task. Second task I can say to you can you count them. This is second task. Third task can you distinguish the animals with you know which are having feathers or not. which are having blue color or not which are having yellow color or not which is which are having maybe a particular type of characteristics or not right so these are the different tasks so it means that whenever you are going to use this machine learning in any of your
20:53 application you need to define the task very clearly okay and third term is performance measure obviously How do you know that what you are learning is good or not? So just think of it if you are making a child to learn, you ask that child many times, okay, what is this? Then the child say it is a cow. If it is cow, you say that, okay, good, very good. If it is not a cow, you correct them, right? So there is some measure to evaluate the performance. So in day-to-day life, Because you are conversing with the child, it's easy to give them feedback. But for the machine learning, for a machine to learn, you need to evaluate the performance in terms of a particular matrix.
21:40 Another example of performance matrix is your marks in the class. So let's say you are having, we are having 100 students here, right? Let's say I give you a set of questions, you complete the questions and I ask you how many questions, means how many marks a particular student get? So I can have an average idea that this class may be, let's say your average class marks is 70%. So I have an idea that this class knows 70% about this course. So this is how you define some measure. Those measures are dependent on the task. For example, let's say I ask you to distinguish between cat and dog. So you will say this is cat, this is dog. So the performance will be different whether it is correct or not. If I ask you that please count the animal in this box. Okay, or
22:26 in this area. So, the performance measure may be different because now counting is the final outcome. So, performance is related to the task and task can be defined on an experience. So, in our terms, we generally use experience or data in the simultaneous way, right? So, here we can say that the ultimate, you know, the god of this machine learning is data. So whenever you see any data, data can be audio, data can be video, data can be text, data can be in any format. You see on the internet, around you, it's all data. So whatever you are experiencing from the data for a particular set of tasks and by defined performance measure, it is called learning. Okay. So
23:16 if that's why it is written here as if it's performance on task in T, T are the task, performance is P and if your performance improves with the experience then you say that you are learning. So let's say you go in class 1, you understand some type of grammar, literature from English and you get some maybe 70-80% marks. In the next class when you go, you get some previous lectures, you get some new lectures, your marks increase by 82 or 85%. Then you can say that yes, I am learning this subject. Right. So in the similar way, we wanted machine to learn for a particular task on a particular data and evaluate them with the performance measure. So these three terms are very important. Whenever
24:06 you are going to use this machine learning, first define data, second define task, objective, third how will you evaluate? If you have three things in your mind, you are set to go. Other things are more technical around it. Okay. So this is what is machine learning is. Now, obviously, why we define task, performance and experience? Obviously, without clear task and performance measures, learning becomes vague and unmeasurable. So, let's say if I give you some data and say, do whatever you want. You will say, ma'am, just let us have some problem to solve. Then we can go by it. If I say, these are the numbers given to you, do whatever you want. This is not a task. If I say, okay, these numbers are there, please sort it in an ascending order. Then you
24:55 can say, yes, I can do that. Okay. So that's why you need to define a clear task and their performance measures as well. So for example, something seems to improve is unclear because, you know, we need very clear things which are quantitative in nature. We go for quality as well for some of the tasks. But generally, whenever you are using computer science and machine learning, it's all in mathematical form. So, something will be quantitative. You need to quantify. Either you are approaching towards maximization. For example, I need to, you know, maximize the profit or minimize the loss. So, ultimately, either you are working on profit or you are working on loss. So, ultimately, something needs to be quantified. So, examples of clear performance
25:44 measure is exam marks for test checking. So, if I define your marks, okay. So, let's say I have given you 10 questions of true and false. So, if you write correct answer, then I will give you one marks. If you write incorrect answer, I will give you zero marks. Then I can get that whoever person scores the maximum marks has the best performance. So, I have a criteria that maximum, I am going towards the maxima of the marks. for this test. So it is now clear to you that I need to go towards Maxima of the marks. If there are 10 questions, I need to guide by 10 by 10, 100% accuracy. So think of learning as a training for a marathon. For example, they have written that the task is clear for a marathon. The task is running. The performance measure is time
26:33 and experience is how many times you have practiced the run. So when you have practiced so many times the run now on the final day of the marathon, when you are running, you need to reduce your time at the finishing line. So this is very clear in your head. If it is not clear, if I say, okay, just go and run. You will ask me now what is the task? Do I need to run the, you know, minimize the time or do I need to have another task in between the running? What is the criteria that I can win this marathon? So similarly, here also in learning also, you need to have an objective. That objective is clearly defined. Over some performance measures, that performance measure, either it is in the minima case or maxima case. Minima means here
27:22 you can see that minimum time. Maxima, it means you can see here the example of test marks. So you need to get the maximum marks. Now, this thing is more like, okay, we have some idea that we can make machine learn over some objectives, defining some parameters. These performance matrices. That's fine. But how many types? How can we make learn the machine? So some scientists say that. Why do we have to think about it? We take the concept of learning from human. So let's see how human learns. So the very first approach the human learn is supervised approach. Okay.
28:07 So supervised means you are coming here. to learn this course and there is a supervision given to you in the form of instructor. Okay. So whenever, whatever you are learning here, it is provided by the supervisor. The supervisor is telling you, this is the question and this is the answer. All right. So basically in this type of learning, you need to come to the class. Some supervisor is giving to you, okay, this is the question, this is the answer. Now you are training your mind over that. Okay. If this type of question comes, I need to solve like this. If this type of question come, I need to solve like this. Okay, that's very easy to go. Next, what you have in this type of classroom
28:52 teaching, after some time, you need to give a test. Okay, in that test, some of the question may be similar what you have learnt in the class. Some of the question may be applicative. Right. The supervisor is very smart. OK. They will not test you on the question. What is already given to your while you're training in the classes. But some new questions will be there. Right. So maybe you are able to solve those questions. If you are able to solve, then you can say you have learned the concept. If you are not able to solve, it means you are just grasping what is going in the class. So what type of learning we generally say it's good that you are able to apply, right? So here the important thing is you know the answers in the form of supervisors or supervisors know the answer.
29:41 So he is there to supervise you, to correct you and give you a feedback that no it is not correct or yes it is correct, okay? So this type of learning is also called supervised type of learning. So here the example is The one picture is given and the network has to predict whether it's a cat or a dog. So while training, the network know it's a cat. The network know it's a dog. So what generally it happens is, let's say, just assume like that. There are 100 pictures of cat and dog. Okay. So what is given here? First picture is shown through this network. And that because... The network don't know anything about this. So network will say randomly it's a cat. So now the
30:30 supervisor is there. They tell the network that no it's not a cat it's a dog. Then the network say okay I got your point that this is not a cat this is a dog. Next picture shown to the network. Again it will predict something. Then supervisor will say either yes or no. Again and again. Till the hundred picture, the network, network here you can understand it as a human mind as well. So the human mind is now capable to match the patterns that what the supervisor is telling that this is a cat, no, this is a dog, that how the cat look like, how the dog look like. It means that how this is different from human learning, we as a human is telling you that look, these are
31:16 the four, look at these four legs of the cat, look at the head of the cat and dog, now differentiate. No, we are not doing this. We are just telling the network that whether it's a cat, yes or no. So this is how you teach a child, right? Look, this is a cat. Other time when he goes to the market, the mother asks the child, what is this? This is the dog. No, no, no. This is cat. Now this is how the child is able to learn, right? So this is called supervision. In other sense, wherever the input and output means the data and the label, for example, the cat, dog is given to you, it is called supervised. So that's how you train the network using the output. Now, another question arises is because this is
32:04 a machine going to do that and it's all in mathematical term, there are two types of supervised learning on the basis of output. Okay. How? First of all, let's say I am asking you two questions. Okay. Just listen carefully. First is I am saying what will be the temperature of tomorrow? So what will be the temperature of tomorrow? The answer will be in a continuous number. For example, it can be 12 degree. It can be 12.5 degree. It can be 12.59 degree. It can be 13 degree. Okay. So it's a continuous number. Next question I ask you. How the weather will be tomorrow? So is it be cloudy? Is it be sunny? It will be rainy? How will be the weather? So these are the three classes that
32:53 I ask that I want to ask. Is it cloudy, rainy or sunny? Okay. So these are three classes. If your output is a label or a class, for example, is it a cat or a dog or a flower or a tree, then it comes in the classification category. So first type of supervised learning is this classification. Another type which I ask you what is the age of this person? What will be the temperature of tomorrow? Okay. How, what will be the stock price, marketing price rate? Right. This will comes under a number. It a continuous number. It means if I am saying tomorrow will be 12.3 degree temperature, it will be different from 12.5 degree temperature. Okay, so that's called regression. Regression and
33:42 classification are two types of paradigms where supervision applies. It means for classification and for regression, you need to have true labels or the actual truth. Okay, according to the data. Generally, regression is first used if you have gone through your BTEC third version second, third year, regressions are there in the syllabus for statistics. So statistical terms, if you have read some statistics. Ma'am, I have one doubt, ma'am. Yes, yes. Go ahead. What is the label, ma'am? Label means your final output. For example, if I show you a picture and ask you what type of animal it is and let's say the picture is of a cat, then cat is a label. Got
34:30 it? Oh, okay, ma'am. Okay. Thank you, ma'am. Associated with the input, the output is called label. It can be called as ground truth. Okay. So, any term can be used. Generally, we call it as ground truth and label we associate with classification generally. So, whatever you want to predict is output, label, this form. And it is present in supervised study. Okay. So, these are the types where supervision is very important. Now question arises is, look, someone asked, is the supervision always applied? So there are a lot of data, there is a lot of data on the internet where this label or ground truth is
35:16 not there, right? Just think about this. Think about you capture a image from an internet and then who will tell you that this is cat and dog? You don't know. There are many images on the internet, right? always human will not be there to annotate it. We generally use this term annotate whenever we want to find the true label. So, let's say if you, let's say I give you a task that collect some, you know, wherever you are living, you have different types of trees there. So, what you do, collect some pictures of leaves, okay, of different plants or trees. You collect the images in a camera. Now, I say, let's do this task that for each image create a sheet with the name of the image
36:02 and what plant that leaf belong to so you as a human because you have collected the data you give some time and write the name of the plant associated with each each image right if you don't know even i ask you to ask some expert and get an idea this is called annotation because you want your data to be supervised you want your data your labels should be there, right? But it is not present every time. Let's say I hired someone. I say, please collect images for me. That person has collected the images to me, but he has no idea which plant it is. So there comes the problem where the labels are not present. What to do? Then comes this unsupervised type of learning. So what
36:48 has been seen as a human perspective also that sometimes what happens is people Our parents are always there with us as a child to correct us. This is wrong. This is right. Look at this. But many of the things we don't need supervision. Right. Excuse me. Could you please repeat what is annotation? Okay. Annotation means if you have the data and don't have the label. The process of getting the label is called annotation. Okay. So let's say I ask you to collect some images of leaves around your area. So what you will do, you will collect the leaves and create an excel sheet. Okay, a sheet with the name of the image that you
37:36 captured. Let's say your image is one dot jpg. What plant does it belong? So this is how you are mentioning it, writing it somewhere, storing it somewhere that these are the leaves from this plant. These are the leaves from this plant. This is called annotation. We can say it as labeling the data, right? Yes, you can say that. I generally would say when you label the data, there are many other tasks for segmentation, for detection. So in complete form, we call this labeling the data as annotation. So unsupervised we were talking of. So just think of if I a very small child right. If I give that child some toys of different color. What we
38:25 generally observe that that child make a complete classification based on color. So color is the first thing that we human are capable to capture. So what can the child do is the child could capture. make all the red colored toys in one row or the blue colored toys in another row or on the basis of texture or on the basis of some quite you know the size or the shape it is able the child is able to distinguish between the things okay this does not come on the basis of any supervision no supervision is required for that so here comes this next type of learning which is called unsupervised In other way, we can say that this unsupervised is grouping
39:14 the similar characteristics in the data. So when data is given to you, you find some similar and dissimilar characteristics and according to that similarity, you group the data in clusters, you can say. So that's why one form of unsupervised learning is clustering. So as you can see from the diagram as well, there is there are three four types of emojis okay so on the basis of the distance they have created three clusters of three type of emojis but look here in the first cluster red there is a emoji which should belong to the second cluster right the emoji of sadness right so in data science or in
40:00 in generally what we what we assume that the similar type of data should lie nearby to each other. But it does not, it is not a strict, you know, assumption. It may be that some of the data lie out of that region. It can apply because exceptions are always allowed. But generally, because we don't have any supervision, we don't know the exact label, what we can do is we can apply some similar type of characteristics, go run through the data and find out if this is okay, the group. So one example that I generally give here is, let's say you have a task to, you know, let's say I have given you some
40:48 documents and I tell you that please group these documents into different clusters or based on the similarity of the subjects, group them into some groups so that I can easily evaluate them or read them. So let's say there are the documents related to finance, some documents are related to you know this technology, some documents are related to entertainment. If these three are the categories, it will be easy for you because those are completely different, right? But let's say I have some more detail in the technical domain, there are documents from engineering from civil engineering from computer science engineering from electrical engineering now I didn't
41:35 tell you that these clusters are need to be maintained means these domains are need to be picked to maybe some of you will say okay let's have three main domains of entertainment technology and this finance some of you will say no no this electrical engineering engineering is completely different from civil let's let's separate them all together too Right. Maybe some document is having the like prediction of stock markets. So there the prediction means it is from CSE also and the stock market means finance also. It can be merged in the two documents. Then there is a there is a conflict there. Right. So this is how because we don't know the actual truth. We don't know the actual label. We are just using some optimal way ki okay
42:21 we can make it cluster in this form or in this form. Okay. So, in this way, we are going to learn how this mechanism work, but this is called unsupervised way. So, because I told you about grouping mechanism, that is under the clustering algorithm. So, as we have classification regression in supervised, we have clustering algorithm in unsupervised. Okay. Next is written as association. So association also don't need rules. So what is happened that generally it is seen that from a market basket, let's say we set up a system at a mall where this
43:07 daily use things are there. Okay, daily mart actually not a mall. So I get an idea that in a basket how many items are there and which items are there. Okay, so what they have done is they have created a study where they study about each and every basket of individuals of the person who is coming for shopping. So what they have seen that bread and butter are generally associated with each other. If someone has purchased a bread he also has purchased either a bread or a jam or some you know topping over it. Similarly if someone has purchased a mobile phone He has gone for some other accessories like headphones or maybe some mobile cover. So this is how this association thing comes into picture.
43:54 That some things are not being clustered but can be associated with each other. So what they have used this market basket is for you know having a price of the commodities. So what they generally do is if they increase the price of bread they decrease the price of butter and they advertise it Look, the butter is at this percent discount. So, what people will see? Okay, discount, discount, discount. But how will you eat butter? Without bread? No, obviously. So, you need to take the bread also. So, this is how these conscious decisions are based on this association level. And this association may be defined by looking at some data, but it is not in a supervised manner. It's not compulsory if someone is purchasing a bread, he or
44:41 she will obviously purchase butter? No. Maybe he is purchasing a bread to make a sandwich. So he can club the bread with potatoes as well. Okay. So it is not compulsory. It is not rule. It is not supervised. But yes, it has some association with some commodities. So this is how in this unsupervised way, we have another type of algorithm called association. Most probably, we will not going to discuss association in these lectures. We will go towards different type of clustering as well. Third type is reinforcement. So this supervised, unsupervised is somehow very mechanical in nature. Why? Because as I told you the supervised, it's more like you sit with a child, make it with one, two, three images and then
45:30 make the child learn that no, you are not correct. This is the correct form. No, you are not correct. This is the correct form. Yes, you are correct. Then the child is learning. In this way, human don't work in day-to-day life. That's why a third type of learning comes into picture that is reinforcement learning. So here you can see a robot is learning to play, right? So all the board games starting with the chess and the go player. These are the first games from where these reinforcement learning comes into the market. Okay. And at present, when we deal about this machine learning paradigms, reinforcement learning is quite different from other learning. That's why we keep it as a separate subject as well. So you may find that reinforcement learning is a separate course. So we are not going to discuss reinforcement learning in a big
46:18 detail. Maybe one lecture for that. But supervised and unsupervised are the very basic form of machine learning. So we are going to discuss about these two. So let us understand that what is reinforcement? Reinforcement, reinforcement meaning is feedback. So generally we have feedback and rewards. So how you train a dog? Let's say, for example, so when you have a dog and you want to play with, you know, disc with it, so you flip a disc and you make the dog run to take the disc to you, right? If he do that, you pat his head and say, oh, very nice, good, you have done a good job. These words make animal learn that okay this is a good feedback to me, this is
47:03 a reward to me or you give some biscuits or give something to eat, right? So this is how he says that this is something that I am doing good. Or in the human form as well when we go to a class, when we were in the preschool or nurseries, When we do something, our teacher say, oh, very good. You have done a good job. This is the prize. This is the award for this task. We say that, okay, now we got a star. We will write, our handwriting is good today. Our teacher gives, have given this reward to us. We do those actions again and again. We repeat those actions. Okay. Another thing is, that is called punishments. So the opposite of reward is punishment. Punishment means, let's say your dog has done, you know, something. maybe he tried that
47:50 animal has tries to bite some of your relative. So what you do? You scold that dog. No, don't do that. If you have done anything wrong, your parents say this is what you should not do. They scold you. Now you get to learn that these things I should not repeat in my behavior. So this is how you are getting your feedbacks after taking the action. Okay. So in this way, these robots are trained. So for each and every action, there is a reward associated. If reward is good, they try to mimic it more. If the reward is bad, they try to avoid it. In this way, when many of the games like chess, go player, go, okay, Chinese chess, actually, these games are based on this
48:36 reinforcement. So what they are doing is on the board games, because they have started from the board games. If any of the you know state if they have used any of the state to continue and that lead to lose that lead to make them lose the game it means they will not use that move again if some move make them to help to win the game it's in their memory if we did that okay this helps me to win the game i will choose it more often so this is how after making play these robots the games many many many times they are able to grasp that what will be the next move this this like a human mind does okay it's very automatic in our
49:21 mind but if all of you just sit and just think about how human mind think we are also you know thinking very unconsciously so we are not able to think that why have we have taken this action just a just a simple action if i say whenever you are going near a fire First of all, you will not go at this stage. But if something is like very hot, you somehow do the reflex action and your hand moves out of it. Right. It's not from the first day. Right. So, so these actions you trained yourself. Sometimes you are able to predict, okay, this, this, this may be a law, law, law. lot more amount of heat from this device, I should stay behind, right? So these are the prediction that human mind is making after making a lot of experiences. This is how what we
50:11 learn in this reinforcement type of learning. It's a separate complete different types of algorithms there. So these are the basic three types of approaches in machine learning. In this whole session, of 11 weeks or 12 weeks, we will be keep focusing on supervised and unsupervised. First, we will start with supervised because these are the basic. Then we will go for unsupervised, especially for clustering and one lecture for reinforcement that how this works. Okay. So these are the three basic type of learning. So when we have started this definition of machine learning, we have three terms. First is your experience. It means data. So I get to know that these three paradigms are there. So data will be different. Data with output is in supervised. Data without output is in
50:58 unsupervised. Data with feedback is in reinforcement. That's fine. Second is task. So we define the task as classification, regression, clustering, association. Right. So that's a basic task. Now third important thing. Performance measure. So in supervised, unsupervised and reinforcement, because supervised means you know the actual label. So how far your machine is from the actual value is your performance measure for supervised study. So for example, let's say you created a machine which can predict how will be the weather tomorrow. Okay. And when tomorrow will come, I can check that, that whether it is true
51:45 or false. So let's say I keep testing your machine till 15 days or one month. So I will have an idea that how many times this machine has given me false alarms. How many times this machine has given me good answers. So this is how you can get an idea that from these 30 days, how much the machine is accurate. Okay. Similarly, if I am saying about, uh, What will be the temperature tomorrow? Because you know accuracy is good to go when you have three classes. Is it cloudy, sunny or rainy? So let's say your machine predict tomorrow will be a sunny day. And we get to know that tomorrow the rain has happened. So it means that this machine has given for day one wrong output.
52:32 Means zero. Next day maybe it has given the correct output. It means one. So how many ones you have got from in 30 days this will be your accuracy, simple accuracy. But this cannot be used for the task if I ask that what will be the temperature tomorrow. Why? Let's say your machine is saying tomorrow you will have a temperature of about 12 degree. When we reach tomorrow we found that the temperature is around 15 degree. So I will not say the machine is incorrect. I will say the machine is degree incorrect right so this is how you take the task then define the performance measure so here we can say that the error can be the classification error or the prediction error okay
53:19 so this error it means how many times the machine is inaccurate or how much far the value is from the true label this is from this can be evaluated in a supervised learning Because we have the label, we can label it. But what we have, we don't have the label. For example, I have given you the document. You cluster it and you say, ma'am, this is... One person come to me and say, ma'am, I have made three clusters. Another came say, I have made five clusters. No one is wrong. How can I say that who is correct and who is not? So here we talk about similarity because similarity is a very vague term in mathematics. We call it as probability. Yes. Mom, I have one doubt.
54:05 Yes. Mom, whatever the example you have explained regarding rainfall, I mean that value 12 or 15, that is related to the regression, example of the regression. So, will you explain the example for classification, Mom? Classification just treated as three labels, cloudy, sunny, and rainy. Three classes are there. I want to know whether tomorrow will be clouded, Tomorrow will be rain or tomorrow will be sunny environment. Let's say your system say tomorrow will be rainy. But when I will go to tomorrow, I found out there is no rain. There is pure good sunny day. So it means your system has given the wrong feedback. Wrong prediction, I should
54:52 say. Not feedback, wrong prediction. So it is wrong. Okay, that type of error is classification error. That type of error is classification, right. So, whatever the value you said, no ma'am, like 12 and 13, that is a prediction error. Yes, this is for regression. Okay, prediction error indicates regression and classification error will indicate classification. Yes, yes. Okay, ma'am, thank you. Okay. Please mute yourself. Kashish has raised the hand. Any problem? Yes ma'am. I'm in the supervised learning. Ma'am, we have KNN logistic regression. Do we have to study these all decision
55:40 trees? Yes. Actually, we will be studying different type of algorithms in the next classes maybe. But yes, we are going to see logistic regression decision tree. and these type of algorithms in the upcoming lectures. OK, one time. Hi, Shreya. I'm sorry, I have a question if I can. Yeah, yeah, sure. I wanted to understand, for example, you have a session now from 3 to 5. For people who have also enrolled in other NPTEL sessions like Python and data science, it's an overlapping session from 4 to 6. Is this recording going to be available later on or how
56:29 would it work because there is an overlap of time? Yes, there is a recording. All the NPTEL, these type of sessions in NPTEL, they have the recording in their YouTube platform. Sheet will be shared to you for each course in which this link will be provided for each week. I mean the session that you're covering now, would this be recorded? Yes, yes, this will be recorded. And after this session itself, I will upload it on my YouTube and share the link in a sheet. So once sheet is shared to you, maybe after this week, means after tomorrow, today, and in which these lecture links will be there. So
57:14 each week there will be a new link will be updated. OK. So actually, in our login problem solving session 2020-26 is there, ma'am. There they have given option, like problem solving session recording. When we click on that, it is showing one Excel sheet. In that, I think your ID, your name is there. So maybe their YouTube link is there. So maybe that link will be updated there. So every person can able to see their login. Yes, yes. That sheet I am mentioning that sheet because after this lecture, I will upload the video and generate the link and we'll update the sheet.
58:04 Please mute yourself if you don't have any doubts. Ma'am, in the WhatsApp group, the link that is shared in the email call message, we have to join that group or yes or no. Look, I am not creating any WhatsApp group. And if you people need among yourself, you can go for it. This group is not shared officially. Okay. Can I speak, ma'am? I have a small issue. Yeah, yeah, sure. See, you have given the link for the registration for the examination. But while we are trying to do that, it is not getting uploaded. It is asking for the course selection. And there is no course selection provision for that.
58:51 Sorry, I am not getting your query. I think it is regarding for your portal, right? Portal. Portal. you there was a link there was a link for you know exam registration for the april month examination that link is provided by you on that mail but when we try to do all the things that have been done even uploaded that id and photos everything and uploaded when you submit that you know it is coming as rarer saying that it is was not selected selected okay okay look uh whatever you are facing these problems these are from the these mails are coming from the portal and there is a team which is handling this it's okay whatever problem you are having it just click a screenshot and you have a query mail I think it is given to you support mail or query mail just
59:39 mail this query on that support they will reply you back okay okay And second thing that, you know, second week that your link has come, but again that first week's mail is again repeatedly coming. First week is repeatedly, that's why coming because this is your first session for the first week. Correct. So that's why they are making you reminders for that. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. So So let us continue where we have left. We have this unsupervised type of learning in which we don't have the correct label. So how can we say that this much error or this much thing is there? So we work on similarity. Similarity is a
1:00:25 very vague term. Okay. So it's very clear in a human head but in machine it's more like a quality rather than the quantity. So we work on probability. So what we found is this unsupervised type of learning comes for the data analysis. Okay. So whatever you are going to learn in the subject of data analysis is generally comes under this unsupervised approach. Why? Because I want to learn about data on which I have no labels, no information. So I want to take a raw data, analyze it in the form that how much information I can make out of it. Okay, so there are some performance measures, for example, spread, purity, support, confidence. These somehow sometimes use, you know, in the
1:01:15 similar meaning of each other. But what can I give you an intuition at this time? Because when you will learn about unsupervised, you will get more of it. Just think about this. A data is given to you. The data as I told you can be audios, can be videos, can be images, can be text, can be labels, can be anything, right? In mathematics, in computer science, any data is plotted over a region, okay, on axis. Whether it's a two dimensional axis, is a three dimensional plane, or 4 or 5 or many dimensional plane. That's another issue. Just think of it, a data on a two dimensional plane. Because it's very easy to understand on the human level. Right? You can make a graph with two X's,
1:02:00 X and Y. Put some points there that are called data. Okay? Let's move one dimension ahead. In three dimension, it means wherever you are sitting in your room, you are in a 3D. Right. So let's say there are some object floating in the air. It means you are in a 3D room. Those objects constitutes your data. Okay. Now this data what we understand that comes from our probability distribution. That's why the probability concept is very important which was kept in week 0 for you. So what is probabilities? Probability distribution is we generally say that for example. Let's say if I ask you to create a picture of a human, what you will say, I'm not a good artist ma'am, but
1:02:49 what I know is a human has one, you know, head, two eyes, one nose, two ears, two hands, two legs, one body structure. You know about this structure, right? If I say just create a lion, you know that how this lion can be look like. It may be different, but it has some common characteristics. You find some common characteristics among all lions, among all humans, right? That in the mathematical term, we call it as a distribution. So there are some parameters of the distribution that is learned on the basis of these characteristics. These characteristics of hands, of skin, of texture, We know it as a human because we have a lot more bigger understanding, right? But on the mathematical ground, we call it as
1:03:38 a distribution. So humans are coming from a distribution. Lions are coming from a distribution. Zebras are coming from a distribution. But just let me know one thing. The distribution of a human and a zebra will be closely linked or not? Yes or no? What do you think? distribution of a human and a zebra do you think it will be closely linked it will be similar no ma'am no but if I ask you the distribution of a zebra and a horse or a donkey can you find some similarity there yes ma'am yes ma'am yes so that's why that's is your distribution and
1:04:24 distribution is coming from a property which you don't know which we cannot define. Okay, that's why it is working on numbers. So if you are saying, so let me if I can able to give you this glimpse, just consider that if I am saying the distribution of zebra, this horse and a donkey has similar type of distribution, it means they're talking about the shape, right? But if I am saying you the distribution of a zebra and a human is quite similar, it means they are talking about a similar characteristic that they both shared. Maybe living. Maybe we are talking about only living things. Maybe we are talking about the animals. So this property is not explicitly defined using the distribution.
1:05:12 It's just that you have a data on the mathematical ground on the computational graph and you just find a distribution. So that distribution is important. So if in the data you can find some distribution, strong distributions, it means you are good to go. And if your clusters have no good distribution, it means you are actually having some errors. So how to use that, how to use those mathematically, we will be seeing it afterwards. Next is this reinforcement learning. Here the positive and negative feedbacks are there. So basically total number of penalties, what you are getting out of the system is the performance measure of this reinforcement learning. So we'll see maybe some type of reinforcement feedbacks or something in the last lecture
1:06:01 of these sessions. Okay. So after this, there are some concept of bias and variance, which I will cover through the questions. So let's solve some questions. Okay. So this is your question number one. That which of the following are unsupervised type of learning problems? Okay. So you have to give me this answer. That first is sorting a set of news article into four categories based on their titles. Do you think it's an unsupervised problem? No, ma'am. Because the titles can be categories are given to you. Okay. Okay. Next is forecasting a stock price of a company based on the historical data. It is also
1:06:50 supervised only because already the data is available input and output. So we can able to predict by taking the values of the stock price. Correct. Third is predicting the type of interaction positive or negative between a new drug and the what is written? This is unsupervised learning. This is a set of human proteins. Predicting the type of interaction between a new drug and a set of human proteins, positive or negative. Okay, so you have to... This is a classification. This is also a classification. Yes. And last is identify close-knit communities of people in a social network. Unsupervised learning. Yes, because you may not have
1:07:40 the idea that how to define this close-knit communities. Okay, so the people who are nearby to each other on the graph may belong to the close-knit communities. The people who are farther away may not be. Last is learning to generate artificial human faces using the faces of former facial recognition data set. Okay, supervised. And this is supervised. Supervised. learning to generate artificial human faces using the faces from a facial recognition data set. Because already faces are stored in the database domain. So data is already known. I mean label data. Is it correct man?
1:08:29 Look, if you say that this facial recognition data is there, But what type of you know the generation is not a supervised task. Because I am not saying to predict or classify whether the faces from the recognition data set or not. I am saying generating. It means the generated image will not be same as the data that is already there. Okay. Why I am saying it is unsupervised? Because It has no label. How can you say that if the generated face is from the facial recognition data set or not? How will you classify that? Ma'am, and one
1:09:18 more. Facebook. Actually, our profile photos will be there in the Facebook. So, that example will come under which type of learning, like identifying the faces from the Facebook. Identifying the faces from the Facebook. If your faces are already there, you have already a label dataset in which you are saying that this is my friend, then this is a supervised type of learning. Is there any chances that that example will become unsupervised? Like what type of... Unsupervised... Look, unsupervised come if you are not saying that this can be your friend or something. If you are saying that... maybe you are just classifying it into two classes
1:10:08 or two categories whether the image has humans or some background then it can be an unsupervised way but if you are particularly saying that this this is my friend this is this is my this friend let's say friend a friend b friend c it means you have already labeled it okay I mean, this E option, they have given faces from the facial recognition data set. So facial recognition data set is nothing but already on the face, they will give them all the attributes on the face. But what you are learning, you are learning to generate real human faces, right? So when your system will generate a face, it means it will generate an image. How will you say that this is from the facial recognition or not? How
1:10:55 can you say that? You have no label for that. Means that whatever answer is given from the system is not labeled. It is coming out of something which cannot be trusted. Right? You don't have label associated whether it is true or false. Okay. Right. If the question was if from the facial recognition data set, if I give some image and if the model is saying, yes, it is from facial recognition, this information I have from a set of image, this is present in this recognition data set or not. This is supervised, but it is now generating an image. It is not generating a prediction. It is generating a face. So let's say, for example, you have
1:11:42 a data set in which all your images are there and I ask a system to generate your images. I am not saying generate the same image that is on the facial recognition dataset. I am saying similar looking. Yes, ma'am. Anyway, now, JGPT is providing. Now, when we are posting our image, it is going to view another image by changing some of the features. Yes, that's why it is not supervised. It is unsupervised. Okay, ma'am. Thank you. Okay. So, now, next question is how much time is left? It's 30 minutes. We have time. We have time. We have started 15 minutes late. So we have around 40 minutes. So for unsupervised learning, the randomness is there. Sorry, can you please repeat what you are saying? If the randomness is there, that is unsupervised learning.
1:12:35 I should not say randomness. I should say no specific output. We are saying deterministic. that is super learning. You can't say that is randomness there. Then it is... I think what you are saying in another context, then I can say that yes, I can approve that this is called randomness, but because this word random has different meaning in machine learning, so that's why I cannot say that this randomness is there. What I am saying is, let's say you are making a system to learn some task. If the output of the task is there, so
1:13:22 for the output of the task there is a label on which you can say this is correct or not, then it is supervised. If you don't have a specific
…
