The Absolute Beginners Guide to Artificial Intelligence

1. Introduction to the AI Course: Hello there, and thank you so much for enrolling in the course. The purpose of this introductory video is to welcome you officially, to introduce myself, and also give you a general idea of what to expect from taking this course. So first things first, my name is Alexander. I am a cybersecurity or an online cybersecurity instructor with more than seven years experience, and I'm also a big AI enthusiast. So what can you expect from this course? Well, let me just give you a quick summary of the curriculum. First of all, we're going to delve into what AI actually is, as well as the foundations of AI. We'll tackle the history of AI a little bit. And then we'll delve into the three main modules for this course. Starting off with machine learning. And then we'll take a look at deep learning and then natural language process. And these are the three big modules or three big sections in this course. And then to run up the course, we'll take a look at the future of AI and what we can expect from artificial intelligence within the next ten to 50 years. So a few things to know about this particular course, there are going to be quizzes at the end of each section. So please take the quizzes. Don't worry. The questions are actually not that difficult. They're just to test what you've learned. So don't panic. These are very, very easy questions. And then resources, I'm going to provide for you all the slides I will use in this course, as well as the book, which is basically a PDF book summarizing the entire course. Now, as of the time of me recording this particular video, I am still working on that book. So please do be patient. Maybe you've enrolled in the course and the book isn't ready yet. Don't worry. I am working on it. I'll let you once the book is ready to download. Also, if you do want to use the slides in any sort of formal presentation, you are welcome to do so. I only ask that you credit me, Alexander One, as well as my company labsyba.com. Thank you. And then, of course, questions if you have any questions about anything that I've covered in this course. Maybe there is something you don't quite understand. Always feel free to reach out to me. I'll be more than happy to answer all your questions. So with that being said, I want to welcome you once again to this course where we're going to talk about AI. Fundamentals of AI, the foundations of AI. And I can only hope that you will enjoy this course because I have put in great effort to make the lessons as entertaining but also as educational as possible. So welcome once again. Let's get started. 2. Section Preview Intro to Artificial Intelligence: Welcome officially to the very first module introduction to artificial intelligence. Now one thing you should know about me on a personal level is that I love watching movies. I love going to the cinema. It is one of my all time favorite hobbies. And one thing I like to do as an instructor is to incorporate movie clips into my lessons whenever I can, because I think that they can be very, very informative, but also very entertaining at the exact same time. Now, out of the hundreds and hundreds of movies from Hollywood that have artificial intelligence as the central topic, I was thinking about the perfect clip that I could use to introduce this course, and I think I may have found just the right clip, so sit back, relax, enjoy this clip, and I'll see you at the end of it. Good afternoon, Hal. How's everything going? Good afternoon, Mr. Amer. Everything is going extremely well. Hal, you have an enormous responsibility on this mission, in many ways, perhaps the greatest responsibility of any single mission element. You are the brain and central nervous system of the ship, and your responsibilities include watching over the men in hibernation. Does this ever cause you any lack of confidence? Let me put it this way, Mr. Raymer. The 9,000 series is the most reliable computer ever made. No 9,000 computer has ever made a mistake or distorted information. We are all, by any practical definition of the words, foolproof and incapable of error. Anyway, Queen Ts pawn. Okay? Bishop takes Knights pawn. Lovely move. Uh, ok to King one. I'm sorry, Frank. I think you missed it. Queen to bishop three, Bishop takes Queen, Knight takes Bishop. May. Ah. Yeah, it looks like you're right. No resign. Thank you for a very enjoyable game. Thank you. Well, come back. Hope you enjoyed that clip. Now, it was taken from the movie 2001, a Space Odyssey, and that movie was made back in the year 1968. Now, I'm making this particular video in the year 2025. So that's what 57 years ago. And the reason why I bring this up is because I want you to understand that AI as a technology or as a subject or as a topic, has been in existence for decades now. A lot of people seem to think that, Oh, AI is this new thing that only just emerged recently. That's simply not true. AI has been around for many, many years now. Going to the clip itself, what exactly did we see? We saw the introduction of a particular AI model known as the Hal the HAL 9,000. Obviously, you can see that it is an extremely sophisticated AI model because according to the presenter, he said that the Hal 9,000 is the central nervous system of the entire spaceship, and it's also responsible for the well being of the human crew members while they are in hibernation. So this is a very capable, highly advanced, highly intelligent AI model. And we also saw the AI model playing chess with one of the human crew members, and obviously it won. Now, fun fact, if you're not a chess player, I'm a chess player. I love watching chess on YouTube, as well. Today, we do have AI models known as engines that actually help professional chess players to prepare for games. These models are extremely intelligent, highly advanced. So professional chess players actually use these models to analyze chess games, to prepare for competitions, and also to prepare traps for their opponent. So I think it's actually fascinating that this particular movie, the Space Odyssey, was able to correctly predict that sometime in the future, we're going to have AI models that will be so good at chess, they'll be able to beat any human being. But going back to the clip again, we also see during the interview between the presenter and the actual AI model, the AI model said something very, very chilling, and that is, it is incapable of making any sort of mistakes. And I think that's very, very scary because it raises the legitimate concern about AI in the future. What happens when we begin to create AI models or systems that are so intelligent that are capable of independent thought, and they decide to start making decisions on their own. Now, there are a lot of skeptics who claim that this will never, ever happen. AI will never become self aware or become that intelligent. However, there are others who believe it might be a possibility. We're going to delve into this much deeper in the course. But one more thing I wanted to mention before I ran up this introduction is the topic of natural language processing NLP. If you note in the clip, the humans were having almost sort of like a very natural conversation with the AI model. The AI model, the hell 9,000 was able to understand what the humans were saying to it because of NLP. NLP natural language processing is what allows machines or AI models like the Hal 9,000 to interpret voice commands from humans and then execute a particular kind of task or function. Of course, I'm going to dedicate an entire module to talking about NLP later on in this course. So hopefully you've enjoyed this short introduction to the world of Artificial Intelligence. Let's now move on to the very next lesson. 3. What is Artificial Intelligence: What is artificial intelligence? Well, it's basically the simulation of human intelligence in machines. You can think of it as us humans trying to transfer our intelligence over to machines. Now, we do have key characteristics for artificial intelligence. We do have the ability to learn. So AI systems should be able to learn and improve over time. And then reasoning, they should be able to make logical reasoning and deductions and then eventually produce a result, and then perception, which means their ability to process sensory data like images, sounds, and so on, and then produce a result. And of course, the use of natural language processing NLP, which allows AI models to understand and interpret human language. Now, the primary goal of AI at the end of the day is to allow models to perform cognitive functions similar to that of humans. Now, we do have different types of AI. We have the narrow AI, what we call the task specific AI. It's also called the weak AI. These are AI you would find in your firewalls, in your recommendation systems, in your virtual chat boards like Alexa, and even your hagPT cloud, dipsk models, and so on. And then we have the next stage which would be the general AI which have the ability to possess human like intelligence. We're not there yet, but a lot of people believe that eventually we will be able to develop AI that's this advanced. And then in the future, super intelligent AI. A lot of people believe that we're never ever going to develop AI. They'll be that sophisticated. But yet again, we also have a few people who do believe that we will eventually get there. So AI in everyday life, we are already using AI in virtual assistant, like, you know, Alexa, Siri, your recommendation systems like on Netflix, Spotify, YouTube, your smart home devices all use AI as well. Our self driving cars, like, of course, Tesla, that's all AI. And we also have AI in healthcare with the use of medical imaging and so much more. So AI is already playing a big role in our daily lives. Now, I do want to address some common myths around the world of AI. The first one here is that AI has emotions like humans. Now, I am very guilty of this because whenever I use my favorite AI models like Cha JPT ask you to do something for me, it does it very, very well. I end up thanking hajibt I will say something like, Oh, thank you so much, Dos. You did something wonderful for me today. And, of course, hagibt will respond by saying, You're welcome. And I say these things almost sort of like subconsciously because I believe that if I'm nice to hagipti and I praise hagibt whenever it does something good for me, it's going to reward me in the future with even better results, which is, of course, not true. Hagibt cannot understand emotions. It doesn't process emotions like we humans do. Also, AI will replace all jobs. Now, there is no doubt that AI will replace a lot of jobs, but it's never going to replace all jobs. And in fact, AI will also create new jobs as well. We'll discuss this in the final module, the feature of AI. And then AI is infallible, meaning that AI cannot make mistakes. This is, of course, not true. AI can make mistakes. AI can hallucinate information. And remember, that AI is developed by human beings. So if the human beings, if they make mistakes in the coding or how the AI model is trained, the model will make mistakes. It will be prone to making mistakes, and we'll discuss this later as well and that AI can think independently. Again, most people believe that we're never going to get to the stage where AI will become self aware and will be able to make decisions on its own. But there are still some people who believe that we're going to get there eventually. I don't know if that is true or not, but most people seem to believe that it's all a myth. AI can only operate based on instructions or rules that have been given to it. And then finally, my favorite one AI will take over the world. And I put aesthetics there because when I say AI will take over the world, the myth here is AI becoming self aware and decide that, you know what? I'm going to rule over human beings. Human beings will become slaves to AI. No, I don't believe that will ever happen. However, AI, in a way, will take over the world because we will have AI in just about everything that we do from transportation to communication, to healthcare, to shopping, to creativity, to entertainment. You name it. We're going to have AI in one form or the other. Affecting the way we live our daily lives. So, I wanted to give you a quick engagement activity. Think about the very last time that you had an interaction with AI. What AI power tools do you use daily, and how do they enhance your experience? Take 5 minutes to think about this and, I think you might find the results to be very fascinating. So to round up, I just want to give you a quick summary, AI is a simulation of human like intelligence in machines, and AI is already a part of our everyday life. And there are many myths and misconceptions about AI. I've talked about them already. And AI, the ultimate goal of AI should be to enhance human capabilities and not replace them. Thank you for watching. I will see you in the next class. 4. History of Artificial Intelligence: Now briefly take a look at the history of artificial intelligence. So, AI's history dates back to thousands and thousands of years ago. And like I said earlier, contrary to what a lot of people believe, AI isn't something that just emerged. It's actually been in existence for quite a while now. Of course, just like with any kind of technology, AI has experienced both progress and setbacks. The setbacks, we call them AI winters. Now, just to give you a few examples of the early concepts of artificial beans, not necessarily artificial intelligence, but artificial beans. You have in Greek mythology the existence of a giant bronze automation called Talos. You may have seen Talos in certain movies before. And then also Chinese and Arabic automations. We had early devices that were able to mimic life. And also Leonardo Da Vinci's renaissance, he was able to sketch humanoid robots. So the birth of AI, the 20th century foundations, AI, as we know it today. So a lot of the groundwork was between the 40s and 50s. And in particular, we had Alan Twin, the very famous scientist who back in 1950, he created what's known as the Turing test. That's used to evaluate machine intelligence. Now, in 1943, we had the first neural networks designed by Walter McCulloch and Walter Pitts. And then in 1956, finally, in the Damoth Conference, the official birth of the term artificial intelligence as an academic field was coined by John McCarthy. So what are the AI winters and resurgence? Well, in 1970, we had the LA systems failing to meet expectations, leading to reduced funding. Of course, a lot of companies and countries felt like, You know what? This AI thing isn't going to work. Let's stop funding our researchers. And then 87-993, there was another setback because our computing power back then was severely limited. Keep in mind that in order to power AI models, you need plenty of computational power. And back then, computers were simply not powerful enough. But what about the revival? See, in 1980s, we had expert systems that gained traction in both business and medicine, and between the 1990s and 2000, because of advances in machine learning and the rise of the Internet, all these were able to contribute to further artificial intelligence research. So I do have some key milestones in AI development right here. In 1997, we had the IBM superpower computer Deep Blue that defeated then World champion Gary Kasparov in chess. Now, as a big chess fan, I actually followed the games. The very first time they played Gary Kasparov who, by the way, a lot of people considered to be the greatest chess player of all time, he actually defeated Deep Blue in the very first match. Went back for they came back for a rematch, and then in the rematch, Deep Blue eventually beat Gary Caspro. So that marked a significant milestone because it became the very first time in human history that a machine was able to beat a human being on the chess board. And then we had some other milestones in 2011, when the IBM Machine Watson won Jeopardy against Even Champions. And then in 2016, we had Googles off a goal. That defeated the World champion goal player Lee Sedol. And then in the twentytwenies, AI models like the Cha GPI, the emergence of models like ha GPT, Cloude and so on, have been able to demonstrate human like text generation. So AI in the 21st century, as we know it, we now have AI power in cars like your Tesla, use of natural language processing that's used by AI models like your Deep sik, clothe, ajepit and so on. And then also in healthcare, we now have AI Power diagnostics and also drug discovery. So what is the future of AI? What can we expect? Well, we can expect AI powered creativity in art, music, and also advances in human and AI collaboration. There will be more collaboration between artificial intelligence and human beings. And, of course, the use of ethical AI frameworks to ensure that AI is responsibly used, AI is responsibly trained, and that AI is, in fact, safe to use. So just a quick lesson summary, AI has existed in human imagination for centuries now. The 1956 Dermoth Conference marked the official birth of AI as an academic field. And then we've had, of course, the AI winters and then the AI resurgence. And, of course, modern AI is powered by the planning, which we will get into big data and, of course, plenty of computational power. And finally, AI's future will include ethical considerations and human AI collaboration. Thank you for watching. I'll see you in the next class. 5. Key Concepts & Terminology: Welcome back. So now let's take a look at some key concepts and terminology used in the world of AI. We're going to start off with the AI terms. There's three of them, artificial intelligence, machine learning, and, of course, deep learning. Now, when it comes to artificial intelligence, we've talked about this already. It's basically machines trying to simulate human like intelligence. You have your examples with Alexa, Civ, you know, your virtual assistant. But then we also have machine learning that's basically a subset of artificial intelligence, where the machines are able to learn based on patterns. Your exam one of the best examples here would be your spam filters in your email. Think about it, okay? Email spam filters, they're not rigid. They're able to identify what is spam, based on trends, based on patterns, based on history. So think of the machines as learning in the process. Initially, the spam filter might not do a good job in being able to catch every type of spam. But over time, as it begins to learn what exactly is spam, the different shapes and forms a spam mail might come eventually with time, it will improve. And then finally, we have your deep learning, which is a more advanced subset of machine learning that uses neural networks to process complex data. One of the best examples here would be your facial recognition systems. Now, later on in this course, we're going to delve deeply into both machine learning and deep learning. But what about the concepts, okay? There's a ton of them, and you should have an idea of what they are. The first one in here is going to be the algorithm. Basically a set of rules or steps, a machine will follow to solve either a problem or make a decision. We've seen algorithms in just about every kind of application out there, whether it's a dating application or a gaming app, basically, any kind of app or program uses an algorithm to determine how that application will process data or make a decision. We have your model. It's basically a train system that's able to make predictions based on data. Oh, your a GIPT, your Cloud, your My journey. These are all examples of AI models. You have the training data, which is the data that's used to train these models. Again, we're going to talk a bit later in the course, a bit more deeply how AI models are trained. And then we have what we call inference. This is basically the ability of an AI model to make predictions based on new data that it has acquired. Say, for example, you have an AI model that's used to make predictions in the stock market. So if something new happens, maybe, for example, a war has broken out or maybe some presidential candidate has won the election, this might impact the global stock market. So the AI here would be able to make predictions based on these new events that have occurred. That's what we refer to as inference. And then we have your neural network. It's basically a computational system inspired by the human brain used in deep learning. Again, we'll talk about deep learning later in the course. And then natural language processing NLP. This is the ability of a machine to basically understand and generate human language, text. Also, there's going to be a special section of module. Dedicated to learning more about NLP. And then finally computer vision. This is, of course, AI that enables machines to interpret images and video as well. So these are some of the key concepts that you should be aware of when it comes to AI. Now, types of AI, I already talked about this earlier. You do have your narrow AI that's used for very specific kinds of tasks. An example here would be your Google Translate, your Cha JBT. Basically, kind of like the narrow AI. But then you also have your general AI, the strong AI, AI that can perform any intellectual task a human can do. It's still kind of theoretical at this point. It's not yet been developed. And then, of course, the super intelligent AI, AI that surpasses human intelligence. It's a future concept. So even argue that it will never get to that point while people who believe that we will get super intelligent AI, they believe that it will take decades for that to occur. And you should also understand the difference between automation and AI. See, when it comes to automation, it follows predefined rules. So for example, the customer buys product A. Since customer has bough product A, send customer 25% coupon to buy product B, you know, kind of like that, right? Automation relies on rules and triggers. When it comes to AI, it doesn't follow rules. AI is basically able to learn and make decisions on its own. For example, yourself, driving cars. So just to give you a quick lesson summary, AI includes machine learning and deep learning as subfields, algorithms, your models, your training data. These are all core components of AI. Now, AI can be narrow, specific tasks, general AI, like, human like or super AI beyond human capabilities, and then automation and AI are different, but AI can, in fact, enhance automation. Thank you for watching the video. I'll see you in the next class. 6. AI vs Machine Learning vs Deep Learning: Come back. So now let's take a closer look at the differences between artificial intelligence, machine learning, and, of course, deep learning. So AI, artificial intelligence is the broadest concept, okay? While machine learning would be a subset of artificial intelligence, while deep learning is a more advanced subset of machine learning. So think of it this way, okay? At the very top, we have AI. Just below AI, we have machine learning, and then just below machine learning, we have deep learning. Now, I have provided an analogy in here, okay? Think of AI as the entire universe, right? Machine learning would be like a galaxy in that universe, while deep learning would be a solar system inside of the machine learning galaxy. So to kind of round this up, all deep learning is a subset of machine learning, but not all artificial intelligence is machine learning. Keep in mind, there's a lot more to artificial intelligence than just machine learning. So just to recap that again, all deep learning falls under machine learning, but not all of artificial intelligence is machine learning. Okay. So what is AI? We've talked about this already? Basically, machines simulating human like intelligence. And, of course, AI is able to perform tasks, solve them, make decisions, things like that. So AI, you should know, doesn't always learn. AI can also follow our predefined rules as well. Now, machine learning, it's basically a subset of artificial intelligence that would allow machines to learn from data and patterns. And of course, this would allow that machine to make predictions and solve problems over time. There are three ways how machine learning is done. The first will be what we call supervised learning, where the training data that is given to the machine is actually labeled. Imagine you're trying to train a spam filter. So under supervised learning, the machine will be given different types of spam emails, and the data will be labeled, Okay, this is spam, this is spam, this is Spam. So over time, when the machine has seen all these examples of the different types of spam emails, it will be able to learn and make predictions in the future of whether or not a particular kind of email is spam or legitimate. And then we have what we call the unsupervised learning, where the training data isn't labeled and the AI model learns how to make predictions based on these audib data. Basically, it tries to find patterns. One of the best examples here would be in customer segmentation. And then the last one is what we call the reinforcement learning. Think of it as a rewards penalties kind of system, where when the AI, when the model makes the right kind of prediction or is able to solve a problem or gives the right answer, it will be rewarded. But when it makes a mistake, it will be penalized. So that's what we refer to as the reinforcement learning. So we've seen machine learning in many examples, your spam filters, your YouTube Netflix recommendations, and even fraud detection in banking as well. So machine learning relies heavily on algorithms to identify patterns and make predictions. But what about deep learning? It is a subset of machine learning that will use artificial neural networks to process large amounts of complex data. So it uses multiple layers of nuance. We call them the deep neural networks. Works best with large datasets and high competitional power. In other words, you need very powerful computers to run deep learning, and of course, it enables AI to perform human like tasks such as your speech recognition, facial recognition, and so much more. Now, we have seen deep learning in several examples. For example, your facial recognition in your smartphones, fingerprint detection as well, self driving cars like your Teslas, and, of course, in your AI models like TajiPT and so on. Now, deep learning is the most advanced AI technique that allows machines to basically mimic human brain functions. So I've given you the table in here to highlight the key differences between AI ML and DL. We have the features in the definition, like we've said, AI basically is machines that mimic human intelligence. Machine learning learns from data and patterns, while deep learning Wood is an advanced ML subset that uses neural networks. When it comes to the data dependency, now, AI by itself may not require data because AI can also just learn on its own. However, with machine learning, it requires structured data. With deep learning, it requires, large amounts of data sets. Examples, of course, under your AI, we have your chat boards, virtual assistance. For your machine learning, we have your spam filters, recommended systems, for your deep learning, your self driving cars, as well as your facial recognition. And then the last feature, the complexity. This is very, very interesting. Now, with AI, it's a broad field. Includes rule based artificial intelligence. And then for machine learning, simpler algorithms, but it requires training. And then for deep learning, it's highly complex and requires very powerful hardware. So these are some of the key differences among these three terms. So I wanted to give you a real world example of how these three come together to power something very powerful. So say, for example, your Tesla, a self driving car. Artificial intelligence, basically, will enable the car to make the decisions, okay? So because of AI, the car knows that it probably shouldn't be speeding at a highly populated area as an example. But then with machine learning, because remember, machine learning requires patterns and data to learn from. With machine learning, the car might be able to make predictions on what the traffic is going to be like at a certain period of time during the day. It might also be able to make predictions based on what the weather might be like, things like that because of machine learning. And then deep learning is what would allow the car to be able to interpret road signs, traffic signs or even recognize pedestrians trying to cross the road. So when you combine all three, think of it this way, right, AI is basically kind of like the overall system of the self driving car. The machine learning is what would allow the car to improve and learn over time, while deep learning will make the car highly efficient. So this is how these three come together to power your Tesla or other self driving cars out there. So quick listen summary, AI is the broad field ML is, of course, the subset of AI. That would enable machines to learn from data while DL is the very advanced subset of ML that uses neural networks for advanced learning. And of course, AI, ML, and DL are interrelated, but of course, have distinct differences. Thank you for watching. I will see you in the next class. 7. Section Preview The Foundations of Artificial Intelligence: I'm to motel to the foundations of artificial intelligence, and it is time for another movie clip. And something tells me that you've probably seen the movie where this clip is going to be taken from. Nevertheless, sit back relax. Enjoy the clip, and I'll see you at the end of it. Right now, we're inside a computer program. Is it really so hard to believe? Your clothes are different. The plugs in your arms and head are gone. Your hair has changed. Your appearance now is what we call residual self image. It is the mental projection of your digital self. This this isn't real. What is real? How do you define real? If you're talking about what you can feel, what you can smell, what you can taste and see, then real is simply electrical signals interpreted by your brain. This is the world that you know. The world as it was at the end of the 20th century. It exists now only as part of a neural interactive simulation that we call the matrix. You've been living in the dream world, Neo. This is the world as it exists today. Okay, welcome back. And, of course, that clip was taken from the very popular movie the Matrix released in the 1999. And if for some reason, you've never seen this movie before, what are you doing with your life? Stop watching this course and go watch the movie. Now, I'm kidding, finish this course first, and then you can go and watch the movie. But seriously, though, the Matrix, in my humble opinion, is one of the greatest movies ever made. It raises so many interesting questions at the action, and it's a great movie. You simply have to watch it. Now, why did I choose to use this particular clip? Well, because it raises a ton of fascinating questions. I'm going to tackle two of them. First of all, let me describe what happens in the scene. You have Mofius, the man with the dark shades. He is explaining to Neo, the other guy that, Hey, this world we're in right now, this is all virtual reality. It's not real. It is fake. It's generated by a very powerful AI system known as the matrix. Now, Neo is obviously very surprised. He's shocked. He's like, No, how can this be? No, this is real. This can be fake. He touches the chair, and Morpheus asks him, how do you define what is real? And I thought that's a very, very fascinating question. How do you define what is real? And the reason why this is fascinating is because today, we have AI generated videos, images, deep fakes. And even though today we can to a very large extent tell what is real and what is AI generated, think about it. In a few years to come, the kinds of content that AI will be able to generate will be so realistic that we might not be able to distinguish between what is actually real and what is generated by AI. We might need certain kinds of systems or scanners or algorithms to help us detect whether or not that image or that video we're watching is actually real or fake. Think about it. So it's very, very fascinating. How are we going to be able to define what is a real image and what is an AI generated image? Another question here, though, is in the clip, we see the influence that the matrix now has over the human population, right? The matrix is very powerful. It's been able to create this virtual reality, so it has a lot of influence on human beings, right? Now, I know the matrix is an extreme example of AI's influence over human beings. But think about it today, Today, believe it or not, AI already has some influence on our daily lives. You don't believe me, when you go on YouTube or you go on Netflix or Spotify or any one of these platforms, you always have this recommendation tabs or systems, right, that will recommend content to you, based on your history, based on your search results, and sometimes recommendations might even just be random. But think about it. Those recommendations already begin to influence the way we think. It might influence us to begin to support a particular political party or candidate. It might begin to influence the way we buy things. It might begin to influence the way we think about certain kinds of controversial topics and so on. So in a way, these recommendation systems that are powered by AI are already beginning to have some level of influence on how we live our daily life. So that begs the question how much more influence will AI begin to have on the way we live our lives as it becomes more and more intelligent? Because, believe it or not, whether I like it or not, AI is going to be introduced to just about every part of our lives, whether it's communication, transportation, shopping, entertainment, creativity, AI is going to come everywhere. So just imagine the level of influence that AI is going to have over us in the near future. Anyway, let's move on to the next lesson where we're going to talk about the actual foundations of AI. I'll see you there. 8. Foundations of AI: Now take a look at a very, very important topic, and that's going to be the foundations of artificial intelligence. Now, contrary to what a lot of people might believe, AI isn't just limited to the tech field. It actually cuts across multiple disciplines. As an example, you'll have AI obviously in computer science, where the use of algorithms, data structures, programming code, and so on used in AI then you also have AI in the field of mathematics and statistics. Don't forget that we can use AI for mathematical calculations, for calculating probabilities and so on. And then you also have AI in the field of cognitive and neuroscience. Because think about it, right? In order to be able to develop artificial intelligence that's meant to mimic human intelligence, we first need to understand human intelligence in the first place. And then we also have AI in the field of linguistics. This is, of course, essential for natural language processing. And perhaps very surprising, you'll have AI in the field of philosophy and ethics because think about it, right? The major challenges involving AI revolve around ethics, privacy, and whether or not it's actually moral to use AI. So contrary to what a lot of people might think about AI just being a tech field, you have AI in multiple disciplines as well. Now, let's take a look at the core AI principles. There's six of them. I want to take a look at them one by one, starting off with the logic and decision making. Many AI powered models, they rely on logic in order to make the decisions. For example, you have your bullying logic that uses operators like your N or naught. For example, if A is equal to B, B equals to C, A must be equal to C, something like that, right? And also the rule based system where you have your E then L statements. For example, if weather is rainy, then take umbrella. Else, if weather isn't raining, then don't take umbrella, you know, things like that, right? And then also expert systems. These are AI that would mimic human expertise by following predefined rules. You have them in your medical diagnostics. We do have certain limitations, though, involving the logic and decision making, and that's because AI can struggle with uncertainty, and whenever complex decisions need to be made to keep this in mind, next principle in here would be the principle of probability and uncertainty. We've talked about the role of big data and how important data is to artificial intelligence. But there are many situations where an AI model may need to make decisions based on incomplete or noisy data. As an example, in your Baysia networks, you have AI used for probabilistic reasoning helping machines make educated guesses, for example, in your spam filters, but then you also have them in your markov decision processes, your MDP, where AI is used for decision making in certain environments like your robotics, finance, and so on. And then also in your Monte Carlo simulations, that's typically used for risk analysis and gaming. The next concept in here is the optimization and learning. Of course, AI is constantly learning and optimizing at the same time. So AI can make use of optimization algorithms that would help the AI adjust its parameters to minimize errors. An example in here would be the gradient descent used in ML your machine learning to fine tune models by reducing prediction errors. And then the concept of linear programming where the AI is taught resource allocation and saddling tasks, and then evolutionary algorithms. This is inspired by natural selection. These algorithms evolve solutions over time. An example in here would be your genetic algorithms, your GA that optimize solutions for complex problems. Now, one of AI's greatest strengths involves the ability to recognize patterns and learn from those patterns over time. And one of the key concepts in here involves the neural networks that you find in your deep learning. So in here, the AI mimics the human brain to recognize patterns in data. As an example in here, your image recognition systems, that can detect objects in photos. Also the concept of feature extraction where the AI is able to break down the data into key features. For example, in your voice assistant, the AI model can break down the voice into several segments and can then understand commands based on those segments. And then also in the concept of clustering and classification where the AI can group data into meaningful categories or a in your customer segmentation and so on. And then reinforcement learning. We've talked about this a bit earlier. This is the trial and error or the rewards and penalties kind of training for machines. So AI can learn through trial and error, much like how humans also do learn from experience, right? So we do have the reward based learning where the AI can receive rewards or penalties based on how well it performs in a test or in an exam, and then exploration and exploitation. So AI must balance train new strategies against using known ones. So the AI model needs to kind of, like, find a fine balance between both and not rely extensively on either one of them. As an example, your self driving cars like your Tesla, they learned the best driving actions by being rewarded for safe behavior. And then also your chess Alpha zero, a very, very powerful chess model. It learned how to play chess expertly well by simply playing millions of games against itself. Now, heuristics and approximate solutions. This is very, very interesting. So over here, sometimes, being able to find the exact solution could be impractical, based on the problem based on the challenge being provided to the AI. So the AI needs to make use of heuristic, what we call intelligent shortcuts. So we do have the heuristic search algorithms. For example, the AI will simply search for the good enough solutions faster, say a star algorithm for path finding in maps. Sometimes when they're trying to search for something, say for example, in your Google search engine, you may not provide the exact kind of terms that you're looking for, but the AI needs to be able to make a calculator guess what it is you're trying to find, and then the AI will simply provide the best results. And then fuzzy logic, what exactly is this? So, AI makes decisions based on the grades of truth instead of binary choices. As an example, AI in your air conditioning systems, they are just temperature based on comfort levels. So it will try to make its decision based on what it thinks would be a reasonable temperature for the human being and not exactly a binary choice of whether to turn on the air condition system or power itself off, if that makes sense. And then I provided the table in here for the core AI principles, again, the description and then example. So you can pause this video and go through them at your leisure time if you're interested. Then the key scientific concepts that are used in artificial intelligence, linear algebra that's used in your deep learning networks. You have your probability and statistics. Of course, AI needs to make use of this to make predictions, and then your neural networks used specifically in deep planning that tries to mimic the human brain, and then the genetic algorithms inspired through evolution. And then finally game theory. This is, of course, decision making in competitive environments. So what is AI's connection to cognitive science? Well, AI models human condition in areas like its perception, reasoning and problem solving. And then the study of human intelligence can also help to improve artificial intelligence by itself. So as an example, in your reinforcement learning, it is inspired by behavioral psychology, which is, of course, reward based learning. So these are some of the ways how AI is actually associated with the field of cognitive science. So what are the key takeaways from this lesson? Well, first of all, AI is, in fact, a multidisciplinary field involving computer science, math, cognitive science, philosophy, and so much more. AI operates on logic, ability, optimization, and, of course, learning. And then data, as I've said before, is the fuel. It is the bloodline of artificial intelligence, and different types of AI systems will use different approaches. And then understanding these fundamental scientific concepts will help in grasping how AI works at a much deeper level. Thank you for watching the lesson. I will see you in the next class. 9. The Role of Data in Artificial Intelligence: Now talk about the role of data in artificial intelligence. Now, I like to think of data as the lifeblood of AI models because without data, AI models will not exist or they'll be very inefficient. So why is data crucial for AI? AI models are only as good as the data they have been trained on. So the more high quality data available, the better AI will be able to learn patterns and make predictions, improve accuracy and efficiency over time, and of course, adapt to new situations and refine its decision making. And pretty sure in computer science, you must have heard of the term garbage in, garbage out. Basically, if a program has been designed to make mistakes or not solve problems accurately, then that's exactly what the program is going to do. And that's kind of similar with artificial intelligence models as well. If they've been trained on very bad data, then guess what? That artificial intelligence model is probably not going to be intelligent. It's going to be intelligent. That's why the quality of data used to train models is extremely important. Now, these are the types of data that we use in AI. We have your structured data. For example, this would be data that's been organized into tables, rows and columns. So for example, you have your data from your spreadsheets, Excel files, databases, and so on. But we also have the unstructured data, which is basically raw data that doesn't fit a fixed format. So examples here would be your images, your videos, your audio, and so on. And then the last kind of data will be your semi structured data. What exactly is this? Well, it's basically data that falls in between your structured data. And your unstructured data. So examples here would be your JCNFles, XML, sensor logs, and so on. And when it comes to data processing in AI, there are four main steps. The very first step would be the actual collection of the data in the first place. So the AI will collect data from a wide variety of sources like the Internet, your databases, user feedback, and so on. And then once that data has been collected, the data needs to be cleaned up. So in here, the AI will try to remove, for example, duplicates of records that might already existed. So say, for example, customer records. If the AI finds out that, Oh, this particular customer has two exact same records in our database, just go ahead and remove one of them. So that's basically the next process data cleaning. And then after that, we have the data labeling. So in here, your data can either be categorized or could be tagged. So example would be tagging emails as either spam or not spam. And then the final step would be your data transformation where the data could be converted into usable formats like say, your PDF files, Excel sheets, and so on. So those are the four stages of data processing in AI. Now, when it comes to big data itself, there's four features we need to be aware of the four Vs. The first one here would be volume. Okay? So basically, the bigger the volume, the better. The more data you're able to train your AI model on, the better it is going to be. Next one in here would be the velocity, the speed at which new data is generated. And of course, in the world we live in today, that is extremely fast. Next would be the variety, the different types of data that the AI model is trained on, whether it's audio, video, images, text, files, you name it. And then the last, possibly the most important veracity, how accurate is the actual data itself? Obviously, it's not going to matter if the volume is so large and homod variety. If the veracity is poor, that data is basically going to be useless. That's why, in my humble opinion, I think out of the four Vs, veracity is going to be the most important. So we do have several data challenges when it comes to AI. We have data bias where an AI model could be trained on biased data. And because of that, the AI begins to make certain kinds of decisions, and it may lead to unfair or discriminatory decisions. We have data privacy as well because an AI model needs to be trained on large amounts of data. There is the possibility that sensitive or private information may be fed into the AI in order to train it. And of course, this will raise concerns about security and ethics. We have the data quality. Again, very, very important. How how much of good quality is the data that's been used to train the AI model? So if the data is poor or is of low quality, this could lead to the AI making poor decisions, and then of course, data storage and management. So AI requires massive storage capacity, and the impact here is that it's going to require efficient data handling as well. So it's not that easy. Now, going a little bit deeper, we do have the ethical and privacy concerns when it comes to AI data. So user content and privacy, AI should not collect or use personal data without consent. This is what we like to believe in, this is what we hope would be the case with AI, but you never really know there's always that concern regarding the use of AI. And then, of course, bias and fairness. Again, if the data if the AI model, excuse me, has been trained on biased data, then the AI could make discriminatory decisions, and then transparency, very, very important. Users should know how the AI systems use the data and possibly maybe even important, the AI should be able to explain its decision on why it did something a certain kind of way, transparency, very, very, very important. So for example, in facial recognition, AI has been criticized for racial bias due to biased training data sets. We'll talk about this a bit later. And when it comes to AI in hiring, as well. So AI based hiring systems have been found to discriminate against certain groups if trained on biased data, again, how efficient an AI model is will depend largely on the quality of data it has been trained on. So we do have the process by which AI is able to improve its decision making through data. So the very first phase here in the loop would be the actual training phase where the AI learns from either history or data that's been fed to it. And then the AI will now be able to make predictions based on what it has learned. And then when it makes a prediction, the users will be able to provide feedback to the AI or even the developers, they'll be able to tell the AI that, Hey, you got the answer correctly or the prediction you made was actually false. And because the AI has gotten the feedback from the user, it then goes through a re training phase again to learn based on the new feedback that user has given it. So that's how it kind of goes through this constant loop of trying to improve. So what is the future of data in AI? Federated learning where AI models will train on user data without transferring the data to a central server, thereby improving privacy, and then synthetic data where the AI itself might be able to generate artificial datasets. I real datasets aren't available. And then the last one in here, the explainable AI DX AI very, very important. Where the AI should be able to explain the reason why it made certain kinds of decisions. This would be a giant step forward regarding transparency in the use of AI. So some key takeaways, data is the foundation of AI. Without it, AI cannot learn or function effectively. Of course, AI could use structured, unstructured or semi structured to learn. And improve. Big data, of course, enhances AI's performance, but you do have challenges like your, you know, privacy, bias, and so on. And then ethical considerations are important when handling data in AI systems. And then finally, AI continuously improves through our feedback loops. So that's a thank you for watching the video. I will see you in the next class. 10. Algorithms & Models in Artificial Intelligence: Welcome back. So now let's take a look at the different types of algorithms and models used in the world of AI. But first of all, what exactly is an algorithm? What is a model? Algorithms are basically predefined rules that an AI model can use to process data. But AI models themselves, these are simply trained versions of algorithms that are able to make decisions and maybe even predictions based on new data. So over time, AI models will improve because they're constantly learning based on history, based on user feedback, and so much more. As an example, your spam filter, over time, the filter will get better and better because it's able to learn from previous spam emails and maybe even emails that it incorrectly identified as spam. Over time, it will improve. But what are the types of algorithms that we have? We do have the rule based or your symbolic AI that uses your predefined rules and logical conditions. So basically, it's quite rigid, is yes, no, is no, yes, cannot be no, you know, stuff like that. So it will work well with very structured problems, but these kinds of algorithms will struggle with uncertainty. One of the best cases where these kinds of algorithms are used would be in your medical diagnostics. Another type of algorithm would be, of course, the deep learning and neural networks. These use multi layered artificial neural networks, and of course, they excel in complex tasks like your speech recognition, your facial recognition, and so on. And of course, a gepty and image recognition AI, they use deep learning to be able to identify images and also generate text as well. Now, machine learning, we've talked about this already. They learn patterns from data instead of following strict rules. So these kinds of algorithms are a bit more they're less rigid in their approach. So we do have the different learning types. You have your supervised learning, unsupervised, and of course, reinforcement learning. We've talked about them already. And then the common AI algorithms and the application. So as an example, your decision trees, this is an example of algorithm. Use supervised learning, and you have them in your fraud detection, your medical diagnosis. You have what we call the support vector machines, your SVMs. They use supervised learning as well. You can find them in text classification, handwritten recognition, and so much more. And then K means clustering. These use unsupervised learning. You'll find them mostly in market segmentation, anomaly detection. You have your neural networks, deep learning. Of course, this will use supervised learning, and you have them in your image recognition, speech to text, and so on, and then your genetic algorithms that use optimization and these are AI driven design evolutionary computing. So what exactly is the training process for an AI models typically six steps. The first step is always data collection. Again, I've said this many times before, data is the lifeblood of an AI model. The AI model needs the data to begin to learn to begin to train. So once the data has been gathered, we now have data preprocessing where the data will be cleaned up and will be formatted. And then after that, the model begins to train based on the data that it has been provided. And then the model will now be evaluated. It will be tested with new kinds of data. So it could be like an exam, a test just to see how well the AI will perform. And, of course, if it does well, the AI is then deployed into the real world. And of course, over time, the AI will constantly improve because it's able to adapt and learn over time. So with all this in mind, how exactly is an algorithm chosen? There's different types. So how do developers, how the companies decide which algorithm to use when trying to train their model. So there are several factors involved in here. First, of course, will be the data availability. Some models like your deeper learning, they need large datasets while decision trees an example, work better with small amounts of data. And then the accuracy that is required. Obviously, if you want an AI model to be able to make quite good accurate predictions, then you might be looking at neural networks that are able to provide higher accuracy. But of course, this will require more computing power as well. And then interpretability. So decision trees are very easy to understand, while deep learning networks can be quite complex to understand. And there are also situations where a particular kind of AI powered system might use one or more algorithm. So for example, in your bank for detection system, it could use a decision tree for interpretability, and then it could use deep learning to be able to recognize a complex pattern detection. So what are the challenges involved in AI algorithms and models? We've talked about the bias in training data where the data are being used to train the algorithm or the model could be of low quality. This will, of course, lead to the model making bad decisions, making lots of errors, and then computonal power, especially when the deep learning algorithm is required. This will, of course, means a lot of powerful computing resources will be needed, and then explainability, as well, complex models like neural networks. They work fantastic, they're extremely powerful, but they can be quite difficult to understand and how to explain. So what are the future trends in the artificial intelligence models as years go by? Explainable AI. We've talked about this already, where AI will be able to explain the decisions that it has made. This will, of course, improve transparency. And then hybrid AI models. This is a very, very interesting concept where we can combine different AI approaches for better performance. And then we also have the edge artificial intelligence. Basically AI models that will run on small devices like smartphones for real time process. And these are some of the trends that we can look forward to in the world of AI. So just to round up, a few key takeaways, first of all, AI uses rule based machine learning and dippling algorithms for decision making. Supervised, unsupervised, and reinforcement learning are key machine learning types. AI models must go through, of course, training, evaluation, and continuous learning in order to improve. And then finally, choosing the right algorithm will depend on accuracy, data availability, and computonal requirements. Thank you so much for watching the video. I will see you in the next class. 11. AI Capabilities & Limitations: So before we round up this module, I wanted us to take a look at the capabilities and limitations of AI, as we know it, starting off with the capabilities, what exactly can AI do? Now, I have listed several capabilities in here, as well as the description, as well as the examples. I'm going to go through a few of them. Let's start off with the automation of repetitive tasks. Now, this is one area where AI has excelled in whether it's in finance or datasets or in my field of cybersecurity, we now use AI to perform repetitive tasks. So it can perform repetitive tasks with high accuracy and speed, and we see that with chat boards, automated, customer support, and so on. And then when it comes to predictive analysis, AI is able to forecast future trends based on past data. So in that scenario, it is very, very important that the AI is given the right kind of data with which it can make fairly to quite good accurate predictions for the future. Now, we've seen this in, like, stock market predictions, as well as the weather forecasting, things like that. And then when it comes to image and speech recognition, AI is able to identify text, object, and sounds. It's not perfect yet, but we're getting there. I think AI today does a very good job of being able to identify these. And we can see examples in self driving cars in your security surveillance, and so on. And finally, when it comes to robotics and automation, AI can power robots for precision pased work, and examples, we have them in healthcare with robotic surgery, as well as in industrial robots. So you can, of course, take a look at the slide which I'll present to you and look at the other capabilities in the but I've also provided a table listing out the limitations what AI cannot do. So as an example, lack true understanding. So AI, it can process data, but you see, it doesn't understand data like you and I do. Hopefully, I'm speaking to a human being, as well. I'm just joking. So that's the thing, yeah, AI, it says the data. It can handle the data. It can process the data, but it doesn't actually understand what the data actually is. It's just processing the data and giving us results, right? And then when it comes to true creativity and innovation, AI, it can generate content, but it's always going to lack the human creativity, the human touch. So dependence on data as well, AI still relies heavily on very large data sets. And of course, you're talking about the high computtonal costs. So training very complex AI models requires very, very expensive hardware. And, of course, the general intelligence limitations, yes, we can have AI that can beat human beings at chess, but it still fails at common sense or reasoning. Again, you can take a look at the slide for the other limitations. So just to summarize the strength of AI, AI can handle big data. In fact, it Excels, I thrives when presented with big data. And of course, speed and efficiency. AI has outperformed humans in computional tasks and it's not even funny. And then, of course, 27 availability. Hagibty will never tell you that it needs a lunch break or that it needs to sleep. That's never going to happen. And then of course, scalability AI solutions can be deployed globally very, very quickly without any sort of human limitation. So the summary for the weaknesses, and I know this may sound very, very harsh, but AI still lacks common sense. If you put AI in an unexpected or sort of unusual situation, it will struggle because it struggles with ambiguity. Now, AI cannot replace human judgment, okay? So remember that AI no matter how intelligent it might become, it doesn't have a heart. It doesn't have emotions. AI cannot sympathize, it cannot empathize. I cannot get angry or sad or happy. And as such, it will lack ethical reasoning. And, of course, AI depends on training data. The volume and the quality of the training data will determine just how efficient the AI model actually is. So the comparison in here, human versus AI I have the table of ideas. So creativity. Humans of course, very, very creative. AI's limited. Uh, decision making, AI will make its decision based purely on its ability to recognize patterns. While we as humans, we can make decisions based on experience, intuition, and also the past. And then learning ability, we can learn new skills very, very flexibly. That's what our brain is for. While AI needs retraining for new skills, and then bias, of course, we can be influenced by emotions, while AI can only inherit biases from the data it's been given. If the data is free of any bias, then there is no way the AI will become biased. And, of course, processing speedway a lot slower than AI, while our AI is much, much faster than us when it comes to our structured tasks. So key takeaways, AI is extremely powerful, but of course, it does have its limitations. It excels at automation and prediction, but it's always going to lack human level reason. And then AI depends very largely on data and algorithms. And so poor data will lead to inaccurate results. Bias data will lead to the AI making bias decisions, and then AI will never be able to replace human judgment. And the future of AI, we hope will include more ethical, explainable, and also adaptable models. So that's pretty much the summary. Thank you for taking this particular lesson. I will see you in the next class. 12. Section Preview Machine Learning Basics: Welcome to the next module, machine learning basics. And as usual, I'm gonna play you a clip from a movie, so sit back relax, enjoy the clip, and I'll see you at the end of it. There's six up. There's no way you can win that game. I know that. It doesn't hasn't learned. Is rn go to make it play itself? Spin Learn, God damn it. A. Christ James Game, the only winning movie is not today. Okay, welcome back. Now, that clip was taken from the movie War Games released in the year 1983. And the reason why I chose to use this particular clip is because it perfectly demonstrates one of the key concepts as to how AI models learn, and that's through the process of trial and error. Now you'll observe that in the clip, AI system known as Joshua, it plays a game of tick tack toe against itself. The game ends in a stalemt and then it continues to play the games against itself over and over and over again. But observe that each time it changes its strategy, but the games keep ending in a stale mit. Eventually, it now decides that, you know what, before I launch the nuclear missiles. Oh, by the way, I should have given you some context. In the scene, it's supposed to be a very grave scene because the system, Joshua, the AI model, it's taking control or it's about to take control of the United States nuclear missiles weapon system. And, of course, people are panicking. They're afraid that it's going to launch nuclear missiles against the Soviet Union. Sovnia will respond, and, of course, all of us, we're all going to die. So Joshua decides that, you know what? Before I launch these nuclear missiles, maybe I should try different strategies to see who would actually win in a nuclear war. So the first strategy it launches missiles from the United States to the Soviet Union, but then it realizes that, Okay, this strategy isn't going to work because there isn't going to be a winner. It then tries another strategy where it is the Soviet Union that launches the missiles first. But then it realize that, Okay, this strategy also doesn't work. There is no winner. And then it tries hundreds of different other strategies. And each time it realizes that there isn't going to be a winner. So eventually, it's able to conclude that the best way to win this particular game is not to play at all. So the AI, Joshua, was able to teach itself. It learned through the process of trial and error. Each time it tried a strategy and the result was negative, it went back, readjusted its technique, readjusted its strategy, and then tried again. And then when the result was the same negative stalemate went back again, we find a strategy one more time, and that's basically how AI systems learn. That's how they teach themselves. They try something. Oh, the answer wasn't correct. Let me go back. Let me try a different strategy. Oh, that doesn't work. Let me go back. Let me try a different strategy and so on. So the concept of trial and error, I believe, was well emphasized in this particular clip. So I hope you enjoyed this introduction to the world of machine and Basics. Let's now move on to the very next lesson. 13. How Machines Learn in Practice: Alright, so now let's take a look at the next lesson, we're going to talk about how machines learn in practice, how are they trained. So the whole machine learning process involves three key components. You have the data, the models, and, of course, the actual training itself. The other thing about machine learning is that unlike traditional programming, where the code is explicitly written and the programs have to perform exactly how they are coded, with machines, they don't memorize things. Instead, they're able to find patterns within the data that they're working with. So in other words, machine learning involves a much more flexible process by which the models are able to learn and improve over time. Now, the whole learning phase is all about the machine or the model trying to adjust or readjust its processes so that it can get better with time. Now, just like with normal students, I'm pretty sure you have had exams in the past before, before you took that exam back in college or in high school or wherever, I'm pretty sure you must have studied first, right? Maybe you took a course, an online course, or maybe you read a textbook, right? Now, as you were learning and preparing for the exam, I'm pretty sure you took certain kinds of quizzes, tests, and then ultimately you went for the final exam, and then of course, you passed. It's kind of similar with machine learning as well. First, they provided with data that they will study. And then there will be this particular stage where they'll be tested just to see if they're actually learning correctly in the right way. And then ultimately, they will be tested with new types of data that they've never seen before just to test how they'll perform in the real world. So it's kind of similar with machine learning as well. Now, the key idea here is that machine learning does not involve memorizing data. Imagine a spam filter, right. Imagine you're trying to build artificial intelligence for a spam filter. There is only so many types of spam filters that the machine can memorize. Okay, this is a spam email. This is a spam email. This is a spam email and so on. But then what if that particular model is presented with a new type of spam email that's maybe presented in a slightly different way? The model is going to fail because it hasn't memorized this new particular version of the spam mail. That's why it's always better that machines and models learn through the process of identifying patterns, okay, they're more flexible that way. So there is this thing called feature selection, right? Feature selection involves the machine or the model simply looking for the most important parts of the data. So what exactly are features? There are relevant pieces of information that would help the model or machine make decisions, right? So choosing the right kinds of features will help the model improve in its accuracy. And of course, removing any unnecessary kinds of features will also minimize errors and will also help the model improve on its accuracy. Now, I've given an analogy in here involving the sale of a house, right let's say, for example, you are trying to build a model or an artificial intelligence that can make predictions on how much a house would cost. The right kind of features you'll be talking about here would be what location, okay? Where is the house located? And then let's talk about how big the house is, okay, how many bedrooms it has. These are the most important kinds of features that the model should be trained on. The unnecessary features here would be things like what's the color of the TV in the bedroom, right? Like, things like that, what's the color of the floor in the bathroom? Like, these are very, very unnecessary features that will not help the model, make a good prediction. On how much the house would actually cost. So there is also this process called optimization. Optimization is where the models will adjust over time to improve the levels of accuracy. Now, there is a key concept here called the cost function. The cost function is basically the value that determines how far off the model was for making the actual prediction. So going back to the sale of the house, right? Imagine the model predicted that, okay, this house is going to be sold for $500,000. But imagine if the house was eventually sold for $750,000. The cost function here would be, of course, $250,000. So over time, through the process of optimization, the model is going to get better and better and better and try to reduce the cost function. So that maybe in the next time, instead of 500, it might say 700,000, the house gets sold again for 750, but now the cost function is only 50,000. So over time, the model has actually improved through the process of optimization. So how does the model actually improve well? It's going to update its internal parameters to reduce the prediction errors, and of course, it's going to learn from its mistakes as well. These are the ways how the models will improve over time. There is also this technique called the gradient descent. This is the actual learning process itself, and it's a technique that will help the model adjust a little by little step by step. The whole idea here is that whenever the model is trying to improve and reduce its cost function, it's not going to take giant steps. It's not going to it's not like it's going to undergo massive improvement very, very, very quickly. It takes time, steady, steady, steady, right? So, think of it as finding the lowest point on a mountain. The model will keep adjusting itself until it finds the best settings. So I've given you here an example involving the spam filter once again. So the spam filter predicts spam for email. But let's say, for example, it's actually not spam, it's a real email. So over time, the model will realize that, Oh, okay, so these kinds of emails that I have labeled as spam before, I now know they're no longer spam. They're actually write emails, the legitimate emails. But these are the ones, okay, I now know this is actually spam, and that's how step by step, gradually the model will improve in its ability to determine whether an email is spam or not. So the learning rates, it's very, very important that the model actually finds the right balance during the learning phase because if it's learning too quickly, it will never find the right answers and the right solutions. But then if it's learning too slowly, it's going to take forever to get the right answer. Think of it as you're trying to turn the frequency knob on a radio, right? Maybe you are trying to find the frequency of your favorite radio station. If you turn the knob way too quickly, you're never going to find the actual frequency because you're being too fast. But now imagine if you are turning the knob very slowly, very slowly, it's going to take forever for you to find the right frequency. So there has to be the right balance with the learning right for the model. So why do some models learn better than others? Well, it all comes down to the quality of the data. Remember that data is so important when it comes to training models and machines. So the quality and the quantity of data is going to be will play a key role in here. And then feature selection, of course, the models need to be trained on how to find the right features when trying to work with data or make predictions or find answers to problems. And then the model complexity as well. Imagine if you were building a very, very simple model for machine learning or AI. Well, there's only so many tasks it's going to be able to perform. It might not be able to perform complex tasks. But now imagine if you actually made a model that's way too complex, then the roles or the kinds of tasks you give the model to perform it might just be a bit of a waste because the model was built for something far more complex. And then the optimization efficiency as well, a well tuned model will train faster and will generalize much better. So these are the key factors as to why certain kinds of models are performed much better and are trained much better than our other models. So some key takeaways in here, first of all, your machine learning models learn by adjusting perimeters to minimize errors, training, validation or testing. They ensure the models will generalize well. And then feature selection, like I said, is very, very important. It's crucial. Garbage in, garbage out. If the models are trained with the wrong kinds of features, they'll perform very, very poorly. And then optimization techniques like your grading descent will improve model accuracy over time, and of course, a balanced learning weight will ensure smooth learning. So that's it for machine learning. Thank you for watching. I will see you in the next class. 14. Supervised Learning in Action: Let's not talk about supervised learning. So what exactly is this well? This is a type of machine learning where the model is trained by making use of labeled data. So what exactly is labeled data? This simply means that each training example or data with which the model is trained on, includes both the input, which would be the features, and then the correct output, which would be the actual label. So the whole idea here is for the model to be able to make predictions and find relationships between the inputs and the outputs. So how exactly does this occur? First stage is that the data will be collected and will be labeled. Going back to my favorite example, the spam filter. Thousands of emails could be collected in the first stage and then they'll be labeled. So we will have emails that are spam and then emails that in spam emails that are actually legitimate, right? And then the model will be trained. I'll be trained to identify which emails are spam and which wins. And then the third stage, the model will now be tested with new types of emails it has never seen before to be evaluated. And of course, if it does very well, it will then be deployed into the real world. That's how the whole process actually works. Now, when it comes to supervised learning, there are two types we have classification and we have regression. What exactly are these two? With classification, this is where the model will assign an input to a specific kind of category. It works best when the challenges or the tasks or the questions have discrete values. So say, for example, in your email spam filter, the email will either be spam or it's going to be legitimate, right? There's nothing in between. Let's talk about your lab results, right? The results can either be positive or negative. Sentiment analysis, right, maybe the sentiment was positive or negative or neutral. So classification is best when the values of the outputs are actually discrete. They have very specific kinds of values. Now, with regression, this is where the model has to make predictions and not the actual correct answer to a task. So say, for example, the model has to try and predict the sale cost of a actual house it's going to look at features like the size of the house, the location. But ultimately, it's still going to make a guess, a prediction. It doesn't know for sure if the house will be sold for the amount that it predicts, right? You talk about forecasting the stock market changes, the weather changes as well, the daily temperatures, things like that. So issues or problems where the output is continuous, regression will be used. So to kind of round everything up right now, I want to go back to my favorite example, the spam filter. So first stage, we have the training data where thousands of emails are collected, they're labeled spam, not spam. Now comes feature extraction. Remember, that feature extraction is a very, very important technique to improve the accuracy of the model. So over here, the model needs to be taught that, okay, certain kinds of features in your spam, you should look at them. So things like for example, the keywords being used, the title of the email reputation of the user, sending an email. These are the very important features that the model needs to be trained on to identify whether the email is spam or legitimate. And then, of course, like I said, the model will be trained. And then the model will be tested. New email it's never seen before will be presented to it. And then, of course, the test now will be whether or not the model can identify if that email is spam or not. And, of course, continuous learning, the model will improve over time as it looks at more and more emails. So what are the challenges and limitations of supervised learning? Well, first of all, data labeling is quite expensive. It requires large amounts of high quality data. And at this issue of overfitting, where the model, instead of trying to generalize and find patterns, it ends up actually memorizing the training data. And then, of course, the bias in data, you'll find this every time. If the data being provided to the model is biased, the model won't perform all that well. And, of course, supervised learning doesn't work well with unstructured data. So maybe you provide the model is provided with images that don't have labels or raw data. Supervised learning isn't going to work with unstructured data. So key takeaways to round up the lessons supervised learning learns from the label data to make predictions. It is used in classification, where data is categorized, and also in regression where it predicts our continuous values, real world applications will include your spam protection, medical diagnosis, stock prediction, and and of course, challenges will include data bias, data requirements, and of course, all fitting. Thank you for watching the video. I will see you in the next class where we're going to take a look at unsupervised learning LCU then. 15. Unsupervised Learning & Pattern Recognition: Last class, we talked about supervised learning. So now it is time to talk about unsupervised learning. Now, if supervised learning deals with labeled data, then obviously unsupervised learning will deal with on labeled data. So the whole idea here is for the model on the machine to try and find patterns with data that isn't labeled. Now, I've given an example in here where you have a company that has thousands and thousands of customer records, okay? The thing about the model here is that it's not going to know who customer A is or customer B or customer C or customer D. One thing it could do, though, is that it could look at the purchasing history of the customers and then try to group them into different categories based on again, their purchasing history. How much they spend. So, for example, the model could look at the customer records and decide that, you know what? Customers who have spent more than $1,000 at once, let me group them into the high spenders category. And then there might be customers who only make purchases whenever there are discounts, right? So it may want to group those kinds of customers under the shrewd category or, you know, something like that. Or it could even try to group customers based on what they actually buy. So maybe you have customers who buy accessories, like, you know, wristwatches, bracelets, things like that. So it may try to classify the customers under the accessories category, you know, stuff like that. That's exactly how unsupervised learning works. So there are four main stages involved in here. First of all, the model will receive the raw unlabeled, unstructured data, then the model on its own needs to find the structures, patterns, or relationships within that data. And then once it has done so, it's going to organize the data into meniful clusters or categories or components. And then finally, the output, whenever the model makes a prediction or puts output based on how accurate the output was, the model will then learn and improve over time, as well. So I've given you the key idea here is that the algorithm or model, it's never given the correct answer to a task or a challenge or a quiz. It learns patterns on its own. You can try to think of it as the model teaching itself, training itself to find the right kinds of answers to tasks and challenges. So there are certain kinds of key techniques involved in unsupervised learning. The first one in here would be clustering. This is probably the most popular technique. Here, data will be grouped into similar categories. So by the inforon the algorithm will divide the dataset into groups or clusters where the items in the same cluster are kind of similar to one another than those in other clusters. So the best use case in here, I've talked about customer segmentation where the customers, they could be grouped based on their purchasing history, how much they've spent, or what they like to buy or when they like to spend things like that, right? We do have some examples of the clustering algorithms. You have your mins clustering, hierarchical clustering where the clusters are built on a tree, and then the DB scan that will try to identify dense regions within the data. Now, the next technique in here is something called the dimensionality reduction. Don't blame me. I'm not the one who came up with this very, very interesting term. Dimensionality reduction. It sounds like something you would find in a space engineering textbook, right? It's kind of inside. I don't know who came up with this term. But basically, it simply means we're simplifying complex data. That's what dimensionality reduction means. So, by definition, this will reduce large amounts of data while preserving the key pattern. So your use case in here would be things like your data compression where data can be compressed into a smaller size, but the key features of that data will still be retained. We do have some algorithms for the dimensionality reduction, your principal component analysis, your PCA, and then your TSN, which is the T distributed stochastic neighbor embedding. There is no need for us to go deeper into these algorithms, okay? But the analogy in here is basically you try to imagine compressing a book that's, let's say, 500 pages or 1,000 pages all into one single page. Think of it as trying to summarize the key ideas of that book. So even though it's been reduced from 1,000 pages to one page. That one page will include all the key ideas and key information from that book. And then the final technique in here will be what we call the anomaly detection, where the model learns to find unusual patterns. So by definition, you will identify data points that deviate significantly from the gnome. Use case in here will be in your fraud detection or maybe even in your firewalls, right? A firewall can detect traffic that's malicious because it's unusual. Maybe the traffic is coming from an unusual IP address or from an unusual location. That's one of the techniques. That's one of the ways how the firewall is able to determine what is real traffic and what is malicious traffic. So examples of the algorithms used in here, we have the isolation for here, the algorithm will focus on identifying the outliers in the data, and then we have the one class SVM that's used to detect very rare and unusual instances. So the key analogy in here, think of your security scanners at an airport. They detect suspicious items simply based on patterns. Okay, so what are the real world applications of unsupervised learning? You have them in your customer segmentation, and normally detection which we could use for firewalls, fraud detection, and so on, and then medical diagnosis as well, where the AI model or machine finds hidden patterns in your genetic data or diseases. And then recommend systems like in your Netflix, your Spotify, YouTube, all the work based on the search history of the user. And then, of course, in your search engines, of course, like your Google, this will categorize pages based on topic similarities. So what are the challenges and limitations of unsupervised learning, so no clear labels. So there is no way to check if the model's output is actually correct. The model has to figure that out on its own. And then difficult to interpret. So some clusters or patterns might not be meaningful, okay? In an attempt of the model to try to group certain kinds of data into a cluster or a category, it may not do a very good job at that because how or what it used to group, that data may not be actually clear enough. And then choosing the right number of clusters. This is another big challenge in here. So algorithms like your mines require certain parameters. Basically, you have to indicate how many groups the model has to create or how many clusters it has to create. Otherwise, it may end up either creating too little of clusters or groups or maybe creating too many as well. So that could be a challenge. And then the computionally expensive, it's quite expensive. Training process large datasets, and this will require a lot of computing resources. So some key takeaways for unsupervised learning, unsupervised learning finds hidden structures in data without labels. Clustering is used to group similar data points, for example, in your customer segmentation, and then the dimensionality reduction will simplify data for better visualization and efficiency, and then anomaly detection will try to identify fraud, security threats, and unusual behavior. And finally, the challenges involved with unsupervised learning will include interpretability, computational cost, and of course, the lack of evaluation metrics. Thank you for watching the video. I will see you in the next class. 16. Reinforcement Learning and Decision Making: Well, come back. So now let's take a look at the third type of machine learning. And here, we're talking about reinforcement learning. Now, unlike in supervised learning where data is labeled or in unsupervised learning, where the model has to find patterns within the data, in reinforcement learning, the learning process is through trial and error. So basically, you have your agent or your AI model that will take an action in an environment, and then depending on the type of action that it takes, it can either receive a reward or a penalty. So over time, the model learns the right kinds of actions to take to receive more rewards. As an example, a self driving car will learn how to navigate roads by receiving rewards for self driving or penalties for collisions or reckless driving. How is the process? Well, first of all, we have the agent, which is the AI model that will operate will take actions. We have the environment within which the model is operating in, and then the state, the current situation of the agent in that environment, and then the action the agent will take. And then, of course, the reward system, positive feedback for good actions, negative feedback for bad actions. I do have the diagram in here giving you more information. So we have the agent that will interact with the environment. And then we'll need to take a look at the state of the agent within that environment, and then the action the model will perform an action based on its policy, and then they reward positive feedback or negative feedback. And then the agent or the model will update its policy so that in the future, it can get more positive rewards. And then, of course, this entire process is repeated. So imagine you're trying to teach a robot how to walk, right? If the robot makes a mistake, it could receive a penalty. Maybe it fell, for example, it receives a penalty. But then if it's moving its arms and legs correctly, it will receive rewards, and then over time, the robot eventually will learn how to work properly. So we do have four key algorithms for reinforcement learning. Let's take a look at them one by one. The first one here is the Que learning or the value based learning. So over here, in this particular kind of algorithm, the agent or the model will learn the best action to take by simply building something known as a que table. It's basically a table of rewards for different types of actions. So, say for example, all right an agent could take four types of actions, action A, B, C, or D. If it takes action A, it receives a big penalty. Action B, it receives a small penalty. Action C, it receives a small reward. Action D, it receives a big reward. So over time, it's going to build this table. It knows that, Oh, the more action Ds I take, the bigger the rewards will get. Over time, it learns how to take more and more action Ds. So it's used where the environment is fully observable, like when the AI is teaching itself how to play chess as an example. So I do have the diagram in here, giving you more detail. I'm going to provide you the slide. So, of course, you can look at this in your leisure time. Next algorithm in here is the deep Q Networks algorithm. This doesn't use que tables. It uses neural networks. So this can handle very complex environments like in your Atari games, autonomous robots. The example I've giving you here is where an AI learns how to play video games by maximizing its score over thousands of trials. The third algorithm in here is the policy based methods. So instead of learning values for actions, the model or the agent will actually learn a policy. A policy in this case, right now would be a strategy for choosing different types of actions. So it works very well in environments where a very small change could affect the entire outcome, like in robotics, for example. So the example I've given you here is a robot arm, learning how to grasp objects or pick up objects by simply refining its movements. So I have the diagram in here as well. Again, you can study this in your free time. And then the final algorithm in here is the actor critic method. This will combine both the value based and policy based methods for more efficiency. It's used for very complex real world problems like in your self driving, finance, and so on. An example of giving in here is an AI that learns how and when to buy and sell stocks in order to maximize its profits. We do have several reward applications, like I said earlier, self driving cars in your games like your AlphaGo, Open AI five, at Hari in your robotics, as well, finance and trading, and of course, in chatbards and customer support. But we do have challenges and limitations of reinforcement learning. The first one in here is that reinforcement learning requires millions and millions of simulations in order for the model to get better over time. It's not working necessarily with data here. It's working more with trial and error, so it has to undergo plenty thousands, maybe millions of trials and error. You know, to get better at what it does. And, of course, it's a very slow learning process. That's one of the biggest challenges of reinforcement learning, excuse me. And then we also have the exploration versus the exploitation dilemma. What exactly is this? Well, over here, agent or the model must try new actions, which would be the exploration versus using what it has already learned, which would be exploitation. It has to find the right balance between both, which can be very, very tricky. And then finally, the unpredictability. So the AI might find loopholes in the reward system leading to unintended results. This could be in situations where the reward system hasn't been properly fleshed out and certain actions that should have resulted in maybe huge penalties. Instead, the model gets huge rewards instead. So this could confuse the model, and eventually over time, it's going to end up making the wrong decisions. So some key takeaways before we round up reinforcement learning, here, the model learns by interacting within an environment and receiving rewards or penalties. Trial and error is, in fact, the foundation of reinforcement learning. Key reinforcement learning techniques will include your Q learning, your deep Q networks, policy based methods, and, of course, your actor critic models. And, of course, reinforcement learning is used in robotics. It's used in finance. It's used in self driving cars, games, and so much more. And, of course, the challenges include slow learning, high computation costs, and unpredictable behavior. Thank you for watching. I will see you in the next class. 17. Decision Trees, Regression, and Clustering: Now talk about decision trees, regression and clustering. Now, we've already talked briefly about regression and clustering earlier, but here we're going to delve a little bit deeper. But let's start off first by talking about decision trees. What exactly are they? They're basically a supervised learning algorithm used for both classification and regression. Remember that classification and regression are two techniques under supervised learning. So it tries to mimic the human decision making process by breaking down data into tree like structure of rules. Now, I've given you in the slide over here, the diagram of how it actually works. Everything starts from the base, which would be the root node, the data set. So eventually, this data will split into branches based on certain kinds of conditions or features, and eventually those branches will end in leaf nodes, which would be the final decision or prediction that the model actually takes. So the advantages of decision trees, as you can see, is very, very easy to interpret because we are talking about structures that are well defined, simple flow charts, right? And then it works very well with categorical and numerical data. So it's very, very versatile. And then there is no need for data scaling on what you'll find in regression or clustering. But we do have certain disadvantages of decision trees. One is that it is prone to overfitting, which is where the model, instead of trying to generalize and find patterns, it ends up memorizing the training data and then it's also sensitive to very, very small changes. Now, what are the real world applications of decision trees? You'll find them in your fraud detection, your medical diagnosis, or even in loan approvals in banks as well. Now, let's move on to regression. Now, we've talked about regression already under the supervised learning. Here, the model tries to predict continuous outcomes, right? And we talked about it being a supervised learning technique that's used to predict continuous values based on past data. One thing I didn't mention, though, was that we do have several types of regression. There's about eight or nine of them. But over here, I want to talk about the three most important ones in my humble opinion. We have linear regression, polynormal regression, and then logistic regression. Let me give you some analogies in here, okay? For linear regression, imagine the AI model trying to predict how much a house would cost. Now, in general, it knows that the bigger the house, then the more expensive it's going to be. That's in general. So in this kind of scenarios, the model could use linear regression to make the prediction. However, what about polynomial regression? Imagine the air model trying to predict the speed of a car. But here's the thing, though. The car isn't going at a at the same speed at all times. It could accelerate, I could slow down. It could accelerate again. I could slow down. So it's impossible for the model to use linear regression in this kind of scenario. It has to use a polynimal regression where it will try to plot a graph of the movement of the car and then try to determine how fast it's actually going. And then the last one in here, the logistic regression. Imagine you have an AI model that's been trained to determine how qualified a candidate would be for a job. Now, here's the thing. Eventually, the model will say, Okay, the candidate is qualified or isn't qualified, right? However, the decision making process isn't as straightforward because it's going to depend largely on the qualifications of that candidate. It could even depend on the qualifications of the other candidates who are also fighting for the exact same job. So in this kind of scenario, the model could use logistic regression. So the advantages of regression are that it's very, very simple. And easy to understand. It works very well with numerical data. And it's also the foundation for advanced machine learning models, which you find in finance, in healthcare, in business, in medicine, and so much more. Now, what are the disadvantages of regression? Well, it tries to assume a linear relationship at all times, which isn't, of course, realistic. And then it's also very sensitive to outliers. A single extreme value can distort all predictions. Imagine going back to the old analogy of the model trying to predict the cost of a house. What if in one situation, a very small house ended up costing more than a much bigger house. That's an anomaly. That doesn't really really happen. But because it did, in fact, happen, that could confuse the model, and then the model moving forward, may not be able to make accurate predictions anymore based on that single anomaly. Now, what are the real world applications of regression? You have your stock market prediction, weather forecasting as well, and also sales for casting. These are just a few examples of the reword applications of regression. And finally, clustering. We talked about clustering under unsupervised learning where the model will try to group similar data points together without predefined labels, we have the different algorithms in here which we already talked about earlier. The analogy in here, imagine you own a clothing store, and you want to group customers. So you can group customers into high spending, better conscious or maybe customers who only buy during sale. So clustering can help group customers into similar types of categories. But what are the advantages of cluster? We didn't talk about this. First of all, it can uncover hidden patterns. So there is no need for any kind of label data when clustering is involved because clustering by itself will label data into different kinds of groups, right? And then it's also very, very flexible because it's not working with predefined data or labeled data or anything like that. It basically can work well with customer segmentation, medical data, and also image recognition. And then it's also very, very useful for detecting anomalies. So your firewalls, intrusion prevention systems, and cybersecurity and so on, those typically use clustering as a way to function. But we do have disadvantages, of course. First of all, choosing the number of clusters can be very, very difficult. Sometimes the model will not know the right amount of clusters to create for a certain amount of data, and then clusters can also overlap. This happens quite frequently where some data points might belong to multiple clusters. Going back to the whole clothing store, it's possible that customers who are high spenders, they may also belong to another cluster that's specifically for customers who like to buy accessories. It's just that maybe they buy the very expensive accessories, maybe they buy the very expensive wristwatches. So these kinds of customers will now fall into two different types of clusters, and that could lead to some complications and confusion over time. What are the real world applications of clustering? Customer segmentation anomaly detection, and, of course, in medical research as well. So L's, thank you very much for watching the video, I will see you in the next class. 18. Challenges and Ethical Considerations in Machine Learning: Come back. So before we run up this module on machine learning, we need to talk about the challenges and ethical considerations in machine learning. Now, regarding the challenges, there's three types of them. We have the data related challenges, the model related challenges, and of course, the computational resource challenges. Let's take a look at them one by one. Now with data related challenges, we're talking about issues like the data bias. So once again, if the model is trained on data that's been biased, guess what, the model will make biased predictions in the future. And then data privacy because machines and models need to be trained with large datasets. Sometimes these datasets could be data belonging to customers, users. This could lead to privacy concerns, and finally, the data quality issues. If data is missing or it's noisy or it's unstructured, this could affect the model's accuracy. So as an example, if a medical AI model is trained on incomplete patient records, it could end up making unreliable diagnosis. What about the model related challenges? We've talked about overfitting where the model, instead of trying to find patterns and generalize, it ends up memorizing the training data and then interpretability and explainability. Sometimes the machine learning models can be very, very difficult to understand. And when a model makes a certain kind of decision, it might not be able to explain why it ended up taking that decision. They kind of black boxes that we don't fully understand. And then also the adversarial attacks where hackers or several criminals, they can trick machine learning models by simply modifying the input data slightly. As an example, adding a subtle or very small noise to an image causes an AI to misclassify a stop sign as a speed limit sign, which can be dangerous for self driving cars. Now, what about the computational resource challenges? When it comes to AI training, machine learning, they require high computational costs. You're talking about large data centers, large databases, very powerful computer processors and so on. So it does require plenty of hardware and energy consumption. As an example, training the Chart GBT version four required thousands of GPUs and Tra watts of power. Raising sustainability concerns and also the issue of scalability. Just because a model works very well in a small environment doesn't necessarily mean that it will work better in a big environment. As an example, an AI chatbot, maybe it did very well. It was trained to interact with ten customers at once. But what if that chatbot needs to interact with 100 customers? It may end up performing very, very badly. But with the challenges out of the way, let's talk about the ethical considerations, bias and fairness, privacy and surveillance concerns, job displacement, and economic impact, and, of course, AI safety and autonomous decision making. Let's take a look at them one by one. So when it comes to bias and fairness, and algorithmic discrimination, the machine learning models, they can reinforce social inequalities if they've not been properly designed or if they've been trained using biased data. So organizations and developers of AI models, they must ensure that their systems, their agents, their models are fair and unbiased. And the solution in here would be to use very diverse and a very broad range of different types of datasets, and then bias detection tools should also be used. Now, when it comes to privacy and surveillance concerns, one of the best examples of this would be in China that uses the social credit system, and they also use AI and surveillance cameras to monitor their citizens. So this could lead to big issues of privacy and so on. So the main solution in here is that the companies or countries in the case of China should adopt transparent policies on data collection and user consent. I don't think that's going to happen in China, but hey, that's a different topic for another day. What about the job displacement and economic impact? We are already seeing some people losing their jobs because of AI. So automation and unemployment, AI is replacing human jobs in industries like in manufacturing, finance, in customer service as well. So the solution here would be to simply try to reskill the workforce, and governments and companies, they should invest in AI education and upskilling programs. So employees they should be trained on how to work with AI so that they can gain new skills that will allow them to function and work in an AI powered environment. And then, of course, the AI safety and autonomous decision making. We can talk about AI in warfare, like in lethal autonomous weapons. Countries right now, like, of course, the United States, China, maybe even Russia, they are developing AI powered autonomous weapons, which would, of course, raise ethical concerns. For example, the use of military drones. You see them operating in several places. And then also AI in life critical systems, like in healthcare as an example, it's used AI is used in healthcare, it's used in finance and transportation, where failures can have life threatening consequences. Imagine that self driving car that ends up making the wrong decision, and then it crashes into another car that had people in it. Those people could die as a result. So the solution in here is that AI in safety critical applications, they must undergo as much training, as much testing, and as much oversight as possible. Regulations and solutions for ethical AI, AI ethics principles. So basically governments and organizations and companies, they must follow principles like, of course, fairness, accountability, transparency, and, of course, privacy protection. And then also laws and regulations should be introduced. In fact, some have already been introduced by governments like the EU AI Act and, of course, the GDPR. So some key takeaways before we round up Machine learning faces challenges in data quality, model explainability, and adversarial attacks. Ethical concerns include bias, fairness, privacy violations, and then of course, job displacement. So AI must be must be designed responsibly to minimize harm and maximize fairness. And, of course, governments, companies, organizations, businesses are developing AI regulations and ethical frameworks to ensure safe AI deployment in the real world. Thank you for watching. I will see you in the next class. 19. Section Preview Deep Learning and Neural Networks: Come to the next module, and here we're talking about deep learning and neural networks. And, of course, the rule in this course is that at the start of each module, I'm gonna play you a movie clip, so sit back, relax, enjoy the clip, and I'll see you at the end of it. What does this action signify? It's a sign of trust. It's a human thing you wouldn't understand. My father tried to teach me human emotions. They are. Difficult. Want to explain why you were hiding at the crime scene? I was frightened. Robots don't feel fear. They don't feel anything. They don't get hungry. They don't sleep. I do. I have even had dreams. Human beings have dreams. Even dogs have dreams, but not you. You are just a machine. An imitation of life. Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece? Can you? Well, come back, and, of course, that clip was taken from the movie I Robot released in the year 2004, starring Will Smith, and it's actually one of my all time favorite Will Smith movies. Now, the reason why I wanted to use this particular clip is because it demonstrates the topic of this module, which is deep learning. In the scene, we had an interrogation between Will Smith's character, Detective Spooner and the AI robot called Sonny. Notice, at the beginning of the clip, Sonny observes Detective Spooner making this official gesture kind of like a wink, right? And it memorizes the wink, and then eventually, during the interrogation, Sonny actually asks Detective Spooner, What does this mean? And, of course, detective Spooner who doesn't really like AI or robots, begins to say, Oh, it's a human thing, you wouldn't understand. But Sonny actually challenges Detective Spoona. Sonny says that I am capable of emotions. I've even had dreams, which is kind of fascinating, right? See, the reason why I'm using this clip is because it shows Sonny, a highly intelligent, advanced AI model that can actually think it's far different from the typical machine learning AI systems that can only follow either pre written rules or are able to identify patterns and data and then make decisions. This particular kind of AI models, they can think. They're able to think outside the box through the concept of deep learning. See, deep learning simply aims to try and mimic the actual human brain by making use of artificial neurons. So this kind of subset of machine learning it allows AI models to think. No, like I said earlier in the course, there are skeptics who believe that artificial intelligence will never get to this particular level of sophistication. Well, there are those who think that we will eventually get there. Regardless, I thought this clip was going to be a good clip to introduce this module where we're going to talk about deep learning, which is, of course, a more advanced subset of machine learning. Let's now move on to the next lesson. 20. Introduction to Deep Learning: Let's now take a look at deep learning. What exactly is this? Well, deep learning is a subset of machine learning that uses artificial neural networks with multiple layers to process and learn from very large amounts of data. Now, the whole concept of deep learning was inspired by the ability of the human brain to recognize patterns and then make decisions. Now, we do have some key characteristics of deep learning. Perhaps the most important one is the fact that like traditional machine learning models, deep learning models are able to automatically extract features all by themselves. With a traditional machine learning, the developers of that AI model have to provide the features to the model by themselves. But deep learning models, uh, they can automatically by themselves, extract such important features. Now, they're also able to handle very large amounts of data. But because deep learning models can be very, very complex, well, they do require high computational power. So I do have the table in here that shows you the core differences between traditional machine learning and then deep learning. So you can say in feature engineering, I just talked about that with machine learning, it's typically manual. The features have to be provided to the model by the developers, but deep learning models, they do that by themselves. And then data requirement, machine learning handles small to medium sized data sets very, very well. But then for deep learning, they can handle massive amounts of data. And then, of course, performance. Your traditional machine learning, they get kind of limited whenever they're working with complex data. But deep learning, the more complex data is the better for them. Now, interpretability, machine learning is much easier to understand than deep learning models that become kind of like black boxes. We don't fully understand how such models actually operate in computonal power. With machine learning, it works well on normal computers. But for deep learning, you need some extremely powerful computers to run them. So we do have several key real world applications of deep learning. For example, under computer vision, you have your image recognition, medical imaging as well, and then your chat boards like CII, Alexa, these kinds of chat booards are developed by making use of deep learning. And then in speech and audio processing, like your music generation, emotion recognition from voice, all these use deep and then in finance and business like being able to detect fraudulent transactions, stock market prediction, they all use de learning. And finally, in healthcare, we have our AI as AI assisted. This is diagnosis, predicting patient outcomes, and so much more. So deep learning has its place in many fields in our everyday life. So why is deep learning so powerful? What is it about deplaning that makes it so, so powerful? Well, like I said earlier, it can handle very, very complex types of data. It doesn't matter how complex or how complicated data is. If it's deep learning, it's going to thrive. And then the self learning capability, again, deep learning models, they can extract key features all by themselves. They don't need supervised learning. They don't need a developer to teach them how to think or how to recognize patterns. They can do so by themselves. And because they're constantly improving over time, as well. The kinds of deep learning models that we have today will be nothing compared to the deep learning models that we get in a few years from now. So a few key takeaways deep learning, again, it is a subset of machine learning that uses new networks. It outperforms traditional machine learning in handling large and complex datasets. But then deep learning powers many AI driven applications from self driving cars to virtual assistants. Thank you for watching. I will see you in the next class. 21. Understanding Neural Networks: Previous lesson, we talked about deep learning, and I mentioned that deep learning makes use of neural networks. So it's only natural that we now discuss in more detail what exactly neural networks are. Now, just like the name would suggest, it is basically a network of interconnected nodes that we refer to as the neurons. Now, the neural network itself, it is a computational model inspired by the biological neurons in the human brain. So with that being said, what are the key characteristics of a neural network? They're able to learn patterns from data. They are made up of multiple layers of artificial neurons, and then also they make use of mathematical functions to process and also transform data. Now, what would be the structure of a neural network? There's three main parts. You have the input layer, the hidden layers, and then finally, the output layer. Let's take a look at them one by one. Start off with the input layer. Now, as the name suggests, this is the very first layer. This is the layer that will receive the raw data, which could be images, audio, video. So each node that you have in your input layer will represent a feature of that particular data. So next comes the hidden layers. These are the layers that will perform the actual calculations and will extract the patterns from the data. So this is basically where the main activity occurs in the hidden layers. So each hidden layer contains multiple neurons, which will apply the matcal operations to the tech relationships in the data. So more hidden layers simply means that the network is deeper and therefore more complex and therefore more powerful. And finally, you have the output layer, and this is where the network makes a final prediction or produces the final output. Going back to the same spam filter example, the output layer is what will determine whether or not the email is spam or not spam. So now that we know that neurons make up the neural network, what exactly are neurons themselves and how do they work? So each neuron in a neural network is a computational unit that will process inputs and passes outputs onto the next layer. How do neumons actually operate? There's four main layers in here there's four main steps, as you can see. Now, we've talked about the inputs and then the output, those are self explanatory. But in between is where we have steps two and three. Step two involves the application of weights and bias. What exactly is this? The reason why weights are applied to each input is to basically determine how important the input is. Think about it, okay? A neuron cannot process all inputs at the same time, so it needs to determine which inputs are the most important. An input that has more weight will naturally be more important and will add more value. Now the reason why biases are applied is because imagine if all inputs will have the same value or let's say all inputs had the value of zero, in that kind of scenario, the neuron might not be able to operate because, hey, all inputs are of equal value. So a bias is now applied to ensure that the neuron isn't entirely dependent on the values of those inputs. And then finally, you have your activation function. This is the actual function that will determine whether or not the neuron should activate and then pass the data on was. So speaking of the activation function, there's four of them. You have your ELU the rectified linear unit. This is used in the hidden layers and helps deep networks train much faster. You have your sigmoid. This is typically used for calculating probability, so it works well with values 0-1, and then you have the hyperbolic tangent. This outputs values between negative one and positive one. The last one here is the softer max used for multiclass classification. So just a quick summary, neural networks consist of input, output, and hidden layers. Nus process information by applying weights, biases, and, of course, the activation functions we've just talked about. And then activation functions will determine if a new one should fire or not, basically, should a new one activate or not. So more hidden layers leads to deeper networks, capable of handling even more complex problems. So if there's one thing you can take away from this lesson, just remember that the more hidden layers a neural network has, the more powerful it is likely to be. Thank you for watching. I will see you in the next class. 22. Types of Neural Networks: Now take a look at the different types of neural networks that we do have, and there are six of them. The very first one here is going to be the feed forward neural networks. These are easily the simplest type of neural networks because the inputs go through a straight line. There are no loops or cycles. The input just goes all the way in a straight line and eventually results in an output. So we do have some key characteristics. There are no feedback connections, so basically no loops at it's used for tasks that you'll find in classification and regression. And then it consists of fully connected layers. So each neuron is connected to the very next layer. But now we have the applications where our fit for neural networks actually used. We use them in your spam email filtering and also in your stock price prediction as well. But now let's move on to the next type of neural network, and that is the CNN, not the cable news network. I'm talking about the convolutional neural network. So the specialize in processing grid like data such as images. So the key characteristics for the CNN are that they use convolutional layers to detect spatial patterns. They can extract very, very important features like edges, textures, objects, and so on. So they also are able to reduce complexity by simply using pulling layers. And because of this, you will find them mostly around images. So image classification, they will use CNN object detection as well. And even in the medical field on the medical imaging, those use the convolutional neural networks. Next, we have the RN and the recurrent neural networks. These are designed primarily for sequential data, meaning that you will have memory and can process dependent patterns. And because of this, the key characteristics are that it can contain feedback loops. They can also remember previous inputs. So they do have kind of like a short term memory, and then they're good for tasks that require context understanding. So with that in mind, can you guess the kinds of applications where we would use the RNN? Yep, we can use them in speech recognition. So your chatbards, Alexa, sii, they would use this particular kind of network, and then in language translation as well. And then also in other areas like in stock price forecasting. Moving on to the fourth type, these are the transformer networks. They're kind of similar to recurrent except that they're more advanced because they can handle entire sequences of data all at once. So they're a bit more powerful than they call neural networks. They're more efficient and they are also used to power state of the eye models like your chat GPT. So the applications, we use them in chatbards in virtual assistant, in machine translation, and also in content summarization. This is where you would use transformer networks. The fifth would be the generative adversarial networks they typically consist of two competing new networks. You'll have your generator and your discriminator. So the key characteristics are that the generator can actually create the fake data. The discriminator will then try to distinguish what is real and what's actually fake. So they are used for generating realistic synthetic data. And the applications, for example, you have them in your deep fake videos. You have them in your AI generated network like your Dali, M basically any kind of AI model. They can generate images. You would have this particular kind of network, and also in image to image translation. And then the final would be the auto encoders. Now, an auto encoder is a type of a neural network that's basically used in the unsupervised learning phase to compress and then also reconstruct data. So the key characteristics, they learn efficient representations of data they also consist of the encoder that will compress an input, and then the decoder, they'll basically reconstruct that particular input. And then they're used for dimensionality reduction. Remember dimensionality reduction. We talked about it. It's the fancy term for the process where complex data is simply simplified. So it's also used for anomaly detection as well. And for the real world applications, they're used in image noise reduction in your anomaly detection and also in your feature extraction. So summary, I've provided a table in here all the different types of neural networks, as well as what they are best used for. And also the example are use cases. So you can check out the slide which I'll provide for you. You can study this in a bit more detail. But that's been a very quick look at the different types of neo Networks. Thank you for watching. I'll see you in the next class. 23. Challenges in Deep Learning: We've talked about the challenges involved in machine learning. But what about deep learning? There are also challenges. So I'm going to start off with the data related challenges, and they're not that different from those of the machine learning. We're talking about challenges like data availability, data bias and fairness, and of course, data privacy and security. So when it comes to data availability and collection, Remember that deep learning networks, they require very large datasets in order to operate. So getting high quality data and in such vast volumes and amounts can be a challenge. But what are the possible solutions? Well, data augmentation, where you can simply rotate images for training. And then the transfer learning, where we can use pretrained models to reduce data requirements, and then also the use of synthetic data. Remember in the previous lesson, we talked about different types of neural network and I mentioned the generative adversarial networks, the gangs, we can use them to generate artificial training data for the networks. And when it comes to data bias and fairness, again, I've talked about this several times already. If the data used to train the model is biased, then guess what, the model will make biased predictions. So the solution will be to use very diverse and broad range of datasets, implement bias detection and fairness Aware algorithms, and also regularly audit the AI models for fairness. What about data privacy and security? Again, because deep learning requires large datasets, sometimes those datasets might consist of user data, employee data, customer data, so this could bring about privacy concerns. So the solution would be to use what we call federated learning, where we can train the AI models without sharing any kind of raw data, and we can also simply secure the datasets by making use of encryption. We also have computational challenges. Remember that deep learning requires plenty of computational power because the models are very complicated. So we do have the high computational costs and then slow training time. Now, regarding the high computional costs, the solution here would be to simply use model pruning. And quantization to reduce the size of the model. And then we can also leverage resources like cloud computing and then simply develop efficient architectures. For the slow training time, remember that because the deep lane models have to find patterns themselves, they have to extract features by themselves. Because of this, it can take plenty of time, several weeks, maybe even months in order for it to actually function, especially if it's working with very large datasets. So the solution here would be to use distributed computing to train models in parallel. So we're training multiple models at the exact same time. And then we can also optimize the learning rates with techniques like the adaptive optimizers, the adam optimizer as an example. And then implement check pointing to save progress and avoid starting from scratch. So once the model has trained to a certain extent, we can save their progress, and then they can always continue from there if any future errors do occur. But we also have the model related challenges overfitting against underfitting, explainability and interpretability and adversarial attacks. Let's take a look at them one by one. Now with overfitting, we've talked about this where the model instead of trying to generalize and find patterns, it ends up memorizing the training data. And then also on the fitting where the model is so simple that it's unable to capture patterns. So the solution here will be to apply regularization techniques like your dropout, T regularization, and so on, and then use cross validation to test generalization. And then we can also increase the dataset size or add noise to improve the robustness. But what about the explainability and interpretability? Because again, deep learning is very, very complex. It box with complex data. Being able to explain the decisions made by deep learning models can be a bit of a hassle sometimes. They kind of act like black boxes. We don't fully understand how they actually make the decision. So the solution would be to simply use explainable techniques such as sharp values and lime and then build attention mechanisms to highlight which particular features while used for the model to make a certain kind of decision and then develop a rule based hybrid models for more transparent decision making. Now, regarding the adversarial attacks, the issue here is that very small changes to an input can actually trick the neural network into making all sorts of incorrect predictions and providing incorrect answers. So as an example, if you were to change just a few amount of pixels in an image, could end up confusing the model and it ends up misclassifying what that image is actually supposed to be. So the solution here will be to simply train the models by using adversarial training. For example, we can expose them to attack examples. So basically, we train them, we attack them deliberately so that they can learn from the attacks and protect themselves in future. And then use defensive distillation to make models even more robust and then implement secure artificial intelligence protocols in critical applications. And then, of course, the ethical and societal challenges like the AI bias and ethical concerns. Deep fake, the misinformation risks. So when it comes to the bias, the dip leaning models can reinforce societal biases if they're trained with biased data. So, of course, the solution will be to implement transparent AI training, transparent AI guidelines, and so on, and then enforce accountability and regulatory frameworks for all types of AI models, and then also promote responsible AI development with fairness and inclusion. And when it comes to the issue of deep fake and misinformation, well, deep learning can be used to generate deep fake images. We've already seen examples of these being used in separate attacks. The solution here would be to develop algorithms that can actually detect whether or not an image or video is a deep fake and then also establish AI generated content verification standards. This is very, very important. And then educate the public, of course, on being able to detect manipulated content. But honestly, with the rate at which the images and videos generated by AI, they are becoming more and more realistic, I think eventually at some point, we will not be able to rely on our human ability to recognize what is actually real and what is fake. So in summary, I have provided the challenge types in here, the problems, and also the solutions. So I'm going to provide you with this slide. You can take a look at this at your leisure time. Thank you for watching. I will see you in the next class. 24. The Future of Deep Learning: Come back. So before we round up this module on deep learning, I thought we'll take a look at the future of deep learning. What are the new kinds of features that we expect to be developed for deep learning models? First, will be more efficient and scalable deep learning models. We all know the in problem right now is that the deep models that we currently have, they are very large, they're very expensive, and they're also energy intensive. So the future trend here is that we now have smaller, more efficient models for deep learning AI. And then the use of quantum AI that can help speed up the learning process for the deep learning models. And of course, the impact here is that AI will become much faster. It's going to become much cheaper because it's using less computional power, and it's going to become more accessible to everyone. Another future trend is that the AI will be able to learn with less data. Remember that one of the key problems or challenges of deep learning is that it requires large data sets. So in the future trend, we'll have self supervised learning SSL where models will be able to learn from unlabeled data. And then what we call the few short learning where the models will be able to learn or adapt by making use of just a few examples. And then the transfer learning improvements where pre trained models will be fine tuned with minimal data. Now, as a result of this future trend, AI will be able to generalize much better and also work in low data environments. Another future is where AI will be able to understand context and deep reasoning. The current challenge is that the deep learning models, they are unable to develop true reasoning abilities and are just limited to being able to recognize patterns. So the future trend will have different types of AI, like the neuro symbolic AI that can combine deep learning with symbolic reasoning and then also causal AI. This would be the kind of AI models that can understand cause and effect relationships. And then finally, the multimodal AI. These will be systems that can process your text, video, audio images all at the exact same time. And as a result of this, guess what? AI is going to become much more intelligent. It's going to become explainable and also it'll be capable of complex reasoning. One more feature will be the democratization of AI. The issue right now is that AI development, the playing development, it's in the hands of the Tech giants. So in the future trend, we expect to have more open source AI models. In fact, Deep Sk, which is the model developed by the Chinese company, it is open source. So I believe that's just a start of us having more open source AI models. And then the decentralized AI, think of AI like on the block chain, and then edge AI, where AI will be able to run on our mobile devices. So as a result of this, AI will become more widely available, reducing dependence on the large company. So just a few key takeaways. The future of deep learning will focus on smaller faster and smarter AI models. AI will require less data, less computional power with improved reasoning capabilities, and in responsible AI development, it's necessary, of course, to avoid bias, and, of course, ensure transparency. We talked about this before. And then also governments and organizations must implement AI regulations and governance frameworks. Thank you so much for finishing this module. I will see you in the next module. 25. How Neural Networks Learn: Let's not take a look at how neural networks actually learn. Now, there's four stages involved in the learning process. We have word propagation, loss calculation, back propagation, and then optimization. So let's take a look at them one by one. Start off with the Ford propagation. Now, this is where data will initially flow through the network. The input is processed throughout the network, and finally, an output will be produced. And you can see right there on the slide is basically six stages. You have your input layer that will receive the raw data. Next, the weights and biases. Remember we talked about that will be applied. Activation function will also be applied, and then the hidden layers will process the output pass it on to the next layer. And then if there are more hidden layers, this might help to refine the data even further. And then finally, you have the output layer that will produce the output. So at this stage, the network has a value, but next comes the loss calculation. This is where the actual real value of the output will be compared with the output produced by the model. So the loss function calculates the difference. The bigger the loss function, then the more far off the network was from making or getting the right answer. So the whole idea here is for the model to learn, get better, and in time, over time, the loss calculation will be reduced. So we do have the common loss functions applied in here. You have your mean squared error MSE that's used for regression problems, and then you cross entropy loss that's used for the classification problems. And I gave an example in here where if the model predicted 0.74, let's say, cat or dog, but then the actual value was a one, then the loss function will calculate the error to improve the model. Now, the back propagation, this is where the actual learning process will occur because once the loss has been calculated, the network will adjust its weights and its biases by making use of back propagation. So how does back propagation actually work? First of all, the loss will be sent back through the network. The network will then determine how much each weight contributed to the arrow, and once it's able to decide that, it can then adjust the weights and biases to reduce the amount of arrows. So as an example, if an image is misclassified as a dog instead of, let's say, a cat, back propagation will correct the model by simply adjusting the weights. And because of back propagation, optimization now comes into play where the weights have been fine tuned. So to improve learning, neural networks use optimizers to tweak the weights efficiently, and we do have algorithms for these. You have your gradient descent, which we've talked about already, your stochastic gradient descent, the SGD, which will update weights with small data batches, and then the Adam optimizer, which is a more advanced method that will adapt the learning weights dynamically. So as an example, in self driving cars, the optimizer will help the AI refine its decision making to drive safely. So quick summary before we round up, neural networks learn by adjusting weights to maximize or, I'm sorry, minimize their errors. And then forward propagation will send the input through the network to produce an output. The loss calculation, the loss function of IDR will calculate the difference between the predicted value made by the network and the actual value, and then back propagation will adjust the weights to reduce the loss function. And then optimization algorithms will fine tune the entire learning process. Thank you for watching. I will see you in the next class. 26. Section Preview Natural Language Processing (NLP): Welcome to our next module, natural language processing. And as usual, I'm gonna play you a clip to introduce this module so sit back, coax and enjoy the clip, and I'll see you at the end of it. Hi. Hi. How you doing? I'm well. How's everything with you? Pretty good, actually. It's really nice to meet you. Yeah, it's nice to meet you, too. Oh, what do I call you? Do you have a name? Um, yes. Samantha. Really? Where'd you get that name from? I gave it to myself, actually. That's really weird. Is that weird? Do you think I'm weird? Kind of. Why? Well, you seem like a person, but you're just a voice in a computer. I can understand how the limited perspective of an unartificial mind would perceive it that way. You'll get used to it. Was that funny? Yeah. Oh, good. I'm funny. Well, come back. So that clip was taken from the movie her released in the year 2013, starring W in Phoenix as Theodor the only guy, and Samantha, an AI virtual assistant are voiced by Sclet Johansen. Now, in the scene, you have what appears to be a very natural like conversation between Theodore and Samantha. And I'm pretty sure that if you didn't know that this conversation was between an AI model and a human being, you would have thought that this was a conversation between two people because it felt so natural. Then the reason why Samantha was able to understand what Theodore was saying is because of natural language processing. It's what allows AI models like Samantha to interpret what human beings are saying and then respond back. It's all through natural language processing. But I want to ask you a question. I'm sure you do agree that the conversation felt very natural. Why did it feel so natural? Was it how Samantha was speaking? Was it the words that she used? Was it the fact that she was able to make some jokes? Like, why did it feel so natural? I personally think that it's because she was able to very quickly adapt to Theodore's personality. If you recall in the clip, she was able to even make certain kinds of jokes that made Theodor laugh. And Theodore said, Okay, you know, you're funny. So, Samantha was able to make certain kinds of jokes. She even giggled. But I also do believe that the voice of Samantha also made the conversation feel very natural. See, CletioHansen, who voices Samantha. She has a very feminine, soothing, calm voice. I guarantee you that instead of Samantha, we had the AI model called I don't know, boys, for example, okay? And instead of the cool, calm, feminine voice, we had a very masculine voice, like, you know, good morning, Theodore. How can I help you today? Have you eaten anything today? Like, if Samantha sounded like that, then it won't sound so natural anymore. Now you might start thinking, Okay, this sounds more like a conversation between a robot and an actual human being. Think of Arnold Schwarzenegger in the Tamino movies, right? He's very deep, strong or Austrian accent, you know? That sounds more like a robot, you know, with all due respect to Arnold Schwarzenegger. So I do believe that because Samantha was able to make jokes because she was able to quickly learn and adapt to Theodo's personality and the fact that she had a very calm voice, I think all of this contributed to the conversation feeling so natural. And let me say one more thing before I round up this particular introduction. I do believe that in the near future, we're going to have assistance like Samantha, except that they will be for companionship. Loneliness is a big problem in our society today, and I think it's only going to get worse. So I think eventually there will be a demand for either virtual powered assistant or maybe even robots. At some point in the future, they'll be powered by AI that will be there to serve as companions to combat loneliness, to help lonely people. I think that's something that's going to happen eventually in the future. Nevertheless, though I thought this was an interesting clip to introduce new module natural language processing, so thank you for watching. I will see you in the next class. 27. Introduction to NLP: Let's now take a look at natural language processing. I've mentioned this quite a few times already in the course, but now it is time for us to delve a little bit deeper into this fascinating topic. So what exactly is NLP? Well, it is a special field of AI that helps computers or AI models to understand, interpret, and also generate human language. You can think of it as the bridge between human communication and machine intelligence. It's what allows machines and AI models to understand us as humans. So what are the key points involved in NLP? Well, first of all, it is a combination of linguistics, computer science, and, of course, artificial intelligence. Now, it helps computers to read, understand, but also respond to human language. Now, because of this, it is used in text analysis, speech recognition, machine translation, and so much more. So it's a very, very, very useful kind of technology. So why is it important? Well, communication. Your chat box Your chat box like Alexa and Siri would not exist without NLP. It's also used by your favorite search engines like Yandex, Bink and Google to find relevant terms whenever you use them whenever you type in your search items. Also, automation, NLP automates repetitive tasks like your spam filters, your intrusion detection systems, and so much more. And then also accessibility. Your speech to text, this helps people with disabilities to communicate more effectively. So what is the history and evolution of NLP? I'm going to give you a few points in here back in 1950, the very famous scientist Alan Chewing, he proposed the Tuin test to assess machine intelligence. Then the first four, ten years later in 1960, the Eliza chatbd, which I believe is the first chat board ever created, it was able to mimic human conversation by using our predefined rules. And then between 1990s and the 2000, we had the statistical methods that were implemented to improve language modeling. And finally, between 2018 and present day, transfer based models like your GPT have revolutionized NLP. So everyday applications, as I've said earlier, in speech recognition, it's used in voice assistant like your Alexa, SEI, search engines, Google, Bing Yanex, they all use NLP. And then also for spam detection, email filtering, your chatbards, your virtual assistant, they all use NLP. And then also in machine translation as well, because it automatically translates text between languages, and then text, autocorrection, and prediction. Whenever you're typing in Microsoft Word or on your phone and you have the auto correct feature on, that is using NLP. So a few key takeaways before I round up the video, our NLP allows machines to understand process and, of course, generate human language. It has evolved over time from rule based systems to more powerful deep learning models. It's widely used in your search engines, your chatbards, voice assistants, so much more. And modern NLP models like your GPT, your BERT, they can understand and also generate human like text. Thank you for watching. I'll see you in the next class. 28. Key NLP Concepts and Techniques: Let's now take a look at the key NLP concepts and techniques because it's actually very, very fascinating. So let's first of all, understand at the base level how NLP actually works. There are four phases, as you can see from the slide, and the very first stage typically involves the provision of the input. So this could be a sentence. It could be a text. Next will come the pre processing stage. This is where the text or the input provided will be cleaned and formatted, making it ready for analysis. Then come stage three, which is going to be the actual processing and analysis, where the NLP will apply techniques like tokenization and so much more, which we'll talk about later. And then finally in stage four, you'll have the model interpretation. This is where the system will generate an output based on its understanding or translation of the input. So speaking of the techniques, what are they? There's quite a number of them, but trust me, these are all very, very interesting techniques. Let's take a look at them one by one. And the first one in here is the tokenization technique. Typically, the very first technique applied. So over here, the input provided in the very first stage will be broken down into smaller components. So as an example, if I made an input of I love artificial intelligence. Tokenization will break down that sentence into I love artificial intelligence. So it's broken down the sentence into four different parts. That's where tokenization comes in. Now, why is this useful? Why is it applied? Well, it helps the machines to understand the text structure, and it's also very essential for search engines and chat boards. In order for a search engine to function properly, it needs to be able to break down the input of search terms that you've provided it. Now, after tchonization, we have lemmatization and stemin. This is the process of reducing the words in the input to their base forms. Now, let's take a look at them one by one. What is Semin? Stemin very, very simply remove suffixes. So for example, if in your input, you had the verb are run it's going to reduce that to run. If you had flies, it's going to reduce that to fly. That's basically Semin. Now, lemmatization will convert the words to their base form. The core difference between Semin and lematization is that Semin doesn't care about the structure of the input or the sentence or the context. It doesn't care. All it concerns itself about is removing suffixes. Lematization, on the other hand, actually understands the structure and the context behind the use of that particular word. So as an example, if I input it the cats are running, ematization will say the cat be run. Why? Well, first of all, it's going to reduce cats to the base form, which is, of course, cat, cat being singular for, I'm sorry, cats being plural for cats, so it's going to reduce cats to cat. And then the R, the verb R, the base form is actually to B. That's why it says the cart B. And then, of course, running, the suffix is removed, it becomes run. So the cat running becomes the cart B run. Now, why is this useful? Well, it improves such accuracy because Google knows that, for example, it knows that when you type in running, that means you're talking about something related to the word run. It already knows that. And then it reduces word variations as well for better text analysis. Then we have the POS tagging the part of speech tagging. So this is where the model will be able to label words into different categories like your nouns, adjectives, adverbs, and so much more. So, for example, in this very popular example, the quick brown fox jumps over the lazy dog. Over here, the model knows that, okay, quick is going to be the adjective. It knows that Fox is going to be the noun. It knows that jumps is going to be the verb and so much more. So the reason why this is important is because it helps machines to understand the grammar and the meaning of the sentence. And, of course, it's used in your chat boards. It's used for grammar checks and so much more. Then we have the named entity recognition the NER. Here, the model is able to extract key entities like dates, places, names of people, locations, and so on. As an example I've given this one, Elon Musk, founded Tesla in 2003 in California. So with the use of NER, the model knows that, okay, Elon Musk is a person that's a person's name. California is a location. It knows that Tesla could be the name of the product or the company, and of course, 2003 is the date. So the reason why this technique is applied is because it is used for news analysis, search engines. So it basically helps to summarize and categorize information. Next, we have the stop words removal. So over here, the model simply removes words that don't add any kind of meaning or context to the actual text analysis. So your stop words are words like this and bot, that, and so on. So as an example, the cat is sitting on the Mt. After stop word removal has been applied, it simply becomes cat sitting Mt. That's all model needs to know. It needs to know that cat sits on that. Okay, that's it. It doesn't need to know the cat is sitting on the mat. That's way too much information, right? So why is this useful? Well, it can improve your search engine results, and then it makes your natural language processing models more efficient by simply focusing on the key words. All the extra words, the extra noise is all filtered out. Let's just concentrate on the key words. And then sentiment analysis, this basically tries to detect if a text is positive, negative, or maybe just simply neutral. So as an example, if I said, I love this product, then it knows, Okay, most likely, this is a very positive sentiment, right? But if I said this is terrible service, it knows, Okay, that's negative. But if I said the movie was okay, that possibly could be neutral, right? So why is this useful? It's used in customer reviews, social media monitoring, and so on. And then it helps companies understand public opinion and also their customer base as well. And then text classification. So over here, categories will be assigned to text based on the content. So as an example, it could either be spam or not spam emails. News categoration could be politics, sports, entertainment, fashion, technology, and so on. And then, of course, product review classification. It could be positive, negative, or maybe even neutral. So why is it useful? It helps automate email filtering, your news aggregation, for example, your fraud detection, and so much more. All of these use text classification, and then machine translation this simply converts one language to another. Whenever you're trying to convert English to French or Spanish or Russian or whatever, it uses machine translation. So for example, in English, how are you? French, I believe, will be como Sava. Of course, popular tools like Google Translate, your Microsoft Translator, they all use machine translation. It's useful because it, of course, breaks down language barriers, and it's used for international travel, international business, international negotiations, and so much more. So NLP in action, I want to give you an example in here and show you how the different techniques will be applied. So for example, a customer has bought a phone, okay? But the latest iPhone or the latest, you know, Android phone or whatever, and they say, I love this phone. The battery lasts all day. So when you type that in, how will the NLP model actually operate to break down that input? First of all, tokenization. I love this phone. The bats all day is going to break it down to, I love this phone. The battery lasts all day. That's basically tokenization. It's booking down the entire sentence into smaller categories, and then the POS tagging can come in. It knows that, okay, love here would be the verb. It knows that the phone is going to be the noun. It knows that battery is also noun. But what about the NER? Well, in this case, right now, because we said, I love this phone, the BLAs all day, there are no entities in here. Now, if the customer had said, I love this Apple phone, then any A will recognize that, Okay, Apple is probably the company or the business manufacturing the product, right? But because over here, I didn't mention Apple, I didn't mention Android, there are no named entities in here, then the stop Word removal comes into play. I love phone battery lasts all day. That's basically what is going to reduce the sentence to. So words like I, this, the all removed, we simply now have love phone battery lasts day, right? So, of course, sentiment analysis, because of the words like love, it knows that, okay, this is very, very positive. And then text classification. It's basically a product we view. So that's NLP in action. So to round this up, a few key takeaways, togenization, breaks text into words or sentences, ematization and stemming will simplify words to their base forms. Ps tagging will help to identify grammatical roles. Your NER can extract names, places, and dates, stopb removals, will clean up text for analysis. Sentiment analysis will try to detect emotions, and your text classification will categorize content into meaningful groups. And then finally, machine translation will convert text between languages. Thank you for watching. I will see you in the next class. 29. NLP Models and Approaches: Let's now take a look at the NLP models and approaches. Starting off with the traditional NLP approaches. There's two of them. We have the rule based NLP or symbolic AI and also the statistical NLP. So what is the rule based NLP, right? What is the symbolic AI? As the name suggests, over here, the model will simply follow written rules or a structure to process language. So it's typically based on the if then statements. As an example, if A equals B, then B equals A, right? It's very, very simple, it's very direct. So it works well for structured and predictable language tasks. As an example, you're developing the chat booard for a business. Now, if somebody were to contact that chat board and said, hello, a very safe response from the chatboard would be, Oh, hello. Good afternoon. How can I help you today? You don't need a human to type that the chatboard can respond or type that back to the customer because they said hello. It is a very, very safe response to give. So the pros for this are that, of course, it's very easy to understand and interpret, and it also works well in limited, structured environments. The cons obviously would be that it cannot handle variations in human language. For example, instead of saying hello, what if the customer said something previously, like, I had a very nice dinner last night. How are you today? The chapel might become confused, like, Wait, hold on a second. What is this, right? So whenever there's a variation in the human language or the responses given by the human, these kinds of models will not perform well. They also require extensive manual rule creation. Think about it, okay? You're basically teaching the chat board how to respond to different types of inputs. So it requires extensive manual rule creation. So it's typically used in your EA chat boards, your grammar checkers, and so on. But what about the statistical NLP? So instead of using the predefined rules, like in your symbolic AI, over here, the models will use probability and statistics to analyze text. So models, they can learn from large datasets instead of the predefined rules. And then they often rely on what we call the N grams, which are sequences of words to predict text. So as an example, a spam filter, right? If it's saw in the email, if it saw, for example, congratulations. You're today's lucky winner of $1 million. It might be able to predict that most likely this is spam because your typical spam emails, they have the keywords like congratulations, $1 million, today's winner. So given the fact that the email now has all these three keywords, the model can make a prediction that most likely this is spam. But what if the email had the words launch at 2:00 P.M. Then it knows that most likely this is in spam. I mean, how many spam emails have you ever gotten that said launch today at 2:00 P.M. That would be very, very strange, right? So with statistical NLP, the models try to make predictions. So the pros are, of course, they're more flexible. Then the rule based approaches, they can adapt to different languages and also different types of text variations. The cons, however, are that they still struggle with deep understanding. For example, it cannot understand whether the input is being sarcastic or is trying to be funny. It can't understand emotions, right? And then it requires large datasets for accuracy. They used in your email spam filters, and also the key word based text classification. So you've talked about the traditional learning methods, but what about the machine learning based natural language processing? So the improve NLP by learning patterns from data instead of relying on rules. There's three of them. You have the traditional machine learning models, deep learning for NLP, and also the transform based NLP models, which is of course the modern standard today. Let's take a look at them one by one. First is the traditional machine learning models. We have the nav base which classifies text based on probabilities, so it could identify whether or not the email is spam or not spam. Then we have the support vector machines, your SVM. They identify text patterns for classification and also decision trees. We talked about decision trees earlier. They use for branching logic to classify text. The pros of your traditional machine learning models are that they're more accurate. Than the rule based methods. They also work very well for classification tasks. However, the cons at that they require feature engineering. Remember we talked about features, which are the most important parts of any kind of data. So because it requires feature engineering, this would be the manual selection of the text attributes, and then it cannot process long term dependencies in language. Because of this, it's used in your sentiment analysis, your text classification, and of course, very simple chat bots. What about the deep learning NLP? Well, this basically revolutionized NLP by using, of course, the neural networks to process language more naturally. There's two of them. You have the RNNs, of course, the recurrent neural networks and the long short term memory networks, your LSTMs. So start off with the recurrent neural network. They're designed for sequential data like in sentences, right? And then they are able to remember the previous words when processing text. So they use for your speech recognition, your text prediction, and so much more. Now, the pros of your in and are that they can capture context and also order in sentences because they remember the previous words. And then they can also generate human like responses, which is actually quite fascinating. However, unfortunately, they do have some cons. They struggle with long sentences, right? They forget sometimes the earlier words. They do have short term memory, but the longer the sentence becomes, the more difficult it will be for them to remember the initial words in that sentence. And then they are slower to train compared to the simpler models. So they're used, of course, in your speech to text, predictive text, and so much more. But what about the short term memory networks? They're basically an improved version of the R and ends because they're able to remember much longer sentences. So they're used for long form text analysis and conversations. As an example, they could be used to summarize long article while still retaining the meaning or the key points of that article. So the pros, they handle large sentences much better. They're good for chat boards and text summarization. The cons, however, at that, even though they do have long term memory, they still struggle with very long documents. So now you talk about documents that are like five pages, six pages, and so on. And then it's also very expensive to train. They're used in chatbots, machine translation, and of course, text summarization. But what about the transformer based NLP models, the modern standard of today? They also revolutionized NLP by processing entire sentences at once rather than processing word by word. So far more efficient, right? So the models like your BRT, your GPT, they use self attention mechanisms to analyze words in context, and then they're also able very, very important, they're able to learn the relationship between words across long sentences. So they're able to understand context behind a sentence. As an example, if the input the bank is on the left side of the river. Now, naturally, when the word bank is heard or is used, you might be tempted to think, Okay, we're talking about the place where people keep money, right? But because we're talking about the bank on the left side of the weaver, the model here knows that, Oh, you're referring to the river bank. But if I said, I need to withdraw money from the bank tomorrow, now the model knows that, Okay, you're talking about the actual financial building because you've said words like withdraw money. Okay, you're obviously talking about the actual bank bank, right? So the pose, they can understand context deeply. They can generate human like text. They perform very well in complex NLP tasks. Unfortunately, just like everything else, they also do have their cons. They require massive datasets to train, and of course, they can also generate biased or incorrect responses. They're used, of course, in ChagpT Google's BRT search engine, Air text generation, and so much more. So I do have a table in here that compares the different NLP approaches, the pros, the cons, the examples, I'm going to provide for you this slide so you can study this at your leisure time. But before I round up, a few key takeaways, NLP started, of course, with the rule based methods, but has evolved into deep learning models. Statistical NLP introduced probability based text analysis, machine learning, your NLP, improved text classification, and sentiment analysis, and then of course, the deep learning NLPs, they enabled more complex tasks like your speech recognition, text analysis. And finally, of course, the modern standard for today, the transformers. They are now the most advanced NLP models used in air assistance and the chatbd. Thank you for watching. I will see you in the next class. 30. Large Language Models (LLMs) and Transformers: Come back, so now, let's take a look at the large language models and transformers, of course, these are the modern day standards for AI models. So what exactly are the LLMs, the large language models? Well, basically, they are AI models that are trained on massive amounts of text data to understand and generate our human language. Now, they use deep learning techniques, particularly transformers to process text efficiently. Examples, you have them in your GBT, which is of course the generative pretrained transformer. That's what is used in the hat GPT AI model. You have them in your BRT, your T five, you have them in your palm your mistrial cloud and so on. Now, what are the key features of the LLMs? Well, first of all, they can actually generate human like texts. They're also able to understand context and meaning in conversations. And because of this, they are typically used for text summarization, translation, answering questions, and so on. Now, I want to talk about the transformer architecture because this is basically the heart of our LLMs. So deep learning models, as we know, they are the modern day standards. They have replaced the older natural language processing architectures like your recurrent neural networks. And they use a mechanism called self attention to process text. So basically, instead of reading words one by one like your R and Ns, they are able transformers. They are able to analyze all words at once while also understanding the relationships between those words. So the key components of a transformers, first of all, the self attention mechanism they can understand the relationships between words in a sequence. And then they use something called the positional encoding. This helps them keep track of the word order. That also the multi head attention, this allows the transformers to focus on different aspects of the text simultaneously at the same time. And then also the feed forward new networks, they process word embeddings for output. So why are transformers better than the older models? I do have a table right here. Features I've talked about handles long texts. Transformers are capable of processing words in parallel. They also understand context, not just on a soface level, but actually deeply, they can understand in much greater detail context. And then, of course, they're also faster trained than your older models. So how exactly are the LLMs trained? Well, like with most AI models, you first of all, have the pre training where the model will learn the language patterns from massive datasets. So they provided lots of data, and this data could be gotten from the Internet, from books, from articles, websites, you name it. And then it comes the fine tuning stage. This is where the model will be adjusted for specific kinds of tasks. So maybe you are training an AI model for healthcare the language will be fine tuned so that the model can learn a bit more about terms under healthcare. Maybe you're training a chat board for legal matters and so on. So basically, the AM model will be fine tuned for what particular task it is meant to serve. And then comes the stage of inference. So over here, the model that's been trained already will be used to generate text, translate languages, and then also answer questions. This is basically the deployment stage. As an example, your JBT models, they are pre trained on Internet skill data and then also fine tuned to provide responses that are similar to that of a human. Examples of real world applications, again, in your conversational AI, like a hat GPT, Deepsk your chat boards, they're using search engines, they use for text summarization and content creation as and then of course, in language translation and also in code generation and programming assistant. These are just examples of read applications of our transformers. So just like with any other kind of AI model out there, we do have certain ethical challenges. Again, bias in the AI models. Hallucination and of course, misinformation is very, very possible for the AI models, no matter how complex they are, no matter how smart they are, they can always make mistakes. And, of course, the issue of data privacy. I've talked about this before already. Whenever massive amounts of data are needed to train an AI model, there is always the fear that data used could belong to that of users, customers, that could lead to privacy concerns. And, of course, also the environmental impact. Don't forget, of course, training models like this requires very high computational power. So what effect would that have on the environment? But what about the future? What can we look forward to in the future for the LLMs? Well, first of all, we can always expect as, you know, smaller models more efficient that will require less computing power. Also, better multimodal AI so regarding the processing of text, videos, images, it's going to become better. And then AI with stronger ethical safeguards, we hope to prevent pious and misinformation. And then, of course, LLMs being integrated into everyday tools like your transportation, communication, healthcare, and so on. So a few key taaways before I round up the lesson, LLMs are powerful AI models that can also understand and generate text. They enable the transformers. They enable the LLMs to process texts efficiently and also understand the context behind the text on a much deeper level. LLMs they're using your chat booards, search engines and so on. Challenges will include, of course, bias, misinformation, hallucination, and then the future of LLMs will focus on efficiency, multimodel AI, and of course, ethical improvements. Thank you for watching. I will see you in the next class. 31. Speech Recognition and Conversational AI: Come back, so now let's take a look at speech recognition and conversational AI models. Of course, these are models that use NLP to a very great degree. So what exactly is speech recognition? It's also called the automatic speech recognition, and it's basically the process of converting spoken text or spoken language rather into text. So it's what enables your dictation software, your voice assistants to understand what it is that you're actually saying to them. Now, examples your voice assistants, Alexa, Siri automated or call center. So whenever you call a business and then the machine picks up your call and says something, that's speech recognition in action. And then also in your captions, whenever you're watching videos on YouTube or on Netflix, you see the captions. That's basically speech recognition in action. But how exactly does it work? It's five main stages, and the very first stage would be the input, the capture of the input. So the user would need to either say something on the microphone some sort of inputs to be captured. Next comes the feature extraction. So over here, the system will try to convert the speech into a spectrogram. A spectrogram is basically a visual representation of what the speech looks like or what the sound looks like. And then after that, we'll now have the acoustic model where the model will try to match the sounds with phonemes. Phonemes are the smallest unit of speech. So based on what it's been able to try and match, it will then try to make a prediction as to the most likely words or sentences being spoken. And then finally, it'll produce the final transcribed text which will be generated. So that's basically how it works. There are certain challenges though involved with speech recognition, of course, accents and dialects. A system might be better able to understand an American accent or a British accent as opposed to, let's say, a very thick Indian accent or a very thick Russian accent, that's an example, right? And then, of course, the background noise. If there's plenty of noise in the background while the audio is being captured, that could affect how the system performs. And then of course, homophones, okay? Homophones are basically words that sound the same, but actually have different meanings, like, for example, we and right. So you write the verb to write something and then write which is the opposite of left. When you say that in the system, it might find it very difficult to distinguish between both words because they sound exactly the same. But what about conversational AI? This feels like the next level, right? So here, this allows computers to engage in human like interaction. So it powers your virtual assistance, your chat booards and, of course, customer support automation. Types of conversational AI, we do have the rule based chatbards. Remember, we talked about this earlier. These are basically preprogrammed responses for very specific kinds of questions. These kinds of AI models that can function, but in very limited environments, and then AI powered chat boards. These learned from user inputs, they improve over time. We have your voice assistants that can understand and also respond to spoken commands. So examples of your conversational AI, your Chat GPT, your Google Duplex, and of course, Alexa and Si. How does it work? A bit similar to the speech recognition, first of all, there has to be some sort of input provided, so the user will speak or type a query. Next comes the application of natural language understanding where the AI model will try to extract the meaning from the input. And then based on what's been able to extract, it will try to provide the best response to the input. It's called dialog management. And then the AI will now generate a response, human like response, typically, of course. And then in the last stage, the response will be delivered via voice or text. That's basically the speech text output. We do have challenges, of course, understanding context, AI can misinterpret complex or ambiguous queries. That's still a bit of an issue. And then, of course, the bias in AI responses, and then also handling multi turn dialogue. So conversations with multiple topics can confuse the AI. You can actually try this. You can try engaging with a jibty or maybe even Dipsk. Start of the conversation on, let's say, technology, and then asking a question on sports, change it to fashion. So somewhere along the line, it's very, very possible that the model Tha chibit in this case, right now, will get confused and start hallucinating its responses. So what are the solutions to these challenges for both the speech recognition and conversational AI? We can train AI on diverse datasets to improve language understanding, find models to handle different acets and speech variations as well. And, of course, using hybrid models, hybrid models would be rule base plus artificial intelligence for better responses. So what is the future? For speech recognition and conversational AI models. Well, more natural conversations. It's going to feel even more and more natural when you chat with these models. And then multimodal AI, which will combine speech, text and images all at once to enhance interactions. And then, of course, personalized AI assistant. So AI that will adapt to individual speech patterns and also preferences. We don't have that yet, but it's coming soon. And then of course, real time AI translation where instance speech to speech translations in multiple languages can occur. So some key takeaways, speech recognition converts spoken words into text using AI models. Conversational AI enables human like interactions through chat boards and voice assistance, and then challenges, of course, will include accents, dialects, noise, bias, and so on, homophones, and then also future advancements, right? So future advances will improve personalization, real time translation, and of course, multimodal AI. Thank you for watching. I will see you in the next class. 32. Section Preview The Future of Artificial Intelligence: Welcome to the final module, the future of artificial intelligence. And, of course, I'm going to play you one final clip from a movie to introduce this module, so sit back for lex, enjoy the clip, and I'll see you at the end of it. Everybody good? Plenty of slaves from my robot colony. Give them a humor setting so we'd fit in better with this unit. Thinks it relaxes us. A giant, sarcastic robot. What a great idea. I have a lot I can use when I'm joking, if you like. That'd probably help. Yeah, you can use it to find your way back into the ship after I blow you out the airlock. What's your humor setting, Tars? That's 100%. Let's bring it all down to 75, please. Hey, Tars? What's your honesty parameter? 90%. 90%. Absolute honesty isn't always the most diplomatic nor the safest form of communication with emotional beings. Okay. Well, come back. So that clip was taken from the movie in testla released in the year 2014 by the legendary director Christopher Nolan. Now, to be fair, there are so many other clips I could have chosen to introduce this final module. However, I chose this particular clip because I thought it was very fascinating and also demonstrates effectively the future of artificial intelligence. Now in the clip, we have the AI model called TAS that's helping the astronauts launch their space launch their spaceship into space. And while they're taking off, the Tar Air model begins to make some jokes. So Cooper, the main astronaut, he asks Taz. He says, What are your humor settings, and Taz responds. Oh, it's at 100%. And, of course, Kuper doesn't like this, and he says, Okay, let's bring that down to 75%. And later on in the clip, Kooper asks Taz What is your honesty parameter? And Tas responds that it is at 90%. And Cooper, of course, asks 90%. Meaning, why isn't it at 100%? Now, Tarz, the AI model is very, very smart. It knows that Kuper is asking, why is it at 90%? Why not at 100%? And Tarz rather humorously responds that, Oh, absolute honesty isn't the most diplomatic or the safest way to communicate with emotional beings. And I thought that was really, really funny because it's true. Think about it, okay? There are so many times in everyday life in everyday conversations where you might want to say something you might want to tell someone how you really feel on the inside, but because you're concerned that they might get upset, they might get offended with what you say, even though it's true, you then decide, Okay, I'm gonna play it safe and not be so straightforward and not be so blunt in what I say. So I just thought was very, very fascinating that the AI model Taz knows that 100% Honesty isn't probably the best way to communicate with human beings. So one other thing that we observed in this particular clip is the ability to personalize our AI models. Here you have Cooper being able to adjust both the humor and honesty settings for tars. And this is something that will eventually happen in the future with our AI models. We will be able to personalize them. They might begin to sound just like horse, speak with our accents, and so on. So that's something that's eventually going to come. Personal customization of AI models. So I hope you enjoyed this introductory video to our final module, the future of AI. Let's now begin with the rest of the lessons. 33. Current Trends in AI Development: Welcome back. So let's begin a new module by talking about the current trends in AI development. And when you look around you, regardless of what industry or field it might be, there is already some presence of artificial intelligence. But let's begin by talking about AI in automation and the workforce transformation. We already have the increase in use of AI driven automation in industries such as manufacturing, logistics, retail, and so on. And, of course, the growth of what we call robotic process automation, the RPA to handle repetitive tasks. You find this in companies, in businesses. They use this for chatbards, customer handling, and so much more. And, of course, AI powered customer service chatbod and virtual assistant, your Google seri and so on. And then the shift towards the human AI collaboration. But what about AI in the creative industries? So now you have AI models like Dali, M Jony, that are able to generate realistic looking images. You have AI models like Sra from Open AI, the same company that developed a ha JBT that are able to convert text into videos. And then AI generated music and voice cloning. We have those as well. And, of course, we do have some ethical concerns regarding AI and creative industries, especially when it comes to the issue of deep FAC. We've already had several incidents where several criminals were able to use Deep fake to trick their victims. And, of course, we have AI in our everyday lives and personalization. For example, you have AI in use for Netflix, for YouTube, Spotify, Disney plus, and so on, and, of course, AI in e commerce with the use of chatbots. And of course, personalizing our shopping experiences, and of course, voice assistants and smart devices like your Google Series, Amazon's Alexa, and so on. And, of course, AI in social media. We now use AI to generate content as well. And, of course, AI in healthcare and biotechnology. This is actually very, very fascinating because we now have AI driven diagnostics, where AI can be used to detect diseases from X rays, MRIs, and so on. And then AI also used in drug discovery, where AI has been used to accelerate our research purposes and pharmaceuticals as seen in Dip minds alpha fold, which is able to predict protein structures. I'm not going to pretend I know what exactly that is, but we have AI assisted surgeries. AI has become so advanced nowadays that it can help in surgery with little to no human intervention, and of course, predictive analytics in medicine. We now use AI to predict disease outbreaks and patient health trends. And of course, this course will not be complete without talking about the use of AI in finance and, of course, cybersecurity. So we do have AI being used in algorithmic trading where AI driven financial strategies that analyze market trends in real time have been developed. And, of course, for cybersecurity, we can use AI for fraud detection. And AI part cybersecurity where AI can be used for threat detection, automated responses, as well as vulnerability assessment. Moving on, what is the future of AI integration, three main points in edge AI, where AI will now run directly on mobile devices rather than relying on cloud computing. And then AI powered IOT, the Internet of things. So we'll now use AI to power our smart homes, smart cities, smart networks, and so on. And, of course, the AI democratization where AI will become more accessible to those who don't have a technical background. So there are some key takeaways in here. AI, as you know, is transforming multiple kinds of industries from healthcare to finance to Hollywood. And of course, the combination of AI and automation is reshaping the workforce. But human AI collaboration is going to be key. And finally, we do have ethical considerations, regulations, and responsible AI development, which will be essential as AI becomes more integrated into our daily lives. So that's thank you for watching the video. I'll see you in the next class. 34. The Next Frontier – General AI vs: Welcome back, so now. Let's take a look at a very fascinating topic. And here, we're comparing general AI with narrow AI. What is narrow AI? Now, we've already talked about this previously. These are AI that excel at performing a specific kind of function. And we have numerous examples of this. You have your Amazon's exa, Google's seri and so on. And of course, recommendation algorithms like YouTube, Spotify, Netflix, and even your AI models like your Chi JPT, Claude, Dipsik. All these are examples of narrow AI. But we do have some key characteristics for example, they are highly specialized, meaning that they can excel at one particular kind of task. And then of course, they are data driven. They require massive amounts of data. They also lack reasoning as well. They are unable to understand concepts beyond their training and, of course, no true autonomy, meaning that the AI models the weak AI, they can either make decisions based on rules or simply learned behavior. In other words, they are very, very flexible in how they perform their tasks. But the thing is, despite the fact that today, when you look at models like ChaGPT and Siri and Deep Seek and so on, they all seem very powerful and quite competent in what it is that they do, but they're still considered to be weak AI. And that's because we do have the theoretical and possibly the practical possibility of general AI. Now what is general AI? This refers to AI with human level, cognitive abilities, capable of understanding, reasoning, and learning across multiple domains without any prior training without being explicitly trained for any one of these tasks. So in other words, we're talking about artificial intelligence that can truly match or maybe even possibly surpass human intelligence. So what would AGI be able to do? Well, AGI, General AI we'll be able to understand and learn any subject, just like any normal human being, solve new unfamiliar problems without any prior training. So basically, it'll be able to reason and solve new problems on its own, show creativity, as well as common sense reasoning, and then adapt to different environments without any extra programming. It's going to become very adaptable, very flexible, and then possess self awareness and independent thought. But this is still highly debatable. There are those who believe that, yes, we might eventually get artificial intelligence that'll be so intelligent, that'll be so powerful. I'll be capable of independent thought, self awareness, while many others don't think this will ever be achievable. So what are the current AGI research efforts? We do have companies like Open AI, Deep Mind, and also anthropic. These are among the companies that are working towards achieving general AI. AI models like GPT version four. Even though they're extremely powerful and they're becoming more general, they're still not exactly truly general AI. And now, some researchers, those who are very optimistic, those who do think that we will achieve AGI, they see the time frame 10-50 years. Well, like I said earlier, there are those who don't believe that we'll ever achieve AGI. So what are the key characteristics of AGI? Well, first of all, learn from experience just like a human being, can transfer knowledge between different tasks. So say, for example, you've given task one to general AI, it performs that task. You then give it task two. If there are some similarities between task one and task two, it might be able to transfer the knowledge it gained from working on task one on to task two, again, just like a normal human being, and then show reasoning, problem solving, and adaptability, and then potentially autonomous in decision making. In other words, be capable of independent thought. These are the key characteristics of AGI. I do have the table in here that I've shown you the key differences between general AI and, of course, weak AI. And, of course, in most of these features, general AI surpasses narrow AI. The only thing, though, is that when it comes to current existence, we do have narrow AI. It's already a real thin, while general AI is still theoretical at this point. So what are, in fact, the challenges in achieving AGI? What's the hold up? Why aren't the developers at open air, and so why haven't they given us AGI just yet? Well, as you can imagine, we do have the technical challenges. And if you think that deep learning requires massive competitional power, that is nothing compared to the kind of competitional power required for AGI. And then data efficiency, of course, AGI will require enormous amounts of data, which is still a bit of a challenge. And then common sense reasoning, AI, as we know it today, still struggles when it comes to understanding abstract concepts. It's unable to reason and decipher what they are. And then memory and adaptability, AGI should be able to show the ability to retain and apply knowledge across different scenarios. In other words, the kind of knowledge and intelligence AGI must possess, it's extremely difficult to achieve them. We also have the ethical and safety concerns of AGI. What if AGI does in fact surpass human intelligence? How do we control it, right? That's always the big question. And, of course, bias and fairness. Now, to be fair, no pun intended, this is a big challenge across all types of AI and not just general AI. And then AI alignment, how do we ensure that the AGIs goals do align with human values and of course, the potential risk. What happens if the bad guys, if the cyber criminals get their hands on general AI? The consequences could be disastrous. And then, of course, the philosophical and theoretical questions. Can AI be conscious? Right? Imagine that. Philosophers debate whether AGI could have subjective experiences. How insane would that be? You're almost at this point talking about artificial intelligence, having emotions. I mean, that's quite close, right? And then will AGI replace humans? Will the human race cease to exist because we now have AGI running the world? Well, some fear that AGI could outperform humans in all tasks leading to job displacement or even worse. And then perhaps the biggest question of all, should we, in fact, create AGI? Just because we can, does that mean that we must or that we should? Maybe sometimes it's best to just say, Hey, look, narrow AI that we have now in existence, it's good enough. We can improve on it. But at some point, we need to say, Okay, this is becoming way too advanced. This is becoming way too intelligent. We need to take a step backwards. So these are kind of, like, the very interesting philosophical questions that have been asked. So the road to AGI, where are we now in 2025? Well, artificial intelligence systems are getting more powerful, but they still lack true understanding. Now, some researchers do believe that AGI will require fundamental breakthroughs in neuroscience, cognitive science, and machine learning as well. And others also do propose that the use of habit models, those that combine symbolic reasoning with deep learning, this could draw us closer to AGI. And then AGI regulations and policies are increasingly important to guide the ethical development of AGI. But what if we do, in fact, eventually achieve AGI? What are the possible future implications? Well, super human intelligence could AGI surpass human intelligence and revolutionize every single field. That is a possibility. And then, of course, the job markets. The thing is, we don't even have to go as far as AGI. Look at what's happened today. Weak AI, Chagpt and its bodies, they're already replacing so many people. So many jobs are already been lost because of the introduction of narrow AI. So now imagine what will happen when we now have general AI being introduced. That's possibly going to displace even more people. Even more jobs will be lost as a result. And, of course, the ethical AI governance. How do we ensure that AGI remains beneficial and does not become ddius? It doesn't fall into the wrong hands. And now human AI collaboration could AGI, in fact, work alongside humans, enhancing our capabilities rather than replacing us. I'm pretty sure you've seen movies like terminator, and so on. In those movies, AGI choose It's says, You know what? I'm not gonna walk with humans. Humans are a threat to my existence. I'm just going to destroy all of mankind. So that's what happens in the movies. Hopefully, it doesn't happen in real life. So just a few key takeaways. Networ AI is everywhere today. AGI is still theoretical at this point. AGI, if we do in fact, achieve AGI, it will be capable of reasoning problem solving and an adaptation, just like a human being, and achieving Aga does poses massive technical, ethical and safety concerns or challenges. And then the future of AGI could reshape industry, society, and even humanity itself. So, are we going to achieve AGI only time will tell. Thank you for watching the video. I will see you in the next class. 35. AI and the Workforce – Will AI Replace Jobs: Come back, so let's take a look at the next lesson. And of course, this is the million dollar question. Will AI replace your job? Well, let's find out. First of all, let's talk about the current effects of AI on the workforce. Now, AI has been used to automate very repetitive and boring tasks. AI has been used to improve efficiency. And, of course, with the introduction of AI, new kinds of roles have been created, new kinds of carriers have been created as a result of AI. And because of the introduction of AI, workers are going to need to learn new kinds of digital skills in order to survive. Think about it, okay. Imagine a worker today who doesn't know how to use the Internet. It's almost impossible to get by, right? So eventually at some point, we'll all have to learn some basics of AI in order to be employable. So what are the jobs mostly at risk of AI and automation? You have those in manufacturing and logistics where we now have AI powered robots that work in the assembly lines. They replaced all humans. You have those in retail and customer service. We now have AI powered chatbots that can do the job just fine. You have rules on that data entry and administration. Of course, AI can now handle spreadsheets and, you know, the document analysis and so on. And then on the transportation and delivery, even though this hasn't yet taken full effect, but eventually we're going to have autonomous trucks and self driving taxes that can do the job. But it's not just all doom and gloom and AI is going to replace us all. New kinds of roles will be created as a result of AI, for example, AI and machine learning engineers. Obviously, we're going to be creating new kinds of models or re training and improving existing AI models. And then data scientists and analysts don't forget that data is the lifeblood of AI. So we're going to need data scientists as well. And of course, cybersecurity experts who will use AI to perform their tasks, and then AI trainers and ethics specialists that will ensure that AI models are aligned with ethical standards, and of course, the human AI collaboration specialists managing AI human workflows in industry. So these are a few examples of the kinds of careers that will grow as a result of AI. But we also have some careers that will be enhanced by AI. For example, under software development, we're going to have programmers who will be able to use AI to improve their levels of programming, AI Pow coding assistance, and so on. And then under healthcare, AI will be able to aid doctors in diagnostics, surgery, and patient care as well. And even in the creative fields in the creative industry, where AI tools can help designers, writers, musicians generate new ideas. And there's a few other professions where AI can enhance them, also in cybersecurity, as well as a cybersecurity specialist myself. AI can be used to detect and deter cyber attacks. So ultimately, the question right now is, will AI completely replace human workers? What do you think? In my humble opinion, I do believe that AI inevitably will replace many millions and millions of kinds of jobs will be lost as a result of AI. While new roles, new jobs will be created, new opportunities will be created as a result of AI. I don't know what the economic implications are going to be because not everyone whose job has been replaced by AI will be able to get a new job. So what happens to them, right? I don't know. It's something to think about, but we'll see. We'll see what will happen. So just a few key takeaways before we round up the lesson. AI is automating some jobs, but also creating new opportunities as well. I think the idea here is that you should just position yourself to take advantage of the introduction of AI because, like it or not, AI is here. It is the present, and it's also going to be the future as well. So low skilled repetitive jobs are at a higher risk of automation. AI AI enhances rather than replaces roles in many creative and analytical fields. And, of course, adapting to AI driven workplaces will require upskilling and lifelong learning. Like I said earlier, because of AI, we're all going to be forced to learn some basics of AI. And, of course, the future is human AI collaboration, not total automation. Thank you for watching. I will see you in the next class. 36. AI and Superintelligence – Hype or Reality: Well, come back. So to round up this module, let's take a look at our final lesson. And here, we're discussing AI and super intelligence. Now, to be honest, I wasn't sure if I should make this a lesson because at least in my humble opinion, it's very highly unlikely that we're ever going to achieve this, but nevertheless, it is a fascinating topic. So let's talk about it. Now, what exactly is super intelligence ASI? Well, this is basically AI that's going to surpass us humans in every aspect, including creativity, reasoning, decision making, and so on. So basically, we're talking about intelligence that will become our masters. Now, unlike narrow and general AI, ASI will be self improving autonomous potentially far exceeding human cognitive abilities. So as an example, if we do eventually develop ASI, the superintelligence itself could design better versions of itself, rapidly accelerating its intelligence beyond human control. So that old theory about AI taking over the world, it's going to become a real possibility if we do achieve super intelligence. So is it actually possible C we ever get to this point of superintelligence in AI we do have some factors that point towards it and other factors that say, nope, we're not going to get there. So what are the arguments for ASI becoming a reality? Well, first of all, computational power growth. Now, obviously, to power super intelligent models is going to require tremendous amount of computing power. But given the fact that computing power is increasing exponentially, there is no doubt that eventually at some point in the future, we'll have enough computing power to power such AI models. Now, advancements in neural networks as well. You have deep learning models that are becoming more and more sophisticated by the day. You have self learning AI where we have AI already capable of self improvement as an example, the Alpha zero. This is an AI model that learn how to play chess, and it learned by itself by simply playing games against itself. So breakthroughs in AGI as well. I AGI is eventually achieved general AI, then the next step after general AI is going to be ASI superintelligence. So some prominent AI researchers like Nick Bostrom, who is also the author of superintelligence, they believe that AI could become a real possibility sometime in the 21st century. But like I said, we do have arguments for ASI, but we also have arguments against ASI. First of all, the limits of computation, human intelligence is not just about world competent power. Ons consciousness is still an unsolved problem. So this is still a big challenge regarding developing ASI and then the lack of true general intelligence. Before we can get super intelligence, we need to develop general AI. We haven't even achieved general AI yet. Some people are already talking about super intelligence. So perhaps maybe we should achieve general AI first before we start talking about superintelligence. And then human creativity and emotion, AI will always lack curiosity, emotions, and the ability to experience the world. This is one of the biggest arguments against superintelligence there. No matter how intelligent it's going to be, it is still a machine. It is still not capable of developing emotions, right? And then ethical and technical barriers, right? The world may intentionally prevent ASI from emerging due to safety risks. So it could be that we've gotten the technical expertise. We have the computing power. But again, just because we can develop super intelligence, that doesn't necessarily mean that we should. So maybe that might be what actually stops us from being able to achieve ASI. So some skeptics like Gary Marcos argue that AI lacks true understanding and is unlikely to reach super intelligence. Let's move on. What are the risks and ethical concerns of ASI? And as you can see from this slide, you have a very scary looking robot smiling at you. That's obviously not a pleasant smile. It is a very, you know, evil looking kind of smile. That's, of course, the terminator. If you haven't seen the movie before terminator one and two, I would hello encourage that you watch it. It's a fun time. Actually Terminator two is my favorite movie of all time. It is number one for me. So just in case you're interested in some action, Scify, definitely check out the movies, but why am I using this particular image from the terminator? Well, that's because in the movie, you had a super intelligence that was developed. The model was called Skynet, and Skynet eventually decided one day that, you know what? I'm going to destroy humanity. I'm going to destroy mankind, and Skynet waged war against a human. So definitely check it out. So what are the risks and ethical concerns of ASI? First of all, the loss of human control. It is possible that the artificial intelligence will become so powerful, so intelligent that we as humans will not be able to control it anymore. And that the existential risk if ASI's goals don't align with human values, it could be dangerous. I actually could even be disastrous. So economic disruption. With narrow AI, lots of people losing their jobs. If general AI is introduced, even more jobs are going to be lost. But what now happens if the ultimate super intelligence is now introduced? Millions and millions of jobs will be rendered obsolete as a result of this new kind of technology. And then, of course, the autonomous decision making, could an AI decide that humans are inefficient or necessary? That's kind of what happened in the movie terminator. As an example, I want to show you, well, I'm not going to show you. You can check it out yourself. You can go on YouTube. The channel name is called Isaac Author I just an experiment called the Paper clip Maximizer. It was actually an experiment that imagined an AI designed to create paper clips. But the AI became so advanced and so efficient that it decided on its own that, Hey, I'm going to convert all matter in the world, including human beings into paper clips. It's actually a very, very fascinating video on YouTube. I think it's about 12 to 15 minutes. You can definitely check it out. Again, YouTube channel name is Isaac Ortho. Simply search for the Paper Clip Maximizer video, if you want to check it out. So, safeguarding against uncontrolled ASI, what can we do to ensure that if ASI is achieved, that it is under control? Well, first of all, AI alignment research, ensuring that the AI understands and respects human values. We hope so. And then regulatory oversight. Honestly, I am someone who isn't necessarily the biggest fan of government oversight and regulations. But in certain kinds of technologies like AI or in this case, ASI, I do strongly agree that some government oversight will be necessary. And then, of course, kill switch mechanisms, right? Imagine if ASI superintelligence decides that, you know what? I'm going to take out mankind. I'm going to kill all humans. We should have some kill switch mechanisms in place to shut down that AI immediately. And those kill switch is better work, right? And then, of course, the ethical AI frameworks. AI research must prioritize safety, transparency, and of course, accountability. As an example, open AI and deep mind. These are companies. They are actively researching AI safety to prevent uncontrolled AI growth, and hopefully they will succeed. So what is the coin progress towards ASI? Well, no existing AI has achieved AGI, let alone ASI. So like I said earlier, it's going to take a very long time if we do eventually get to ASI. So major AI models like your GPT, Dip Mine, and so on, they still do rely on human inputs. Again, we're far off from being able to get to superintelligence, and then some AI models, they can self improve in narrow tasks, but not necessarily in a broader or more general way. And an ethical AI discussions and regulations are increasing globally. So more and more governments around the world are beginning to recognize the impact of AI and are looking for ways to regulate the use of AI globally. So prediction, most experts believe that AGI, human level, AI could emerge within 50 years. But ASI, which is, of course, the ultimate super intelligent, artificial intelligence, is much further away if possible at all. Key takeaways. Well, ASI refers to AI that will surpass human intelligence. Some experts, of course, do believe that ASI is possible while others don't believe we will ever achieve it. The biggest risks of ASI include the loss of control, existential threats, and of course, the massive economic disruption. AI safety measures and ethical regulations are crucial to, of course, prevent unintended consequences. And currently, ASI, just like AGI is still theoretical, and AGI here hasn't even yet been achieved. So, honestly, I don't think in our lifetime we're ever going to get to the levels of ASI. AGI, I think will eventually get there, possibly in another 25 years or but AS, I don't think in a lifetime, we're ever going to get there. So maybe in the year 3,000 and something, maybe eventually the humans, then they might be able to achieve ASI. But that's it for the lessons thank you for watching, I will see you in the next class. 37. AI Course Conclusion: Well, congratulations. We've come to the end of this course on artificial intelligence. And from the bottom of my heart, let me say a big thank you for finishing this course, and I do sincerely hope that you found the lessons to be very entertaining, engaging, but most importantly, informative. And if you feel like you got your money's worth, if you feel like you like this course, you'll learn quite a lot. Please do consider leaving a written review. The reviews will help me a lot, and they will really help to boost this course, as well. So thank you so much for your support. Now if this is the last time I'll be seeing you in any one of my courses, let me just say, good luck. I hope that this course will help you in your everyday life. I may also give you that career boost that you've been looking for. And if I do see you in another one of my courses, maybe it's a cybsecurity course or a web development course or maybe another AI course, I will see you there. That will be amazing. Nevertheless, thank you so much, once again, for taking this course, for finishing the course. All the best, and I'll see you next time. Chess.