 Okay, going live. Okay, hello everyone. So welcome back on Sanjay Gupta Tech School. So as you all know, last year we started a bootcamp on AI associate certification. So we just want you to prepare for AI associate certification. So last year like we did two sessions on two different topics and this year we have restarted this bootcamp and like this month and next month, all the sessions we will be conducting so that you can learn everything about associate AI so that you can clear that certification, right? So I have Nikita with me. So welcome Nikita on the channel. And she will be delivering like one more topic today that is neural network. And before starting that, she will be giving you a quick recap what we did in the day one and day two session, right? So over to you Nikita. So just do a quick recap so that we can connect those previous sessions and then just start the session. Over to you. Sure. Thank you so much, sir. Okay, so we are here back again on the third session of artificial intelligence program which is strictly on the grounds of AI associate program that Salesforce has offered. So here we started discussing about some points which are called supervised learning, unsupervised learning. And then I've also taught you about some things called classification, regression and clustering. So let's have a quick recap of these three things. So first of all, I had taught you about supervised learning. Supervised learning means when you have got labeling in the output. There are two things, input and output and when you have labelled data in the output that's called supervised learning. That means you are netting the machine, know about the output and that's how the machine is having the work done. However, in clustering in the unsupervised learning what happens was I had given you an example. In that example we had mentioned that there are people who are the buyers, who are the frequent buyers and we and the company were giving credit cards to only those buyers who were very good in their civil score, who were very good at their loan repayment. So this is how the machine was clustering all the buyers and the machine was doing all this work all by itself. So that's how the algorithms of machine learning were working. So two things were regression and classification. In classification we had labels where I told you that if a student is passed or a student is failed that is based on the percentage criteria. So these two labels we granted them pass or fail so we classified students in the pass category or fail category. Similarly in regression supervised learning we had given an example where salary was discussed and there were people who had some salaries in some pointers. So when numerical data is the output and there is lot of numerical data that is the output we categorize it into the regression algorithms and the regression part of the supervised learning. So these were the three sections that we had a good discussion about. Then we went on to having a classification of regression also in simple linear regression multiple linear regression polynomial regression. But as Sir already mentioned that we are strictly onto the path of AI associate program. However, all these things with the mathematics will be discussed in the coming up syllabus and course that we would launch of data science and artificial intelligence where you will be learning statistics, mathematics all around for linear regression, multiple regression and polynomial regression. So all the math will be discussed. Here the math is not discussed because in associate program there is not too much of mathematics that is going to be there. So hence we pause that this particular sitting and we thought that we will initiate a different program for it which will be around more than six months to eight months where you will be delving into the mathematics then you will be also going through the Python language because one of the prominent language that we are going to use for the execution of different algorithms would be Python only. So first there'll be maths Python and then you will be accustomed to the different algorithms that we are going to discuss. Meanwhile, what you can do is we have a group on LinkedIn and there is where I have put lot of mathematics about these regression algorithms and you can just have a look if you are interested in learning the math behind the things. Otherwise, if you are only looking for an associate program then definitely you can just skip the maths as of now. But it would be better that you do a holistic development rather than just a pinpoint development. Then I discussed about simple linear regression and I had taught you about that there's a regression the line of regression that is on that path which is the most or you can say that the least error-fied path and this is our line of regression on the scatter plot looks like. So we plotted data points and a line of regression was plotted. We found out the different least square errors and then we minimized them and wherever the errors were minimized this is the line of regression that we found. On the y-axis it was dependent variable because these were only for the linear regression. So linear equations was executed. Similarly, as you can see in the right hand side these the connections between the dots and the lines show that these are the, this is the difference between the point that you got and the point that is predicted. So output that is predicted it has, it's different from the output that you've got. So this generates a little error and wherever it is less there is the line of regression which is executed. So then we went on to a little bit of theoretical aspect of simple linear regression. However, in the associate program there's not a glance of this linear regression. You just have to have a simple idea about it. So you can just read a little bit and then go for this program. Yes, just to add a little bit. Sorry to interrupt. Just to add one thing like guys if you want to understand this simple linear regression so you can watch the previous session that we conducted like day two. So there Nikita already explained everything about the simple linear regression. So you can go through that so that you will be having all the information. Okay, go ahead. So today we had for you was difference between the hand coded algorithm and train models. So we already had a discussion about these theoretical parts where I told you there's a difference between automation and a trained model. Automation depends on how we control a computer. We control an algorithm or we design something for a computer that this is the work that has to be done and automated. That's called the hand coded algorithms. However, what trained models would do is we would just give them the data and they would be trained on that data and then they will have a execution of whether or not the task is to be performed. So this will be a quicker way of execution and you can see that the trained models what all they can do is they can learn patterns, they can have relationship from the data. So they can derive a lot of things from the data that you provide. All you need to give them is just the data set and they will, the models will be capable enough to execute the things. Then you have got that the choice between these approaches depend on the factors such as the nature of the task. Of course, we are going to, not everything can be trained. So there are restrictions in that part as of now because AI is not fully developed. So there are little consequences and things that we have to keep in mind. Then you have the availability of data and the need for interpretability. Of course, if the data is not available you can not train your machine or your model if you have a lack of data because for small data, the machine would not work as good as for the large data sets. Hence it is advised that you should always have a large data set for your models to be known. And we'll have a look at number of data sets that we are going to discuss in the future. Then you have the need of interpretability and the level of expertise required for the algorithm design. Now everybody cannot design the algorithm. You need to be specific enough to have a derivable relationship between the data and the execution of it and hence the results. So there are a lot of things that have to be kept in mind when you are training the model and that's the difference between when you hand code the algorithm that means you are controlling the machine. Here you are making the machine self dependent to think of itself and the integral part that the machine is going to use is what we are going to learn today that is deep learning and a part of it that is neural network. So we can interchangeably intermixingly we can use these two terms called neural networks and deep learning. So let's see what is there in the neural networks today. So this is a previous slide that I had discussed about. We did a glance discussion about the AI and ML. However, I did not delve into the deep learning because of obvious reasons that we had to discuss that in a later stage. So today we'll look at the orange section where deep learning as a subfield of machine learning will be discussed. So this deep learning is based on the first thing which is neural network and the first neural network that was coined was perceptron that we will have a look at. So deep learning is a subfield of machine learning that uses artificial neural networks with multiple layers to model and process a complex data and enabling the breakthroughs in the task like image and speech decommission. So we are here to obviously make our things or the world a better place to be there. So what happened was there was a case where an image was there but was too distorted and an image that had too many lines, too many segmentations and it was like an image which had no pixels left, no very less pixels left to be cleared up. So it is just because of this learning and advancement in the neural networks theory that we got to know that that image could be taken out and you can have a great picture of the past. So whenever like previously it was obviously not advisable or it was not very handy for us but now if you have any previous pictures which are totally distorted what you can do is just get to the study of neural networks and machine learning and of course they are going to do the task. So the things have become very easy. There are various applications that have been used in past to take out these images which were totally distorted and today they are of humongous one. So AI hence is a very high-dub term and should be in the near future. So for 10 years if you are getting into this stream so nowhere the jobs are too many in the market let's see what neural networks have. So we started from the artificial intelligence. We went on to learning the machine learning and we also had executed some of the supervising unsupervised so all of it was belonging to this bone of animal. Then today we are moving on to the neural networks where we will study that what weights would do what the biasing would do what the summation of the inputs and the weights when you multiply the inputs with the weights what will those things do. So this is what the study would be however a very short study has been given in the link of Salesforce trial blazers. So we'll just as I mentioned we will just go in accordance with that as of now. So what are neural networks and why you might be interested in them? So neural networks which were initially called artificial neural networks but we dropped off the name artificial for some reason. And now we are going on to just with the neural networks things. So neural networks are a type of machine learning often combined with the deep learning. Now why is it interchangeable is because neural networks we used to have only two, three layers but when we go beyond two, three layers and there are a lot of hidden layers in between around about 150. So at max I mean the 150 layers can be accommodated and that's why deep learning went into as a coin term because there was a lot of data and a lot of execution and if you will use a lot of hidden layers in the deep learning aspect so what will happen is you will have have the outputs that will be more which will be more executable and this sort of feature engineering will happen in between the layers. So in deep learning wherever you have many layers, many hidden layers and a lot of feature extraction will happen and hence the output would be likeable, mostly likeable. Hence the small perceptron is of no use as of now but large networks combined complex deep learning networks are of most usage as of now. So defining the characteristic of a deep neural network is having two or more hidden layers and these hidden layers are ones that neural networks control. So it's reasonably safe to say that most neural networks in use are a form of deep learning. So here this is where the neural networks or the term neural has been derived from if you know the basic terminology of biology the neuron and this was a class six, class seven thing that we were taught about 10 rides and we had nucleus in between and then you have cell body, axon. Axon is where the transmission of the information happens and then you have axon terminals that are there. So this has been the inspiration point of a neural network and we have derived the neural networks from this particular aspect of biology. So it's a very basic term. So what happens in the den rides and the nucleus and then axon? So it's not the exact thing that is happening in the neural networks. However, as I mentioned that the biological term neuron has been the inspiration for the neural networks. So as things go ahead in this biological term called neuron what happens is the information is transferred and you know there are a lot of neurons in the human body and you require like too many neurons for the transfer of information in the human body. So when you touch anything and that's hot so your body sends the different signals to the mind that you have to perform a reflex, right? This is what we were taught when we were young in the science lessons. However, in this neural network, same thing will happen that this X1, X2, X3, X4 these are all the inputs that you are going to have and then these lines are also having the weights that are in accordance with the inputs. However, there was a question in AI associate program that the quiz question was there and they asked if your summation has got anything to do or biasing has got anything to do with the weights. So no, that was a false thing. You don't have anything. The biasing hasn't got any relationship with the inputs and the weights. They are totally different and these inputs are very much independent having their own weights and when you multiply you get the summation of the weights is here and then you have a biasing term over here then simply you have got the axon terminals just the way the axon terminals are here. So this is how a neural network is formed. This is a very initial level neural networks. However, we will study more complex neural networks in the data classes. Anything that you have any doubt about you can just put in the comment section because we can just answer the doubts. So we have a final theory today about neural networks. A neural network hence is an adaptive system that learns by interconnecting nodes or neurons in a layered structure that resembles a human brain. So human brain was a basic inspiration because then as soon as a computer arrived computers arrived, so scientists throughout the world were in awe that how can we make the machines independent and since then they were just moving on and then they were bringing out a lot of things. There have been a lot of AI ventures also because when the AI had a lot of what do we call investments and all, but then in between there were too many AI ventures that we witnessed and perhaps after 1950 to 19, like 2000 the work was pretty fluctuating. After 2000, 2005 this got a new rhythm and then after 2012 to you have seen that there has been a totally different vibe of artificial intelligence, neural networks, AI. So it can learn from data what neural networks neural networks can learn from data. So it can be trained to recognize patterns, classify data and forecast future events. This is the most important thing that the forecast or the future events, of course that's what we are interested in. This is the most intriguing area of human intelligence and computer intelligence that they need together to generate something to come up with seeing the future and predict the future in a very precise and confirmed way. So neural network breaks down the input into layers and these layers are then you have got different convolution neural networks and a lot of other neural networks that are going to be there in the study. So all in all this neural networks, what do they do is whatever the input is that is they're broken down into layers and then we have different, different complexities related to them, different weights and that's how the story goes. So it can be trained using many examples to recognize patterns in such in the speech or images just as the human brain does. The neural network behavior is defined by the way it's individual elements are connected and by the strength or weights of those connection. So as I mentioned that individual perceptron is of no use, we have to have multiple perceptrons or a very complex network of neural networks that are working in day in, day out. You have got GPUs. What are GPUs? The graphical processing units. So you need to have these because you are going to work with the images and a lot of data. Mostly today the data that you're having is in the form of videos and images and some sort of labor you're having some sort of unable data you're having. So you will be an individual who will be providing these sort of information and then executing these algorithms based on different aspects of life. So the last line that I would like to reiterate would be that it can be trained using many examples to recognize patterns. Patterns in what? In images and speeches as the things go and just as a human brain does. So the horizon of all of this is to be exactly like a human brain. Just the way we were taught in our childhood that this is an apple, this is a mango, this is a banana. So all these fruits, nobody or like people were there to tell us that this is what an apple is called, this is what a mango is called. So the labeling was given to us and that's why our brain may be functioning faster. However, if they would give us apple most of the time and they would not tell us, so we would be saying that, let's cluster this sort of fruit in one section, let's cluster mango sort of, so whenever you get mango, you get a cluster of mango that, okay, this sort of fruit is different. So you will yourself give the name to the section that you are clustering. So same way we are expecting now that the machines and robots would do the work. So yeah, I mean, this is how the training would happen. So the neural network behavior is defined by the way, its individual elements are connected and by the strength or weight of these connections. In the end what happens is you are going to add the input times to weight and you will get into the different bias and then the function will run on and that's the mathematics that we do behind the scenes. So here is what the first neural network perceptron looks like. You can have this image where a single layer perceptron is shown over here. You can see that there are a lot of inputs over here, X1, X2 till Xn. You can take any variable of course and these have got different, different weights, W1, W2, W3 till Wn. And now if you are going to sum them up, that is X1 times W1, X2 times W2, X3 times W3, Xn times Wn, you can sum them up and then put it in the activation function and the output. So here also what we are having is we'll have a predicted output and the output from our perceptron. Whatever the difference will be, that will be the error. Now what am I going to change with respect to that? The weights, I'll go and change the weights until and unless I reach close to the predicted output, until and unless the errors are reduced, that's the stage where you will say that okay, now the model is fine, right? So this is how our neural networks work. Let us have a quick recap of things that we discussed today. We started off from revising our supervised and unsupervised and these classification, regression and clustering. So we had a discussion about regression. We had, we know that this is a part of supervised learning and then you have got output as a continuous quantity. As I mentioned that when your percentage is around about 70 to 72% in your final year, then you get some 42,000 per month as a salary, then your percentage goes from 72 to 75. You will get a different salary, which is let's say 45,000. So you have got a continuous quantity in your output and this is what regression would look like. There are a lot of your medical digits that are going to play into regression. So main aim is to forecast or predict whatever you want to predict, whether it can be a salary, whether it can be anything that you have a new medical data output of. So you can see that the predicting the output of the stock market price is one of the things or what you can do is the output as a, if you are a real estate person that you know, who is dealing into the places. So you know that in this area that your housing prices will be, you know, more than two crores, one crore or 1.5 crores, that may be the cost of the houses. So this is also a regression analysis and the algorithm that we are going to use is linear regression. Similarly in class, we have supervised learning, output is a categorical quantity. As I mentioned, yes, no, or days of the week if you have, so Monday to Tuesday, Wednesday, or which day you're going to choose. You have got seven outputs for it. So, you know, this is category. Then one aim is to compute the category of the data. For example, classify emails as spam or no spam. So you are putting the data into, or mails into either spam bin or no spam bin. If it's no spam, it is going to show up in your inbox. If it is spam, it is not going to show up directly in your inbox. You have to check your spam box, right? And the algorithm that we're going to use over here will be logistic regression. And finally, you have got clustering. So clustering, as I mentioned, that you're going to, this is one of the examples, that you are a person and you are delivering some credit cards and you're distributionally based on how many number of people are paying the loan off, paying it very nicely and they are absolutely fine, visible score. So you will be clustering those sort of buyers and calling them and giving them out any sort of discounts or any sort of deals that you wish to. So this is how clustering would work. And similarly, then you have got simple linear, multiple linear, and one of the simple linear has been already explained in the videos ago that we did in the mathematical way. So all of it you can just go through. If you want me to reiterate the simple linear regression, it is just like, it is a way to figure out how one thing can predict or explain another thing. So imagine if you're trying to understand how the number of hours people study relates to their test scores and simple linear regression helps you find a straight line that shows a relationship between the two things, right? So in simple terms, it answers like if someone studies for X number of hours, then likely to do great in the test, right? So we can say that the output would be good if somebody is studying for longer hours. The output will be worst if nobody is studying. So it's either pass or fail. So pass will be, you know, when you are studying for the longer hours and fail will be when you have studied for not even a bit. This is how it works. Then we went on to see the deep learning and the neural networks. So in the neural networks, I believe you will have to just study this much for the program that you are looking for, the AI associate program. And you can just revisit just so that, you know, you have a complete tone of neural networks. So neural networks are, initially, they were called artificial neural networks and they are a type of machine learning often combined with the deep learning. The defining characteristics of deep neural network is having two or more hidden layers. So as many layers, as many feature extraction and as many feature extraction, the stronger will be the output, the more likable will be the output. So it's reasonably safe to say that most neural networks in use are a form of deep learning only. Right, and this is where I told you that the neural network is derived from the human body, the brain cells, the neuron of the, this is the biological term, the neuron, and this is where our inspiration for developing a neural network came from. Then we went on to studying even more into the, into a bit like a neural network is an adaptive system that learns by using interconnected nodes or neurons in a layered structure that resembles a human brain. So as the brain studies, this is what the target of the current world is, that that's the way a machine should learn. And hence we are improvising a lot of algorithms in those aspects. So it can learn from data. So it can be trained to recognize patterns, classify data and forecast future events. A neural network breaks into the, breaks the input into layers and that's how it works. It can be trained using many examples to recognize patterns in speech or images just as the human brain does. So this is what we did. And the final, the first neural network perceptron. The next lecture we start from mathematics of perceptron. So I'll put some details about perceptrons and how the match behind it works because I remember there is a form, please, it is a neural network thing. It's a long formula. You won't have to learn anything about it. Neither will be there in any of the quizzes that you appear for. It's just that I'll tell you a bit of maths behind it that is just to touch, that you have done perceptron in a manner that it should have been done. That's it, but that will be not there. The maths will not be there in the certificate questions. So I think this is it for the day. Yeah, I think you already covered everything. You very well revised whatever we have done so far. And like you explained neural network also. And what I am thinking like maybe in upcoming sessions, we will try to relate our trailhead modules with the topics which we are discussing so that people can go and they can take those quizzes also. So we can guide them like this trailhead you can follow along with this topic and just prepare for the quizzes so that along with the theoretical knowledge and whatever mathematical information you will be sharing, based on that they can go through the quizzes and they can prepare themselves for EI associate exam. I think that will be a good idea. Correct, correct, correct. I think after this neural network questions would be pretty much doable. Yeah, or maybe like we can just share those on the screen because they are publicly available. And we can guide them like if this is the question so what can be the output. So we can ask viewers like what should be the answer and this way our sessions can be more interactive. So we'll try to plan upcoming sessions accordingly. Okay, so anything else that you want to add or you are done? I'm done, I'm done. Okay, okay. That's it for my session. Thank you so much. So guys next week we'll be having two more sessions, right, so like from now onwards every week you will be having few sessions so that if you are following all the videos so you will be able to prepare for Salesforce EI associate certification. Okay, so thank you for watching this video and like if you are watching the recording so thanks for that also because I know people are following this channel from different geographical location so if this time is not suitable for you so you can watch the recordings and anytime if you have doubt so you can reach out to us on LinkedIn so that we can help you out. Okay, and like soon we'll be going to start like in-person training programs also so I will be sharing those details with you so that along with other admin and development programs you can start learning EI also. Okay, so thank you so much and thank you Nikita for sparing some time and sharing your knowledge. Thank you so much. Thank you guys. See you next week with new topic. Bye everyone.