Deep Learning With TensorFlow and Keras [FREE] | TensorFlow Tutorial | Keras Tutorial | Simplilearn

Key Takeaways

Hey everyone, welcome to our deep learning with KAS and TensorFlow course by simple learn. Deep learning is behind some of the most exciting advancement in technology today. And learning KAS in TensorFlow is a powerful way to dive into that field. In this course, we will start by exploring how model predictions work and how sigmoid activation helps you sense of those predictions. We'll also teach you how to interpret probabilities in machine learning so you can understand what your models are telling you. We will cover data prep-processing with numpy which is crucial when working with large data sets. You'll also learn how back propagation works and how it helps train a neural network. We'll also look into how adjusting the learning rate can improve your models and help them learn more efficiently. As we go on, you will understand epox and forward propagation and work with the fashion MNIST data set. A fun example to practice deep learning techniques. We'll also discuss hyperparameters and how to configure models for better performance. Plus, how to debug and analyze errors that may pop up. Finally, you will dive into the convolation operation which is key for image related task and learn how to fine-tune hyperparameters to get the best out of your models all while tackling the challenges that come with neural network training. And by the end of this course, you will have the skills to confidently build, train, and optimize deep learning models using Keras and TensorFlow. So, let's get started. Before we move on, here's a quick quiz question. What does deep learning primarily help us with? Your options are making websites, making predictions from data, creating text documents, or organizing files. Let me know your answers in the comment section below. Before we get started, here's a quick information. If you're interested in boosting your career in 2026, Simply Learn's professional certificate in data science and generative AI program in partnership with Pura University is the perfect place to start with. This program covers everything from Python and machine learning to generative AI and charge GPD giving you hands-on experience with cuttingedge tools and technologies with per university certification. IBM's industry recognized courses and simply learns job assist to help you connect with top employees. You will be ready for high paying roles in tech, finance, healthcare and many more. So hurry up and enroll and you can find the course link below. >> Let's get started with this. So as we understand deep learning is nothing but an extended field of machine learning and it has proven to be useful in the domain of text, images and speech. So the collection of the algorithms implemented and the deep learning have the similarity with the stimulus and the neurons exactly how they are seen in the human brain. and deep learning has extensive applications in computer vision, language translation, speech recognition, image generation and so on. So these set of algorithms are simple enough to learn in both a supervised as well as unsupervised pattern. So if you actually look into it, a majority of the deep learning algorithms are based on the concept of artificial neural network and the training of such algorithm in today's world has been made possible because today number one we have huge amount of data. Number two, we have huge amount of processing power. Number three, we have huge amount of storage as well. Because these three things are there, deep learning actually geared up. Else there was no reason for deep learning to gear up so fast. So deep learning has these all advantages to say with additional data. the performance of the deep learning algorithm keeps on improving only. So if you see over here this is mainly your machine learning where you see as the amount of data increases performance increases but after that you see it stabilizes. But with deep learning initially the performance is not as great as machine learning but as the amount of data increases the performance of deep learning algorithm keeps on increasing. So deep learning loves data. Deep learning basically loves data. That is something which is very very important point to be noted over here. So whenever you have huge amount of data you have deep learning. Then going ahead see the term deep in deep learning refers to the depth of the artificial neural network architecture. Yes something like this. So in any neural network you will have an input and you will have an output. Specifically this is the input layer and this is the output layer in between this layer and this layer we call them as hidden layers. So specifically I can call this as hidden layer one. I can call this as hidden layer two. Each ball is called as a neuron. So this is a neuron. This is a neuron. This is a neuron and input layer. This is the input. There are no neurons in this layer. So although they have drawn circle, most of them draw circle but ideally it's not like that. So specifically, you know, if I have to redraw this figure for you, you can consider this is your input one. This is your input two. I'm just assuming this as x1 and this as x2. And then one layer, right? So this is connected here. This is connected here. This is connected here. This is connected here. This is here. This is here. This is here. This is here. This is here. This is here. Finish. So that's our output layer. That's our input layer. This is hidden layer one. This is hidden layer two. [clears throat] Then yeah. So I say the term deep in the deep learning refers to the depth of the neural network and learning basically stands for learning through the artificial neural network. You know how these networks learn they initially you know imagine this that I have a book to study. So let's say this is the book that you have to study. So consider this network to be nothing but you. What you do? You take this book. You study studied. Then you give an exam on that book. Once your study is over. So here while doing this forward propagation you basically study and while doing this backward propagation you give exam. So you are studying for the first time and you are giving exam for the first time. That is exactly what I say. You have studied once. This is exactly what I say. you have studied once and you keep on repeating this process multiple times. So this is forward propagation and this is backward propagation. So one forward propagation and one backward propagation is nothing but an epoch. Right? So this is nothing but called as forward propagation and this is what we are calling as backward propagation. So one forward and one backward propagation is called as an epoch. So basically what you do is you study for n number of epochs. So I say repeat for any epoch. That's exactly how a human studies anything right? You don't study anything only once and say that okay now everything is saved in my brain. You keep on studying the same thing again and again. So neural networks are also doing the same thing again and again. See these all things are like a story. So you have to understand it step by step slowly and steadily. So one forward propagation is studying once. One backward propagation is giving exam on that. So you become better on it. So you can repeat this study exam study exam study exam again and again. So over here this network is called as a deep network and this network is called as a shallow network. Again you see here we have input layer. Here you have output layer and here is what you have hidden layer. But the difference is in shallow network you have only one hidden layer and here you have more than one hidden layer. Okay. Here you have only one hidden layer and here you have more than one hidden layer. So next so deep neural networks are capable of what? See the deep neural networks they are mainly capable of discovering latent structures or I can also say feature learning from unlabelled data and unstructured data such as images which is nothing but the pixel data. We have documents which is the text data. You have files which can be nothing but audio or video. So basically deep learning neural networks can automatically extract features from the data. In machine learning I remember we had to extract the features like in Titanic data set we extracted features by ourself right so from the data sets itself I remember we engineered the feature for extracting uh price we engineered feature for extracting the passenger class family class etc But in deep learning all of this feature learning happens automatically. So I would say automatic feature learning. It can do automatic feature learning. So how deep is deep? See, we often hear the term deep and we instantly get intimidated by it. But there is not much difference between a shallow and a deep network. So over here you have only one hidden layer and over here you have more than one hidden layer. So a deep neural network is simply a feed forward neural network with multiple hidden layers right multiple basically multiple means more than one that is the basic difference in between the two. Next we get into the basic structure or the history of the neural network. Now here see each unit or a neuron is simple. Now what you see on the screen right now is a biological neuron. Right? So if you consider if you consider a human brain, our human brain it has 100 billion neurons and these are with 100 trillion connections. So the strength and the nature of the connections stores the memory and the program which makes us human. So the neural network ideally is a web of artificial neurons. This is the biological neuron. Then I say a n is nothing but a web of artificial neurons. So here see let's understand how does it work biologically. So let's say uh h this is you. So now let's say here if I'm pinching you you know this is the point where let's say I pinched you what will happen signals travel from here to your soal brain right signals travel from here to your so-called brain oh it's so bad now but when signals travel from here to your brain. Then where do they go? It basically goes to the dendrites. This one, these dendrites receive the signal and then the signals get processed in the neuron. This is exactly where the signals gets processed. Then if the pinch was very hard, hard in the sense outside your bearing capacity, you know, I pinched you really very hard and you just can't resist the pain, then this nucleus will decide that is that pain really harsh. If yes, it will immediately inform the axon to go to your right hand to go to your right hand and quickly go and rub over here. Ah and if the pinch was soft then this neuron decides that you can ignore it. So this is how the neural network used to work. So if you now see in [snorts] so uh I was saying artificial neural network or perceptron was first developed in the year 1950 by Frank Rosenblad and this looks something like this. There are inputs there is an output. So basically it takes several inputs produces a single binary output if the sum is greater than the activation potential. So the neuron is set to fire when the activation potential exceeded and behaves as a step function. So the neurons that fire along the uh signal to other neurons are connected to their dendrites which in turn will fire and that's how it continues step by step. So now this is how it looks like. So now let's understand how does it actually work. It's very very easy. So remember this thing in very simple words we are trying to understand what is a neural network. So the neuron which you saw right so if you see this is artificial neuron inputs come over here get processed and then signals flow from here. This is a biological neuron. In the artificial neuron you have inputs processing which is nothing but the nucleus and this is where the output comes to let's understand this with the help of a simple example. So here uh the neurons can be sketched something like this input in machine learning. If you remember if I was talking about the iris data set then I had seleength sele petal length petal width and I had the class of the flower that's how my data set was and this data set consisted of 150 rows and five columns then we divided this data into train and test. This was independent variable 1 2 3 4 and 5. So this is also very much similar to that only. This is my x1 x2 x3 x4 and you get over here an output. So basically earlier the machine learning model was understanding the uh relationship between independent variable and the dependent variable. But over here this is the independent variable. This is the dependent variable and it is a neuron who is understanding the relation and that do not one time but it does that multiple times or I can say that it does this for many epochs where We know that an epoch is nothing but a forward propagation and a backward propagation. [clears throat] Okay. Now let me uh come to Manu first. Let before I forget these layers are derived from another system. Uh not really Manu. Uh we are yet to get into the mathematics of what you are asking. Slowly and steadily uh you know the story of neural network will be something which will be little bit hazy for you at least for next I would say 6 hours. So you know you can imagine this neural network to be like a uh a movie which goes in the future and comes back just like uh I'm sure you might have seen the movie Three idiots. You remember that right? I hope Manu you have seen that. Yes. So we see that till the last till the last the movie keeps on going sometimes in the future sometimes in the past again it goes into the future again it goes into the college life. So the entire movie becomes clear at the end. So here also in neural network uh the story is there are a lot of small small connected things which you will understand step by step and our end will not come at the end of the syllabus. Of course our end to understanding this working you can expect a total time span of around 8 hours from our today's session to understand how does this thing work. So basically we need to mathematically understand what happens in forward propagation and what happens in backward propagation for us to get an answer to your question. So slowly steadily it will be clear. So don't worry about that. Right? Yes. Now let us actually look into how does a neural neuron basically work. See um let's consider this. So I'm drawing a very basic structure of a neural network and precisely perceptron or neuron. H it has three inputs and it has an output. So the output over here can be either zero or one. So the thing over here is sum if the summation of wj xj is less than is equal to threshold it is zero. Else if the summation of j wj xj is greater than threshold the output would be one. So let's understand how. So let's say um I'll take a very nice example to make this thing easy for you. Remember this example for lifetime. So here suppose uh weekend is coming up. And you heard that there is a Punjabi food festival in your locality. I'll come to it. I'll come to it Raj sometime. Suppose a weekend is coming up and there is a Punjabi food festival in your locality. So now the thing is how many of you like Punjabi food. Now let's say you are trying to decide whether Okay. So over [snorts] here the the task is you are trying to decide whether I will go or I will not go to this Punjabi food festival and your decision depends on the following factor. Factor number one. Is the weather good? Yes or no? Now when I say weather is good means uh see if it is not raining according to you it is good. If it is raining according to you it is not good. Or if it is too hot according to you it is not good. And if it is too cold according to you it is not good. So you decide what is good and what is bad. Second does your partner wants to accompany you? Yes or no? So is your girlfriend is your boyfriend? Is your wife, is your husband ready to come with you to the festival? Yes or no? And the third one is, is the festival near a metro station? So let's assume that you don't have a car. So obviously you want the festival to be near the metro station. So that is the next thing that you are trying to decide. So your decision depends on the following factors. Your decision depends on the following factors. [snorts] So now now you are trying to decide whether you will go or not to this festival. Okay. Now, now a little bit of tantrums, let's say you are such a person. You are such a person that uh let's say your wife says dan uh let's say I am such a person that my wife says that dan I'm not coming and I'm like okay no problem I will go. Uh the festival is not near the metro station and I'm like I will I will travel somehow. Huh? I will go. But if it is raining, if it is raining then dare I keep my leg outside the house. I just hate rains. If it is raining I am not going to step out of the house. Not even little bit. I just hate rains. I just hate rains. If it is raining, I'm not coming. So what does that say? It basically says that for a person like me, even if my wife says dash, I am not able to make it up. It is okay. I mean, I'll manage. It's not exactly near the metro station. That is also fine. I will adjust. But if the weather is not good, I will not adjust. So over here among the three factors, I realize that weather is something which is more important for me as compared to the other two factors at least for a person like me. And that is exactly what we quantify by giving weights over here. Raj. So obviously if weather is more important because these are the three factors right. This is the factor x1. Yes. This is the factor x2 and this is the factor x3. among these three this is more important to me as I told you and this is relatively lesser important. So what I do is I give a higher weight to this and little lower weights to this. Now I can give any number. These all are just all imaginary number. It is just that I have given a higher value and little lesser value. Again this can be two, this can be one also or this can be three. This can be two also. These all are imaginary values which I have given. Are you all with me so far? One are one is the deciding deciding factor and the second is how much that factor is important. Everyone clear so far. [snorts] Right. Yes. Very good. Now over here I also did say I also did say that when I'm pinching you huh let's say let's say a girl pinches my thick skin. I won't feel that pinch a lot and I can just smile and ignore. But if I pinch the soft skin of the girl with the same intensity, she shouts, "Ouch." Am I correct? Am I correct? Right. So essentially what I'm trying to say is the painbearing capacity of everyone's skin is different and so is a similar fun which I am introducing over here which is the thresholdbearing capacity and here I am assuming that the threshold is five assuming this is assumed This is assumed and that is what is the figure right now. So let's try to [clears throat] come to the next slide. X1 X2 X3 that's the output. W1 W2. Now what did I say? The output can be either zero or one. It will be zero. If summation w dox is less than threshold, it will be one. If summation w dox is greater than threshold. [snorts] So over here let's consider a scenario. My the weather is not good. So zero my wife says dashan I'm ready to come with m you and it is next to the metro station. So I want to decide. So for deciding I have to do w into x. You know what is w intox? W into X is nothing but W1 X1 + W2 X2 + W3 X3 which is 6 into 0 + 2 into 1 + 2 into 1 which is which is four. Now four that's that's the output four is four less than equal to threshold or is four greater than threshold. Four is less than equal to threshold. So if four is less than equal to threshold tell me on the basis of this output would be zero or output would be one? Output would be zero or output would be one? 0. So I say output is equal to zero. That means I won't go to the festival. Am I correct? My decision over here is I am not going to the festival. Right? Let's consider the other scenario. My weather is good but wife is like darian I'm not coming I am busy in office work and the third factor it is away from the metro station. So this becomes 6 into 1 + 2 into 0 + 2 into 0 which is 6 + 0 + 0 which is 6 which is greater than threshold. Therefore, the output would be what? Put will be one and therefore I will go to the festival. You tell me if my wife is not coming. Sorry, if the weather is not good, wife is coming but the festival is away from the station. Tell me whether I will go or I will not go to the festival. No. Right. Correct. Tell me that weather is good. My wife is coming but it is away. Whether I will go or not go. Yes, I will go. Right now I'm changing the ink color. And let's say if I am assuming the threshold over here as three. If I'm assuming the threshold over here as three and let's say the weather is not good but my wife is ready to come and it is next to the metro station. Tell me whether I will go or not. Threshold is three. Now I'll go. Yes. Very good. I'm not I'm not computing a Yes. Okay. I will go. If the weather is good but my wife is not coming and this is also not in my favor. Then will I go? Yes. Okay. I will go. So basically what am I trying to say is by varying the weights and the threshold we can get different models of decision making right so in other words it is a different model of decision making if I am changing the things after that let us try to do a little bit further modification now now what was this formula if you recolct So the formula over here was something like this. Output can be either zero or one. It will be zero if summation wx is less than equal to threshold. It will be one if summation wx is greater than threshold. Now what did researchers say that This is complicated for machines. Therefore, we need to simplify this. A so they said consider wx less than equal to threshold. Therefore I can say w dox minus threshold less than is equal to zero. Then they introduced a new term called as bias. and bias is minus of threshold. Therefore, this equation became wx + b less than is equal to zero. So, I can kind of say that now this can be modified where I say output would be zero or 1. It would be zero if w do.x + bias is less than equal to 0, it's 1. Okay. W dox plus bias is greater than zero. Yes. So this is what we understand. Now why is this better? The reason this is better is machines love zeros and ones. It's much easier for a machine to compare against a zero value. If the output is greater than zero. Uh so basically if this is greater than zero then one is the output. If this is not greater than zero zero is the output. This is far far far convenient for a machine like consider an example. So consider a perceptron where the bias is this. I have x1 with the weight minus2, x2 with the weight minus2. That's the output. And what was the formula for output? It can be zero if w do.x + b is less than zero. it would be one if it is greater than zero. So over here let's consider x1 and x2. If it is 0 0 then let's perform one calculation wx + b which is -2 into 0 + -2 into 0 plus bias is 3 3 it will not be minus3. Don't consider the bias as minus3 because I clearly said bias is minus of threshold. So going forward you know going forward I will never ever talk of a term called as threshold. You will only talk about bias. You will only talk about bias. So threshold forget threshold now you will not see anywhere in the documentation as well going forward. Everyone will talk about biases only. So this comes out to be 0 + 0 + 3 which is equal to 3 and 3 is greater than 0. Therefore I will go to the festival 1 1 - 1. Okay. 1 1 - 1 1 is greater than 0. I will go 1 is greater than 0. I will go minus one is less than is equal to 0. I won't go to the festival again. Okay. Now tell me the same calculations with bias uh sorry not bias with the weights as let's assume this to be two and let's assume this to be four. tell me go not go scenario based on this. So this was one thing. Now uh if we uh now a problem. So the problem if I show you your family photograph. Let's say if I'm showing you a family photograph, your family photograph. Let's consider this is your family photograph and I am assuming that this is who you are. Huh? So I ask you that hey you know what tell me where are you? So let's say you know today you are 25 years old and this is when you were just 5 years older. Let's say this is when you were just 7 years old. My question over here to you is will you be able to recognize how you looked 7 years back? Your own photograph in a family photograph. Do you think you'll be recognize? Because you are saying yes. Because you are saying yes. My question is how do you recognize that this is you only and not someone else? How does your brain recognize that this is you and not someone else? Face patterns, features, facialia features. I had seen my photograph before features we have in mind in the memory. So how does your brain store the features? inside this box how does it get saved right I'm getting a little deeper into it memory by seeing multiple times in the neurons cerebellum but how is it saved in the cerebellum how is it saved in the neurons how is it saved in the memory I mean uh what is the format in which the features are saved 1 zero some weights links are made permanent. But how? You all are speaking some technical terms. If I say that in simple words as a layman, tell me how is it saved in your com in your brain? You say I have watched multiple times. I have learned features. How does your brain even know that this is a feature? How does your brain even know that okay this is a nose this is a eyes this is a left eye this is eyebrow this is lips how does your brain know that learning what is learning how much soever you try to justify we ourself doesn't know how our brain recognizes us in a old photograph technically how muchever I say so you will say that okay something is saved in the form of a [snorts] chemical in my brain so how is the chemical storing that we have absolute no answer to all of these questions we have no answers to all of these questions and that is the most important thing we ourself self don't know how our brain is working. Anand says building from correlation or similarity or from what is known. What is correlation and how does your brain know how to correlate? How does your brain even know that this is how I correlate from whatever is last known? So your last known information is somewhere saved in what format is it saved in your brain. How muchsoever we try to justify it we won't be able to actually get the correct answer to that human brain is far more complicated than what we can think of. So at some point of time you are likely to say dash please shut up. I don't know how is it saved in my brain but I am pretty sure about the fact that this is me even when you say electrical switch as a storage have you ever seen that electrical switch manu saw that this is all bookish knowledge which your teacher has taught electrical signals waves traveling electromagnetic waves electromagnetic induction all of these are bookish things we all just understand that how the information is coming out of the brain we don't know but we are like I am sure this is me how I don't know dash so here what am I trying to make you realize is what am I trying to make you realize is we ourself does not know how our biological neural network is working. But we all agree to the fact that we know how a artificial neural network is working. Correct? We didn't know how a biological neural network is working. But we completely understand how does a artificial neural network work because you only solve two problems on it. Yes. And this is the problem. This is the problem. What are we here for? We are here for building a artificial neural network. This network is not even like a network. Anyone can do some plus and multiplication and get the output. So are we really saying that we are able to build a artificial neural network? Answer is no. A neural network is the one which we ourself doesn't know how it is processing but here it is clearly visible. So basically I don't want it to be predictable. So we need to now change a few settings. Now what am I going to do is I want to show you the secret source of the neural network. Another imaginary concept for now we call it as the activation function. This is called as the secret of the neural network. What it does is it converts something which is predictable to something which is unpredictable. Deep learning is entire about fuzzy logic. That's absolutely correct. So I was saying over here now what am I going to do? Now this is what you considered as the secret of the neural network. So what is happening right now? What we learned till this point of time was this part. This is what we learned so far and this is what is the activation function. So till this point of time I said that the output is predictable and I want the output unpredictable. I want the output unpredictable. Okay. So now uh see understand what is the problem. Uh if you recolct uh linear regression algorithm you remember the linear regression algorithm y= mx + c what we studied during machine learning right? Yes. So now I remember taking you to this concept of slope of the line where I said in linear regression algorithm. This is what we did. I remember I had taken several points and for all of the several points we added a neural network. So we have created a equation which is nothing but the equation of the line of best fit whose equation was of the form y = mx + c. In this case, m is 0.5 and that c let's say 2.5 and this was the equation of the line which was used for predicting the future things. Now the problem if you see right now the equation that we are having is output is equal to W into X + B X W M and this and this. So this is the equation which is linear straight line which means that what we have over here is also linear which means that the output of the neural network right now is linear and this linear output cannot be considered as the working of how a brain is. No. So the thing is I don't want the linear output and that is exactly what the activation functions do. What they do is basically whenever you have inputs x1 x2 x3 what output you get is linear. What happens over here? Here w into x + b happens and what you get is a linear output. You pass it through the activation function and the final output that you get is a nonlinear output. Now how well there is a lot of mathematics over here. If I actually take you through the activation function in the TensorFlow libraries. Yes, these are the various available activation functions. So many are there. All of these are responsible for converting a linear output to a nonlinear output. Can you practically learn the mathematics of each of this? Well, the straight answer to that is no. Right? Uh so will we be skipping the mathematical part? No. I will be taking you through the important mathematical part. It's not that we are going to purely ignore it. But how muchsoever you try to discuss that in detail you are definitely going to forget. So you know there is a typical way we need to understand how this works. Well, that's too early to talk of. But yes, we will be going through these all things. But sir, why do we want from predictable linear output to unpredictable nonlinear output? Because because we are here not to implement linear regression once again. Basically if you look into this amito whatever I explained you over here right whatever we have understood on this slide in a simple way I can actually say that I have fooled you all so far because technically what we are implementing is nothing but linear regression algorithm mx plus See now if it is greater than zero or less than zero without knowing I successfully fooled you all because this is nothing but linear regression. Yes or no? This is nothing but linear regression without even you realizing. Our brain does not work linearly. It thinks of n number of possibilities and therefore we don't want the output to be linear just like I explained the example of that photograph where we ourself don't know how our brain is able to predict where we are in a photograph. In the same way here as well we want the neural network to be correct but we ourself doesn't know how is it actually predicting the things that is the reason we have to add this the performance of the neural network if you don't add an activation function could be even bad than a normal machine learning model. it won't be even matching the machine learning model. So the secret source of the neural network is the activation function and hence we need to add this thing getting that so it's a long story as I told you right slowly steadily you will understand it but let me tell you without activation function your neural network will just not function as expected because you are doing nothing but w into x + b which absolutely does not serve the purpose. So to uh convince you more, what am I going to do is once we come back, I am planning to get started with a uh little bit of hands-on to make you understand how does it work by leaving for some point of time theory aside and focusing on the implementation aspect as well. So you know it would be basically covering both at the same time theory as well as the implementation parallelly. So uh how I will let you know as and when we go ahead but yes we will be looking into all these things. So you know you will be little bit more convinced rather than randomly uh going ahead with the things. So but you cannot create a neural network without an activation function for sure. How you will understand as and when we go ahead I'll explain you with the help of very simple problem. Is the activation function a part of the neuron or it is outside it? Yes, that's a good one. So here if you see this is how I wrote it clearly indicating that it is not a part of it but going forward going forward I will never draw a figure like this. I will always draw a figure like this where there will be inputs and there will be output. What happens internally is w into x + b and activation function applied over it. So this is the standard and the default figure of a perceptron from now onwards. Anytime you think about a perceptron, you will always see a perceptron like this. So expect this visualization to be like this only. So the first notebook is a linear classification. The first one is I am importing the framework TensorFlow as TF and then I am printing the version of TensorFlow. It doesn't take so long to import but sometimes it happens. Even if you are having a different version of TensorFlow, that's not at all a reason to worry about. So this is what we have. And let me say what is basically TensorFlow. So TensorFlow is an opensource machine learning framework which is developed by Google. It provides a comprehensive ecosystem for building and deploying machine learning models particularly deep learning models. It supports various tasks such as image recognition, natural language processing, time series analysis making it very versatile for data scientist and the developers. TensorFlow is widely used by both research and the production environment because it is very scalable, flexible and there is an extensive community support as well. That's an advantage that you get with TensorFlow ideally. It allows complex neural network and optimize them for performance and deploy them on different platforms including mobile as well as the cloud services. Then this framework works on tensors which are multi-dimensional arrays and provide highlevel APIs for building and training machine learning models. It also includes tools for visualization such as tensor board which helps in monitoring the performance. So let us get started with it. So the first thing which I want to do is I want to [clears throat] load in the data. So I'm taking a very basic data from sklearn dot data sets. I am importing the load breast cancer data and then I am checking the type of the data. So, so I would say uh once the data set is loaded I say over here data is equal to load breast cancer and this is the type of data. So I can say the type of data is a skarn utils bunch object. The type of the data is a skarn util bunch object. So this is mainly like a dictionary which will have key and value pairs. So if I actually say data do keys this will show all the keys. Now uh see here also you know one of the keys data then target then frame etc. Let let me load this as df then so that there is no confusion. So df dot data will give me the independent variable. DF.target will give me the dependent variable and so on. I can actually uh check it out. So if I say pf do data dot shape. So there are 569 rows and 30 columns. So understand over here we have independent variable 1 to 30 and there are row number 0 to 568 because 569 rows are there. Okay. And if I say df dot target dot shape 569 comma which means there is one more dependent variable. So basically it is 569 rows and 31 columns. Yeah that's what we understand. We have 569 rows and 31 columns. Okay. If I say df dot target and if you actually try to see that see these are the values and if I say df dot target names this will show the target classes. So you can see there are two types of classes. So over here zero is the malignant and one is the benign. >> So it's a multi-class class binary classification problem to predict whether a female has a malignant type of breast cancer or a benign type of breast cancer. So what is malignant cancer? So it is a cancer which is characterized by uncontrolled growth of abnormal cells uncontrolled and in the same way. What is benign cancer? It is not that cancerous and relatively less dangerous. It is relatively less dangerous. It's not life-threatening. So we are trying to do this problem right to identify whether the female has malignant type of breast cancer or a benign type of breast cancer. Now see I said there are 30 independent variables. If I actually say DF dot data these are the ones but instead of saying that it's better that we see df dot feature names so I will come to know what is x1 x2 up to x30 df dot feature names and let us find out the length of it 30 features and if If I say PF dot feature names, these are the ones. So you clearly understand we have the mean radius of the breast. Then you have the mean texture of the breast and so on. There are so many features. Next thing uh that we need to do is train test split. So we say over here that from skarnon dot model selection import train test split then extra x test y train y test df dot data df.target target and uh I'm keeping the test size as 33%. I want to print the shape of each of this data. So we can see we have further divided it into 381 and 188. There are 381 188 respectively. That was uh that's what we understand over here. Now the next thing is I'll be going for scaling. So I say features is equal to DF dot feature names and I don't want it in the form of a numpy array. That's perfect. Then I say extra of features no from sk.processing import standard scala create an object of it by the name standard scala and then I say sc.fit fit transform on this and this what it says ah extend of features that's fine scr extra of features okay extra of features itself is not there because that's not available so yeah that's correct so I'll directly scale it So I say from skarn do this create an object of standard scaler and we have this you want I'll restart your output and run all perfectly now after that let's go ahead and create a neural network. So here how do you create a neural network? See understand over here what what are we mainly going to do? Our process is more or less going to be the same. So understand this without a lot of background I'm actually trying to create a simple perceptron model. So what I want see what I want is I want the model to understand the relationship between this row and this column. Then second row and this value. Then this and this. So on so forth till this. This is what I wanted to study. This is exactly what I wanted to study between X train and Y train. So this is nothing but X train and Y ring. We just have to understand the relationship in between the two. So after this, how do you create a neural network? So pay attention. I'm going to show you two ways of creating a very basic model. You know what am I going to create? Listen to me very very carefully. I have 30 columns. 30 columns. So x1, x2 up to x30 these are available over here. Then after that we are going to create just one perceptron and there will be output. So basically this will go inside this will go here this will go here and there will be multiple weights over here weight one weight two weight 30 right this is the basic neuron which I'm going to create and what happens over here is of course there will be activation function on wx +b this is essentially what we'll be having. So here what I do mainly over here is I just add this and then let us go ahead and create a neural network. So I'm going to build the model and there are two ways of doing it. the post one where I say model is equal to tf dot karas dotsequential and I say tf dot karas dot layers dot I'll explain you don't worry input shape is equal to 30 comma and then name is equal to input layer and tf.karas.layers.tense tense and I want only one thing over here activation is equal to sigmoid right yeah we are done with it so observe over here how am I creating a very very basic neural network here I'm saying tfkaras dosequential it looks like a little bit odd way of creating but this is the one and only way of creating a neural network tfkaras dot layers dosequential So you know what is sequential? Uh see if you are going to your office or you are going to your college wherever you are going that's not important but if you are going to your office or you are going to your college the first and foremost thing that we end up doing is we don't car see if I want if I want to carry my phone I carry phone in my hand Then I want to carry my laptop. So I'm carrying a laptop in another hand. Then I want to carry the charger. So I put my charger over here. Then I want to carry my tiffen. I put my tiffen over here. Is this how I go to the office? No. I put all of my items into a a bag. Right? I put all of these items into a bag. In the same way whenever you create a neural network this sequential layer right from here till here you know the sequential layer can be considered of as that bag only. So whenever you want to create a neural network be it a single perceptron one or be it consisting of multiple other things you will always focus on putting these things into a bag like structure and sequential is for that only so now I say this format is always common tfkaras dot so sequential layer then tfkaras do.input input because we have to create an input layer of 30 neurons. Ideally, there are no neurons over here. Therefore, I didn't draw a circle. So, I say tf.karas.layers.input shape is equal to 30 comma. Basically, 30 values will be coming as input is what I'm trying to say over here. And then the name this name is optional. You can give the name of your choice or you can ignore the highlighted portion completely. Similarly you can create one dense layer which is just of one neuron activation I have supplied over here as sigmoid right just give me 10 seconds so I was saying input layer done output layer tf.karas kas.layers.dense. So see every time you are adding a layer there are only two types of layers. What I said earlier that we have a input layer, we have output layer and we have hidden layers. Now technically there is only one input layer and the hidden layer and the output layer are called as dense layers. There is nothing called as hidden layer. There is nothing called as output layer. All of these are nothing but the dense layers. So we say pf.caras dot layers dot dense one neuron because there is only one neuron which I want to create a perceptron model activation is equal to sigmoid and name is equal to output layer. So let me tell you activation function input layer. The input layer receives the data and there is no activation function applied here because it is the input layer. Never never never. The second is the output layer. Now here I say if we are performing a binary classification problem then the output layer will use sigmoid activation function and B if you're dealing a multi-lass classification problem then the output layer will have soft max activation function and C in the regression problems the output layer will have the linear activation function. Remember these three rules very carefully. Input layer no activation function. In the output layer for binary problem which is in our case malignant cancer or benign cancer output layer will have sigmoid. And for multiclass classification problem like the iris data set where you have the target variable as stosa or versicol or virginica then you will be using the soft solving a regression problem you will be using linear. See right now I would kind of say memorize this for now. Memorize this for now. Right now only. Write now only memorize this. Then after that let me ask you which activation function would you use if you're solving a multiclass classification problem in the output layer. Multiclass classification problem. Very good. It is soft max. And for regression problem output layer will have linear. Which problem are we solving? Binary or multiclass. Binary. And which activation function will we use? Sigmoid. One last rule I would like to add. One last rule. Input layer done. Output layer done. The third is hidden layer. So here use RLU as the activation function for the hidden layers. So see these are the rules. There are many more activation function. I'll show you as I showed you earlier as well. There are many activation functions other than these as well. Elu, exponential, gau, hard sigmoid, hard celu, hard switch, leaky, relu, linear, we saw log soft max. Remember the rules because other activation functions are not very important. When to use it, we will discuss it later on. But for now remember we use a ru we use the linear we use sigmoid and softmax. These are the most important and the basic ones. Okay. So I said this is the input layer which does not have activation function. There is no hidden layer over here because I'm doing a basic implementation of a single perceptron. So you can see there is only one neuron one neuron and you know from where is this coming I'll show you if I actually copy this and I go to Google tf.karas.dens dense it will take me to the tensorflow documentation over here. You can see number of units we gave one activation we gave uh uh sigmoid and there are many other terms which I didn't discuss as of now but remember that they exist. There are many other terms but remember that they exist. It's just that we have not used it for now but they exist. Okay, this is one way of creating it. The same thing I can create it using [clears throat] the second way as well. This is what I like it and I find it more easy as well. Model is equal to TF do.as kas dot models dosequential model equal to tf.tkas kas.sequential and you used to write everything inside this but here I'm doing everything layer by layer model is equal to this done in this model I say add add what tfkaras do layers dotinput shape given and name given so I said add this input layer so here it was everything inside sequential but here I'm adding it layer by layer Further I say model dot add what tf.tkaras.layers dot dense one neuron only which is like this activation sigmoid name output layer. So I do that see this I commented and executed this it works. I can actually say model dot summary. This will display the architecture of the model. So this is the summary of the model. If you see input layer is not even mentioned output layer none, 1 31 parameters. Why 31? Because 30 are the weights and there will be one bias 30 + 1. So this is a very small neural network having 31 parameters. This is a very small neural network having 31 parameters. Okay, before I go ahead, can you please Google for me and let me know how many parameters how many parameters are there in chat GPT3.5? It is 175 billion parameters. This is only 31 parameters. Can you imagine how big that neural network would be? 75 billion. So what we are building is a very very basic one. What we are building is a very basic one. So model dots summary gives us the architecture. Okay. Now you know what what you created just right now. Uh this is like creating of a car body frame. Yeah, this is what is created. Right now it is important that you fit in the engine of the car which is yet not fitted. So I say this step is like designing the body frame of a car. That's what is ready. Now let's fit in the engine to the body frame of the car and optimize the engine either for mileage or performance. So here we have the name of the model by the name model. So I say technically we call it as compiling the model. So I say model dot compile optimize the model using the optimizer atom. Optimize the model using the optimizer atom. >> [clears throat] >> the error in the model or the loss in the model because I'm solving a binary classification problem. I'm using the error function or the loss function as the binary cross entropy and the metric would be accuracy. So model compile optimizer is Adam loss is binary cross entropy metric is accuracy. So basically it is like fitting the engine into the car and then we optimize the engine with some optimizer ensuring that the loss which is quantified using the binary cross entropy is minimized and for that we have the