Goal-Driven Artificial Intelligence and Machine Learning | Alex Castrounis | Skillshare

Goal-Driven Artificial Intelligence and Machine Learning

Alex Castrounis, Author of AI for People and Business

Play Speed
  • 0.5x
  • 1x (Normal)
  • 1.25x
  • 1.5x
  • 2x
14 Lessons (1h 22m)
    • 1. Introduction and Overview

      4:13
    • 2. Business Goals

      1:39
    • 3. Customer Goals

      2:16
    • 4. AI and Machine Learning Definitions

      8:03
    • 5. Machine Learning Types and Algorithms

      7:07
    • 6. AI Types and Algorithms

      8:21
    • 7. The AI and ML Process and Tradeoffs

      12:09
    • 8. Recommender System Applications

      8:01
    • 9. Prediction and Classification Applications

      4:25
    • 10. Computer Vision and Recognition Applications

      7:25
    • 11. Clustering and Anomaly Detection Applications

      1:47
    • 12. Natural Language (NLP, NLG, NLU) Applications

      8:15
    • 13. Hybrid and Miscellaneous Applications

      7:50
    • 14. Summary and Next Steps

      0:22

About This Class

Alex Castrounis, author of AI for People and Business (official book trailer below), teaches goal-driven artificial intelligence and machine learning for executives, managers, and anyone else interested in learning more about these subject areas, regardless of technical expertise.

Are you interested in learning about artificial intelligence (AI) and machine learning (ML)? Have you wondered how these amazing fields can help you or your business? If yes, then join Alex Castrounis to learn all about these topics and more!

Artificial intelligence and machine learning are helping people and businesses achieve key goals, obtain actionable insights, drive critical decisions, and create exciting, new, and innovative products and services.

This class is relatively high-level so that non-technical folks can understand everything. Please note that this class does not include coding or code examples.

This class will

  • Give examples of common business and customer goals that can drive artificial intelligence and machine learning solutions
  • Explain artificial intelligence and machine learning, with a heavy emphasis on why they should be used
  • Provide an overview of the different types of tasks and algorithms associated with artificial intelligence and machine learning
  • Describe the typical artificial intelligence and machine learning process, along with important tradeoffs and considerations
  • Discuss real-world applications and examples of companies and products that are using artificial intelligence

I hope you enjoy!

For the latest from Alex, subscribe to his newsletter and YouTube channel, follow him on Twitter and LinkedIn, and grab your FREE chapter from his book.

-----

Images

  • Simple linear regression model: 
    • https://www.slideshare.net/dessybudiyanti/simple-linier-regression
    • https://image.slidesharecdn.com/simplelinearregressionpelatihan-090829234643-phpapp02/95/simple-linier-regression-9-728.jpg
  • Machine learning process: https://blog.sujeetjaiswal.com/machine-learning-an-introduction-de88d85ebc5d
  • Machine learning process: http://oliviaklose.azurewebsites.net/machine-learning-11-algorithms-explained/
  • Gradient descent: https://sebastianraschka.com/faq/docs/closed-form-vs-gd.html
  • Overfitting 1: Andrew Ng Coursera Machine Learning course
  • Overfitting 2: Python machine learning by Sebastian Raschka
    • Copyright (c) 2015, 2016 SEBASTIAN RASCHKA ([email protected])
    • License: https://github.com/rasbt/python-machine-learning-book/blob/master/LICENSE.txt
  • Decision tree: https://alliance.seas.upenn.edu/~cis520/wiki/index.php?n=Lectures.DecisionTrees
  • Multiple linear regression: http://gerardnico.com/wiki/data_mining/multiple_regression 
  • Support vector machine (SVM): http://dni-institute.in/blogs/building-predictive-model-using-svm-and-r
  • Artificial neuron model: https://commons.wikimedia.org/wiki/File:ArtificialNeuronModel_english.png
  • Artificial neural network (ANN): https://www.extremetech.com/extreme/215170-artificial-neural-networks-are-changing-the-world-what-are-they
  • Biological neuron: http://biomedicalengineering.yolasite.com/resources/neuron_structure.jpg
  • Equation of a straight line: http://starsdestination.blogspot.com/2012/11/conic-sections.html

Transcripts

1. Introduction and Overview: Hi, My name is Alex. Extras Welcome like last goal driven artificial intelligence and machine learning of super excited about this class, and I really hope that you find it to be highly informative and enjoyable. This class is intended to be relatively high level so that a non technical executive can understand everything, which should be very useful to technical practitioners as well. But before diving in the class further, a little on my background and ways to continue learning about these topics after you finish the class, I work many years in the professional motor sports industry, IndyCar racing and the Indianapolis 500 in particular. For those erase fans, I work with several successful drivers and team owners such as Michael Andretti, Al Unser Jr. Jimmy Vasser, willpower, Vanik, A Patrick, Tony Kanaan and Ryan 100 Rada. Name a few in racing, I was data scientist. Vehicle Dynamics is software engineer and race strategist for well over 100 races, including six Indianapolis 500 I wrapped. The live racing career is head of vehicle dynamics, data science and software engineering and been dreading another sport. I left racing to focus my career specifically on the tech industry and have held roles and software, architecture and engineering, data science and product management. I currently run data science and probably management and an I O T company in Chicago. Teach data science and probably management courses for General Assembly and regularly mentor entrepreneurs and other individuals. I also created right for a block called You Know architect at www dot you know, architect dot com, and many of the topics discussed on the architect are related to what will be discussing in this class. I'm always publishing new content, so be sure to sign up for the into architect newsletter for the latest content updates. You can also follow me at at, you know, architect on Twitter last week. If you're interested in learning more after this class, please visit my did. Hub Repository, located at a I dot resource, is dot into architect dot com. There, you'll find a curated side of resource is from data science, machine learning, artificial intelligence, big data, the Internet of things and much, much more. Please start the repository if you find it to be useful. All right, so enough about me again. Welcome to the goal driven artificial intelligence and machine learning class, artificial intelligence and machine learning. Really interesting, exciting and super powerful topics. They're enjoying a huge, a lot of hype. Impressed. So let's take a look at the sections in this class artificial intelligence and machine learning in and of itself a super interesting. But the true power and values in the application of these fields to solving real world problems and achieving goals from a why. How, what perspective It's all about why we use these tools and techniques unless about the way we use them. The how or the final result the what? Well, therefore begin with an overview of common business and constantly goals, their perfect candidates for AI and machine learning applications. Essentially, these goals will give us the why behind used these tools and techniques, and we'll keep these in mind throughout the class. A good example of this is Amazon's recommendation engine, which most of you are probably familiar with. You know, it's the customers who bought this item also bought some other, and one way to think about this is that Amazon built this amazing recommendation system or the what using a combination of sophisticated machine learning techniques or the house. But another way to talk about it is that some estimates indicate that Amazon has been able to increase sales revenue by up to 30% as a result of increasing average customer order, size and user engagement. And they did this by using a combination of sophisticated machine learning techniques in order to build an amazing recommendation system. And clearly there's a huge difference in both of these descriptions. And this class is meant to highlight that difference throughout. Well, then give the definitions of artificial intelligence and machine learning a lot with different types of each and the typical processes involved. The remainder of the course will be a tour of many different real world applications of AI and machine learning by category. There's also a project that companies this class, and you're encouraged to give it a go in order to truly get the most from this class. It involves a data set of movie ratings and other relevant movie information from IMDb Facebook. Thank you again for your interest in this class, and please enjoy 2. Business Goals: all right, so we've had an introduction of the class and we've talked about the different sections that will be covering this one is about business goals. Businesses have many different goals but often share very similar goals. So let's talk about what some of those might be so common. Business goals include things like increasing revenue, increasing profits, cutting costs, customer acquisition, retention and growth, maximizing customer engagement and delight. Now let's talk about that one a little bit more. In the case of product companies, product companies want to make their products sticky. And what we mean by that is that you know, a lot of products come and go And, like, even take APS on your phone, for example. You probably have tons of them, right? But you probably only really use like five or eight of them every day. And those are the absolute sticky. They're the ones that help you do something every day, and you go to them time and time again. And so ultimately, that's the goal is to make products sticky, where people want to use it on a recurring basis for a long period of time. But for our services, type company. Often, you know, you want to really emphasize customer service relationship, building with your customers and building trust and becoming a trusted partner. And so those are some of the very common goals. Another one could be improving operational efficiency. And as you can see, when you keep these goals in mind, you start toe Uncover that why we talked about earlier. And these are the things they're gonna allow you to use that as a driving force for Beacon . When you're choosing what AI and machine learning techniques in order to solve your problems or achieve your business goals in the next video, we'll be talking about customer goals, which will find are somewhat different than businesses. 3. Customer Goals: We've talked about business goals and the many different common goals they're sharing amongst different businesses. So now we're gonna talk about customer goals and you'll find that they could be quite different. So what we're gonna do is we're gonna put customer goals in the context of what's known as the jobs to be Done framework, which is part of a larger sort of concept. No or field known as outcome driven innovation. And basically the general idea is that people hire or buy products and services to get a job done. Products come and go, but the job to get done doesn't. Basically what you want to do is you break the job down in discrete steps. You focus on the job and not on the product or customer. And then you understand what results in a job being done successfully. And so what kind of jobs are we even talking about that people might want to get that? Well, people sometimes need help making decisions or help doing shopping or productivity or staying connected with friends and family and social media or, you know, certain kinds of business applications. But the key here is that you know customers want something that not only helps them get a job done, whether it's a product or service, but also that sticky. And we mentioned that concept in the business goals video. But to mention again, basically, it's the idea that, you know, once you've got something either a product or service that's helping someone get a job done , so to speak, you know they want to go to and use it to get the job done. Time and time again. It's their go to You don't want it to be just like a one off deal, although sometimes people do just need to do one off jobs, and that's fine. In that case, another thing to keep in mind is that customers want tohave a great user experience and want to use prize with great design, you know? And what what do we mean by that? So what it means is that you know the product is self evident. It's easy to use. It's easy to understand there's no extra fluff or unnecessary information, and so these are the kinds of things we want to keep in mind as we move for throughout this class again, because everything we want to address. You know how you use AI and machine learning in the context of these kind of customer goals and also the business goals that we talked about before and again, That's the wine. We want to keep it in the back of our mind. So in the next video, we're gonna dive right in and start learning a bit more about artificial intelligence and machine learning. 4. AI and Machine Learning Definitions: all right, so we've talked about business goals and we talked about customer goals, and we've talked about the why we want to use artificial intelligence and machine learning tools and techniques. So now let's talk about the definition of artificial intelligence. Before we dive into that, I'd like to note that there is no real definition of artificial intelligence per se, and what I mean by that is at least concrete Lee or definitively. If you look out lie, you probably find all sorts of different definitions people talk about in terms of intelligence they talk about in terms of things like thinking in terms of the concept of having a mind and so on. But let's start by looking at the definition of just intelligence non artificial. So often intelligence evolves. Something like learning, understanding and the application of that knowledge learned in order achieve one or more goals. So naturally you can think of artificial intelligence is basically being that where a computer machine does the same thing, it basically learns it understands, and it uses that learning and knowledge gain in order, achieve goals eso we can we could think of that is sort of our loose definition of artificial intelligence here. Artificial intelligence is a broad field and includes many different things. And so machine learning is actually a sub field of artificial intelligence, and we'll give the definition of that in the next video. But you may have also heard of deep learning, which is very popular these days and actually quite amazing. And that's actually a cell field in machine learning, which then in turn, is a some field of artificial intelligence, so it includes it as well. Now we're gonna talk about machine learning, which is a sub field of artificial intelligence. So the key thing here with machine learning is that there's no explicit programming, and whether I mean by that is, you know, before machine learning algorithms got to be used by people you know, oftentimes, uh, software engineers or programmers when needed to write explicit code to solve certain tasks based on some data they had. For example, So, you know, when you're writing code, you write different kinds of loops and things like that. So you say for this, do that if this, then that and selling it self worth. And so it's very explicit machine learning on the other hand is different in the sense that you don't writing explicit programming language stuff like that at all. But rather you hand certain algorithms, machine learning algorithms, some data, and then the algorithm automatically learns from that data and creates a result. And most often that result is some sort of model, uh, and in fact, often a predictive model, which will talk about a bit more later. So what What do I mean by learns from the data will typically, that the learning part involves learning optimal parameters or coefficients to get the best performing model. So some of you may remember when you took algebra back in the day. The equation of a straight line, which was y equals mx plus B i m. Is a Slope B is the intercept. So in this case, M as a slope is a coefficient of acts, and it multiplies against X. And that's how you get the output, which is why it's related to the coefficient m times X plus that intercept. So in this case, you can think of M and B as sort of parameters of the model right, and the model is the model of a straight line. So that's exactly what machine learning does without any explicit programming. You essentially hand data to an algorithm, and the algorithm automatically learns those parameters and coefficients of some sort of model in order to get again. The best performing model and what this essentially is, is an optimization problem and normally what? That how you go about that as you define what's called a cost function. And this cost function is a way to determine you know how well your model is performing compared to actual data that you are a have. So let's say you have some data. Think of it like an Excel spreadsheet, right? Like a table. Each row is, let's say it's It's a database of movies in an Excel spreadsheet, and every row is data about a specific movie, and every column is a feature of that movie. So in these kinds of things, you can think of the columns as either features or attributes of the data, or so our dimensions of the data. But essentially, they describe different things of each of the records of the date of which of the rose so far, given movie, you might have the duration as one feature. Then you might have the category like a drama or comedy on. Then you might have the lead actor or actress as another feature, and so and so. But let's say one of the columns or features is, you know, the IMDb movie ring, and it turns out I'm giving you a little hint here. But this is kind of involving our class project at the end of this course to. But if you have time to be moving rating as one of the columns, that means maybe you can use the other data or features to predict for a given set of values of those features. What the average I'm TV movie rating might be for a given movie. So in that case, the cost function would allow you to know whether or not with the data that you have that already has. I am DB Average costs are person movie ratings. Whether or not your model is close to predicting that rating accurately and by how much. And so the idea is to minimize the cost function, which means that you want your model to be able to predict it almost exactly if possible. And so it's all about minimization, and that's what makes it an optimization problem. You're minimizing the cost function, and one of the most common cost functions, for example, is mean squared error. And one of the techniques are algorithms that used to perform this sort of minimization, or learning, if you will, is called grading descent ingredient descent. Essentially, what happens is you take the algorithm looks for different combinations of those parameters . Remember again in machine learning. We're trying to learn the optimal parameters and coefficients for the model. So think of it like it's tweaking those parameters and coefficients such that it keeps trying to get closer and closer to the final answer, which is the best performing model. So the best values of the parameters and coefficients, and so the reason it's called graining percent is grainy in is like a slope. So if you're walking up a mountain, you know the mountain has a slope, or some people would call the Iranian and the steeper the slope, the steeper the Grady in. So in this case, if you as I mentioned, it's a minimization problem. So you're trying to find the MINIMA in another way to think of this is if you are walking through a valley with some hills and some you know highs and lows, right, valleys and dips, but then also hills and peaks. You're trying to find the deepest dip or deepest valley. And at the very bottom of that is the answer. It's the solution. It is the best performing model, which means you found the best parameters and coefficients for it. And so grating dissent is exactly doing that sort of thing of it walking around this valley , walking up, trying to find the steepest slope downwards and then finally arriving at the lowest point in that valley. And that's the answer. So I know this is a lot, but it is very interesting. And what's really fascinating again is that this all happens automatically by the machine learning algorithm without any extra explicit coding. And all you gotta do, basically, is have some data that handed to the algorithm. And there you go. It's actually a lot more complicated than that, but that definitely gives you a pretty good sense, hopefully of what's involved there. So in the next a couple of videos we're gonna talk about different types of artificial intelligence and machine learning, and also algorithms see that 5. Machine Learning Types and Algorithms: all right, so we've given definitions for artificial intelligence and machine learning, and now we're ready to talk about specific types of both and also some of the algorithms there involved. So let's dive right in, and this video will be talking specifically about machine learning. So in machine learning, you typically have four major primary types of machine learning type tasks, or algorithms. Once called supervised, another's unsupervised, then you have semi supervised and then you have reinforcement learning. So let's talk about each of these in order. So in supervised. Sometimes it's called labeled learning. And what that means is, if you remember in the last video, I talked a bit about the sort of I. M. D B movie data set in a spreadsheet form, and I talked about how you might have an output type variable that you're trying to predict , and so you build a model to try and predict that that variable in this case, or feature more features when you use it to make the prediction. Eso normally the outputs called the output or the target or the response. So, in this case, in terms of supervised learning, those air the labels so you have data that has labels mean it has known values of the output, and so you can train your model toe, learn how to predict that label or output or response, given a set of features so that when you get new, unseen data that doesn't have the label, you can pass it to your model, and then the model is able to make a prediction or apply a label to it. Essentially now, in that case, for supervised learning, there's two different kinds. There's regression on. Then there's classification and regression. The goal there is to predict some sort of numeric value or number. So in this case, I am D B movie rating. Average movie ring is a good example for regression. You're trying to predict natural number out of 10 let's say, and then the next type is classifications. And that's where you're trying to apply a class or category to some data in a predicted manner as well. So a good example of that is take your email client, for instance. Ah, why you may use Gina. So Gmail has a very sophisticated spam filter, and what that does is when e mails air coming in. It runs through basically a classification type algorithm and determines whether or not that email is spam or not Spam. But you can have more than just two classes or categories that you might assign to some data. It could be spam, not spam and unknown. If you're not sure enough to make a prediction or another example is in the medical field. Machine learning classification is becoming more and more prominent all the time in the medical industry, and a good example of that is disease diagnostics, especially cancer diagnoses. So, you know, doing a bunch of tests, taking the data and then saying you either have cancer or you don't have cancer, For example, Uh, so that's another example of classes occasion. Then there's the unsupervised learning, and in this case, the main differences. It's unlabeled, so it doesn't have you know that response variable or target. So you just have a bunch of features. You just have a bunch of features of data and a whole bunch of records. And so the idea is, you apply different techniques like clustering or anomaly detection, things like that to usually create groups from the data groups or segments however you want to think of it, and then you can use those groups to then make predictions. But in this case you're not predicting and out like a label or output like a class or a numeric value, you're just predicting what group a certain bunch of data falls into. So in the case of clustering, imagine you take a bunch of data. You decide you want five groups, you run the model it creates, You know it finds five different groups from the data that they're sort of isolated from one another, and then new data that comes in You can basically automatically assigned to one of those five groups using your machine learning. And that's really interesting stuff. Then they're semi supervised learning, which is a combination of supervised and unsupervised learning. And then finally, there's reinforcement learning as well. And reinforcement. Learning is really interesting in that it's kind of like playing a game where you have what's known as an agent and serve a yeah abstract environment concept. And basically it's trying to achieve certain rewards. And it's so it's all about building up and achieving rewards. Eso that's kind of what makes it sort of like gamey, but it's really interesting. It's becoming very big and having many amazing uses these days as well. But before we end this conversation about machine learning types, it's also worth noting a few other examples. One is recommend er systems, which most you're probably familiar with if you visit Netflix or Amazon ever on bears some of the best known out there in production on and then more specifically, in there you have content based and collaborative based filtering techniques for recommendation systems. You also have a tight known as ensemble methods, which includes bagging and boosting. And one of the algorithms you may hear of in terms of bagging is called random forests. Some of you may have heard of that, but these are interesting techniques in which, instead of having just one model, you create a bunch of different models, and then you use all those models to create the best possible solution or prediction. And they do that by, you know, different sort techniques. In one case, you could think of it is all these models you create sort of vote. They like having election kind of thing, and they vote, and then the voting chooses the best performing predictive model, Let's say, but another way to think of it is you could have all these miles sort of average together, and the average of all of them is the best performing, not serving. But anyway, Osama methods are really interesting and typically could get you some pretty good results if done right. And then the final one is transfer, learning that one's being discussed more and more these days. That one's interesting in that you use models that are good at solving one problem, and then you transfer that solution or knowledge. Or what? However, you want to think of it to solving another problem. So that's really saying as well. And then finally, just to talk a little bit about specific algorithms, um, you know, in terms of supervised learning, you'll hear things like simple and multiple linear regression. You'll hear things like decision trees and so on and so forth. For classification, you'll hear of things like support vector machines, logistic regression, also, neural networks. But we'll talk a bit more about those in terms of artificial intelligence on then, unsupervised learning. You have things like hierarchical clustering and K means and so on and so forth. Uh, so yeah, that's Ah, pretty solid overview. I think of the different machine learning types. In the next video. We'll talk more specifically about artificial intelligence types. 6. AI Types and Algorithms: all right. So in the last video, we talked about machine learning types and different algorithms that you might find in this video. We're gonna focus specifically on artificial intelligence types and also some algorithms that they're coming in artificial intelligence. So in terms of types of artificial intelligence, you may hear different things like strong versus weak ai deeper XYZ narrow ai hard versus soft ai full A I applied ai and generally I or artificial general intelligence abbreviated as a g. I. Now these are all different kinds of art ways of describing artificial intelligence. And if you remember back at the definition when we went over that we basically talked about , you know, artificial intelligence is all about, you know, kind of learning, observing the environment, learning, you know, taking in information and then using that knowledge gained to perform tasks or make decisions, or essentially and ultimately do things that humans can do and display the same kind of logic and reasoning that humans can. Now we're a long way off from full blown, you know, doing everything a human can do, which would be considered artificial general intelligence. As I mentioned most of the AI that you see nowadays is weak AI. And in this case it's all about sort of solving a specific task and, you know, simulating a mind or the ability to kind of learned and then apply that knowledge. Learn to solve a specific task. Artificial general intelligence or some of those other terms, like strong general intelligence, are strong. Artificial intelligence are really Maura about being able to solve any task or multiple tasks at the same time, and being completely indistinguishable from a human that is a machine or computing device. Basically, when a machine or computing devices able to simulate or do things that humans can do, where humans can't tell that it's a machine doing it and they're able to adapt to all sorts of different problems or tasks, sort of on the fly and continue to learn as it's doing these things, that's kind of you know, the future artificial intelligence and, you know, right, get there. But we're not quite there yet. It turns out that one of the biggest challenges with artificial intelligence if you remember we talked about unsupervised learning and unsupervised learning is where you have data, but you don't have the labels, uh, or the output or the response that you're trying to predict. Let's say, if it's a prediction problem now, it turns out that having labeled data can be very difficult to get sometimes especially in the quantities needed to do certain tasks, particularly in the case of a lot of artificial intelligence tests. So, you know, sometimes you could get a lot of labeled data, and sometimes you can't. And when you can't, uh, you know you need to do unsupervised learning. But it turns out that unsupervised learning is hard to do in the context of what we consider to be artificial intelligence. Now there are some ways to do it. But you know, the that's gonna be a core part of the future of AI is getting better at doing artificial intelligence with unsupervised learning methods. And there's this great M I. T Technology Review article that talks about you know, infants and how unsupervised type learning is analogous to infants. So, for example, you know, events learned that objects without being supported by, say, a table or desk will follow to the ground, and we know that's because of gravity. But infants learn that without being told that by their parents. They just learned it from observing their environment. So that's kind of like an unsupervised learning where there's no label in this case being told by your parents that that's true. Another thing that infants learn is that when they walk out of a room, they kind of learn that all those objects and everything that were in that room before they left are probably still there even after they leave the room and they'll expect to find them when they go back into the room again. This is an example of that unsupervised learning that we're talking about. And so the different types of artificial intelligence are all about, you know, to what degree can you solve? You know, just one or multiple problems. To what degree can you learn from, ah labeled data? You know, toe, What degree can you perform? You're all the same kind of logic and reasoning and task carrying out as humans and also not be able to tell that it's a machine doing that. That pretty much encompasses artificial intelligence. So with artificial intelligence, the most common types of algorithms are neural networks and deep learning and Deep Learning is a specific category of neural networks, which we'll talk a little bit more about in a second. Neural networks is this abstract sort of mathematical modeling technique that's modeled on what people believe is the model of the way the brain works and the neural networks in the brain. So the idea there is that the brain has all these neurons. In fact, billions of them. There's tons of them, and your aunts have things like axe odds and danger rights. And they communicate with each other across these gaps called synapses, And they basically fire information off between each other back and forth through an entire network. And you know, that process is what leads the brain to be able to do things like think and reason and perform logic and all that sort of stuff. So neural networks are a our official Neural networks specifically are sort of an artificial or abstraction of that concept, mathematically an algorithmic lee. And so that's a very common algorithm on one of the most common algorithms used for artificial intelligence task. And there's many different kinds of neural networks as well, and as neural networks get more and more complex. You start to get into this realm of what's known as deep learning, which are neural networks, essentially that have multiple, you know, hidden layers, and we're not gonna get into. All the details are not here in this class. If you want to learn more about neural networks and deep learning, there's playing another great resource to do so, including my repository, which complaint in the right direction a swell. But in the case of deep learning, what's really fascinating is not only does it have capabilities around unsupervised learning, which we already talked about, is a bit of a challenge, but also deep learning is ableto learn features automatically. And what do I mean by that? So in machine learning typically, and we're gonna talk about the process more in the next video within machine learning. Typically, data scientist, you know, has a set of data and chooses what features of the data that they want to use in their models. And so it could be kind of manual, actually, like picking what features or maybe creating new features from existing features or new metrics so and so forth. Um, but in the case of deep learning. You know, there are ways in which the algorithm can learn which features to use automatically without , you know, sort of that manual step or data scientists to make the selection. Well, what does that mean Exactly? What it means is that the algorithm actually learns how to learn. And that's kind of D. But it's truly amazing when you're talking about an algorithm, learn not only learning like in the case of machine learning as we talked about we creating dissent and learning those optimal parameters and coefficients. In this case, we're actually talking about learning what features and what parameters and coefficients and all that stuff give you the best performing model, which essentially is saying that the algorithm learns toe learn and that starts getting them or that artificial general intelligence type stuff. And you know what you probably think of like Terminator or something, and again, we're not there yet, but just to give you an idea of kind of how that all comes together. So we've gone over a I types and some of the algorithms involved in the next video. We're gonna look more closely at actual processes involved with artificial intelligence and machine learning. 7. The AI and ML Process and Tradeoffs: all right. So we've talked about different types of artificial intelligence, different times in machine learning and some of the common algorithms involved with each. And now it's time to talk about the process involved with both artificial intelligence and machine learning tasks. And you know, this isn't set in stone process, and it's very cyclical and interpretive, but this gives you a general idea. So it starts with having a goal in mind. And this is why we talked about business goals and customer goals at the beginning of this class and the why. And this is really where that begins in the process and so doing that allows you. Then ask the right questions. So what do I mean by ask the right questions? Well, I already talked about Amazon, the recommendation system. So let's take that. For example, suppose you were somebody at Amazon and you said, Hey, you know, I wonder if there's a way that we can increase average customer order, size and thus increase sales and revenue. Well, what do we have the data to do that, and if so, what data do we have? That's the kind of question that's the beginning of it. But the question is, can we increase average order, size per customer with the day that we have? Uh, and the goal there is to increase revenues, right, increased sales increase revenue, increased profits. And so that's the question. And so the next step is in orderto see if you can answer that question or solve that problem, you want to obtain some data now, sometimes you have the date already, because you might be a lot. You know, you have computer logs, you have CR EMS. You have e commerce sites, you have social media data. You know, you might have data from all kinds of different places, but sometimes you need to also get data that you don't have already. So you may have to figure out how to get the data to solve your problem. And that's the next. Then once you have the data, you have to parse that data. And parse is just a fancy word for kind of read that data from whatever form it's in into the computing language or environment, or whatever form that you need in order to move forward with it, which the next step is processed. That data and process could be filling missing values. It could be cleaning the data. It could be transforming the data. Creating new data or metrics are calculated data from raw data and so on and so forth. And then the next step is performing exploratory data analysis or E d. A. And what that is is basically looking at your data and doing a bit of statistical analysis and data visualization. And this helps you understand your data a bit better before starting to create models. And the reason you want to do that is you want to understand what kind of features you have . The different distributions of those features, um, and so on and so forth and then from there. Typically, you would choose what algorithm or modeling you want, use and choose a performance metric. And the performance metric is how you know how well your model is doing for the task that you're trying to accomplish. At that point, you create a model, train the model, and then you test them all to see if it's performing well, and from there there's usually always room to improve. So you go through sort of this generation process where you're kind of in arraigning and improving your model. You're tuning the model and so on and so forth in with the goal, you know, basically getting a better performing up. And then finally, you either deliver the results. In the case of a product company, you know, maybe you're using artificial intelligence and machine learning to be a part of a featuring your product, so you deliver that feature. But in the case of you know, maybe it's internal inside a company or business, and you're using these techniques there to help answer questions and solve problems and make business decisions using data and so on. Then it's all about how do you communicate the results or present the results on that could be both verbal and written and using data visualizations. And one of the key things to keep in mind is that there's often many trade offs and considerations that you need to keep in mind throughout that entire process because, you know, it's not always super easy to get the best performing model one of the first ones. It's called over fitting, and sometimes you'll hear of it referred to as biased versus Barry. It's and What that means is essentially, you know, let's say you have all this data, and maybe the data contains a lot of outliers or unusual values and noise. You know, if you're trying to fit a model to that data, maybe if you over fit the model, meaning you capture all that noise and out liars and everything in your model, you, then your mom may not do very well. And by do I mean perform predictions or classifications Very well, if that's what you're trying to do on new and unseen data, because you know the model takes all this noise and everything else in your account, so it doesn't generalize well to new data. Now, if you take a model and you completely under Finn, it meaning you don't capture enough of the variance in the data and that sort of thing, then the model doesn't really fit the data very well. It's and its. That's what's referred to as being too biased on In the former case. It's having high variance where if it's all the noise mailed out letters, But in either case, you know, over fitting is a really important and probably one of the most important considerations you need to understand and make. And it's all about the trade off between bias and variance and trying to find the best performing model. That Sir falls as optimally in the middle is possible. Another way that you can under for the model is sometimes you don't have enough data t create a really good performing model. Or maybe you don't have the right data. You don't have the right features or, you know, and so on and so forth and so under fitting can occur as well. So you might need to consider that and figure out waste. Either get more data our data or get the right kind of data. Um, another tea trade off in consideration is model performance versus interpret ability. And what I mean by that is, you know you're always going for the most performance model, meaning the one that can achieve the goal the best if the goal is prediction. It's the most accurate predictive model, let's say, but interpret ability can be a very important consideration, and turning ability means having the ability to describe how the model works to other people, whether even in words or visually for example, um, like take, for example, neural networks and deep learning, which we talked about a bit in the last video. Those air very what's called Black box algorithms, meaning that you know it's very hard to understand or describe what they're actually doing or talk about any of the in your workings. And so it sort of got this sort of magic characteristic to it, and people don't understand it very well. Sometimes that doesn't work out well for companies, particularly companies like financial companies, where, you know, let's say a creditor is making credit decisions based on a predictive model, whether or not to extend credit to, for example, Um, and they reject you. And then, you know, if you were to sue them, you know, and they can't really describe in court what the model is actually doing, how it works, You know, that could be a big problem. So a lot of times interpret ability is key, and so you might choose to give up a little bit that performance. But use, um, or interminable and understandable model and decision trees, for example, and regression like symbol of multiple linear regression are great examples of highly interpret herbal algorithms. Another consideration is model complexity versus simplicity. And sometimes you hear the term parsimony or simplicity, and what that means is model complexity. Essentially, is the idea that, as you add more features to the model so you know, let's say you have 20 features verses only using five features. That's a more complex model. But certain miles are also more complex and other model types. So neural networks is a more complex mile than simple in your regression, for example, and generally the rule of thumb is that you want the most simple model or algorithm that gets you the performance that you need and nothing more. You only want to really add complexity. If you really needed to get that extra performance, and they're making reasons for this one is again. You know, simplicity is just easier to work with. It's easier to understand it's easier to train the models. But another key component worth considering is costs, so cost in terms of computational costs, you know, the more complex them all the more computing power you need, the more resource is you need the more time it taste, so the time cost more time it takes to train your models and so on and so forth. So it's that balance between simplicity and the performance you really need Teoh keep in mind which comes through in terms of Yeah, again, personally, Another thing you wanna consider his curse of dimensionality. And this is a concept that, you know, imagine you have all these features of your day and like we talked about the imdb Dana set , you know, you have duration, you have lead actor or actress. You have category and different features. Well, each of these features can take on, you know, different values, and some features can take on many, many values so less. But as you increase the number of features, the so called feature space and it grows dramatically and essentially, what happens is if you don't increase your training data in in step with the number of features you keep, adding your data becomes somewhat sparks and doesn't really fill in all the different. You know, aspects of that feature space, which means that it's hard to make your model perform well, potentially on you. Carry out the task that you want to do without adding a lot more training data. And so the curses dimensionality is all about that. As you add more dimensions or features on, you make that feature space more complex. It turns out that you need a lot more training data pretty quickly in order to have the same predictive power. Otherwise, without the extra data and a lot more extra data to train your mouth, you lose predictive power. And finally, there's this idea of the global versus local minima. You might remember back in the beginning, this course we talked a bit about Grady and dissent, and I gave the analogy of Imagine, you're walking through a valley with hills and peaks and dips and troughs and all that sort of thing and you're looking for the lowest dip or valley the lowest point. And that's the solution to the machine learning task. Let's say well, sometimes imagine you're walking around and you find a little dip in the ground and you walk down to the bottom of that thing and you get there and you say, Oh, here I am. I found the solution. I'm done. Yea, well, it turns out what you you may have found is a local minima in that Yes, it is a dip, and maybe it's a pretty loaded, but it's not the lowest it in all of the valley that you're in and the lowest it is actually the best solution. That's the global minimum. That's the one you always want to try and find, because that is absolutely the best performing mile possible. Given all the parameters that you're dealing with. And so you know, you always have to keep in mind whether you're finding the optimal solution, which is the global minima or maybe actually only got stuck in a local. So with that, we've covered a lot here and now it's time to actually talk about real world applications of artificial intelligence and machine learning. 8. Recommender System Applications: all right, we have gone through a lot, and now it's time to start looking at actual real world applications off different types of artificial intelligence and machine learning categories. This particular video is going to focus on recommend er systems. And again, let's just take a brief overview of How did we get to where we're at now? We looked at business schools. We looked at customer goals. We looked at the different definitions and types of artificial intelligence and machine learning, a swell a Cem algorithms. And then we looked at the process involved with each and now we're gonna look it applications. And this is sort of that culmination where companies have created really amazing products and or services using everything we've discussed to solve real world goals. And so with recommend, er, systems. Let's start with that now. The two most popular and well known examples of recommend er systems out there in the world are obviously Netflix and Amazon. We already talked a bit about Amazon's recommend er system and you know, it's that whole thing of customers who bought this item also bought that I have. They also have customers who viewed this item also view something else or what other items to customers by After viewing this item, they have a section that says Recommendations for you in and then some category like, let's say, kindle books on and so on and so forth. So they use recommendations all throughout the site. And as we said, you know, some studies have shown that they've increased sales revenue by up to 30% using this type of technique in order to increase average customer order size on also just general user engagement. But Netflix is another great example. And if you think about remember, we talked about that concept of being sticky well, Netflix. Imagine if people went on the Netflix and you know they didn't really use it very often and kind of just used it every once in a while. And that was sort of it and then forgot they even had, or the existence. It wouldn't be very sticky, and it certainly wouldn't help Netflix make any money. Actually, since creating that recommendation, engine wound up being one of the most amazing things that they did because that allowed them to increase engagement and customer engagement and retention significantly, and therefore revenues. And they do that because they're constantly telling you new things to watch. So they're they're taking, like, you know, a lot of people are kind of indecisive, and you don't want to scroll through everything or search through everything. They're sort of taking all that complexity away from you and just saying, Look, because you watch this, maybe you want to watch that or here some topics for you, for example. And they also make recommendations by category, uh, or things like trending now or Netflix originals. Or here Cem recommendations for TV documentaries. And so when you see that, it turns out that people are able to just kind of make decisions much easier, a lot less friction, and they're more engaged, and the experience is quite personalized as well. And so as a result, you know, this is a huge, huge advantage and benefit to them by creating this recommendation system. In fact, there's ah famous Netflix Prize that happened a while back where Netflix had a huge competition and they were going to award the winners a $1,000,000 for coming up. You know, the best possible performing recommendation engine above and beyond what they are a had and you know, this competition went on for years, actually, and eventually a team won. And the the funny thing is, his Netflix wound up not actually using their solution, even though they did win and they paid them the prize money. But, you know, remember we talked about the tradeoffs, and sometimes you do need to consider performance versus things like complexity and everything else. And it turned out that even though this model was the best performing model, it didn't make a lot of sense to actually use in production for various reasons. And so as a result, it's an interesting fact that Netflix didn't wind up using that particular algorithm in production. Uh, so let's now look at some other examples of recommendation systems out there. Okay? Now you remember I created this repository on get hub called data science. AI machine learning resource is on, and basically, if you go visit that you'll see there's all kinds of stuff, their algorithms and tasked artificial intelligence, things about big data and analytics i o t data science machine learning data, visualization, and so on and so forth. So there's quite a bit there a lot of the stuff that we're gonna talk about in these real world applications sections you can also find if you go to the artificial intelligence page and you scroll down, you kind of see a lot of stuff we've talked about in this class, Um, and actually a lot more detail. So here, for a I common algorithms, you could see I've even included a ton of different specific neural networks and deep learning algorithms. But if you go down even further, you see, I have a section of real world application and vendors by category, and we're talking right now about recommend er systems and so long, so feel free to visit that for more information or a recap now, in terms of recommend er systems, some other really interesting recommend their systems out. There are companies like Wealthfront, which you see here and also a company called Betterment, and what they're doing is they're doing something called robo Advisors. Robo advisors essentially find ways to automatically recommend investments or portfolio changes or re balancing or allocations and that sort of thing. Basically personal wealth management in portfolio management automatically, and it's being driven largely by these kinds of algorithms, so that's really interesting. And, you know, in some cases, you know, you get recommendations and then you make the changes. And with some of these Robert Robo Investor cos you just kind of let them make all your portfolio changes automatically, but pretty interesting stuff. Another one some of you guys might be familiar with is Spotify eyes daily mix, which is pretty cool. And basically what they do is they take all the data from the kind of music you listen to in Spotify in your account. And every day they create a set of new mixes that are sort of different. But each one is largely based on the kinds of music you listen to. And so this is sort of like a recommendation engine in the sense that it's recommending, you know what songs to listen to and they're each bundle up in individual mixes. Pretty impressive stuff. And by the way, if you haven't tried it out, I highly recommend it because mixes air always pretty cool. Then another one you might be familiar with is Facebook's news feed, which is very personalized and, you know, it's it's sort of like a recommendation system as well, in the sense that you know, using a lot of techniques that we're talking about here to personalize your news feed, to customize it to you and essentially recommend the kinds of news you should read. And, you know, as they say here, their goal is toe give you news that matter most to you every time. Um and what's interesting? They you know, they have your top stories, the stories you care about page here and they talk about, you know, it's an ever changing collection of all sorts of things on den. On this next page, they talk about the three main ranking factors. And here you get a little bit of a glimpse into, you know, kind of their algorithm, if you will, in terms of what kind of data they're using for these recommendations. And so they're basically saying they look at you know who posted something, the type of the content, any interactions with the post. It's all very interesting. Examples recommend their systems in the real world 9. Prediction and Classification Applications: so we just took a look. It's really a world applications and companies using recommend er systems. Now let's take a look at some involving prediction and classification, which you'll remember eyes often associated with supervised learning in machine learning when we talked about the different machine learning types, so let's dive right in, all right. So there's a lot of amazing things going on with artificial intelligence and machine learning in terms of prediction and classification, as we talked about earlier. Now it's worth mentioning two of the big powerhouses out there. 1st 1 is deemed line, which is a company that was acquired by cool. And they're doing artificial intelligence stuff that's applied to all sorts of different industries, like energy and health and so on. But another one that I'm sure many of you are very familiar with as well is IBM's Watson and you know, there that would became very famous with the Jeopardy Game, where Watson was ableto win a jeopardy, uh, a little while back. And but it turns out that Watson is deployed in production for many, many different applications, and they even note there on the Web page, you know, across 20 different industries, and health care is one of the big ones. Another interesting company is called free No, and they're using artificial intelligence techniques in order to do disease screenings and proactively treat things like cancer and other things. Are diseases pretty interesting there and then a company called Cure Metrics. So some other interesting examples of classification is Gmail. As you could see in the image, they automatically are able to classify or categorize incoming emails in terms of categories like social promotions, things like that. But another interesting feature that they've got is called Smart Reply. And this is getting into the realm of, like, natural language and that sort of thing, which we'll talk about more later. But basically where you know people are writing emails and Google is coming up with some potential short replies that are relevant or pretty decent answers or responses to whatever came in. Another category that is interesting for predictions is, you know, stock market industry and investments and algorithmic trading and that sort of stuff. So some of those companies doing that, our companies, as you can see devout, I believe it's pronounced for car boot, Centeon, genetic and New Marike, which is actually a hedge fund that's sort of driven by data scientists and artificial intelligence on and also another company called Cube LEM that's doing similar things as well. Now some other interesting applications of prediction and classifications are suicide prevention, for example. And there's a lot of companies out there trying to leverage artificial intelligence techniques in order to help with this. One of them is this company here Koji Ito, perhaps, and what they're doing is they're trying to get a picture of emotional health through listening to voices and the way people speak. So it's pretty interesting. And then, you know, obviously, you know, sort of mapping human speech to mental health can then help with things like suicide prevention and so forth. Another interesting application is shown here by a company called Underwrite dot a. I on G. They're basically creating credit myth risk models using artificial intelligence and sort of advanced techniques, which is pretty interesting that lenders can use. Another interesting company is called the CART Labs on, and they're doing really interesting things with geo visual search, Andi, visual detection amongst different scenes and things like that, and also being able to do forecasting using their geospatial platform. And then finally, there's Stanford has a lab called the Sustainability and Art Official Intelligence Lab, and they're doing some pretty interesting stuff with crop yield predictions based on all kinds of different artificial intelligence techniques. 10. Computer Vision and Recognition Applications: all right, so now we've looked at recommend her systems and various applications of prediction and classifications. Now let's talk a bit about computer vision and recognition. Computer vision typically involves recognition of things like objects on scenes and other things like that, but also motion analysis and seeing reconstruction and image restoration. For example, um, and computer vision is really important in a bunch of different AI applications as well discuss, but also in terms of robotics on that sort of thing. Another thing, as I mentioned that will go over now is recognition. And artificial intelligence is being used across the board for lots of different interesting recognition tasks. So recognition can include speech recognition, image recognition, video recognition, sound recognition, music recognition on so worth. So let's look at some of the examples in the real world on companies doing some of these really amazing techniques. Okay, so let's talk about some real world companies in real world applications using artificial intelligence, particularly in computer vision and recognition. Ah, one such company is Jen Attackers. You could see here on there's this technology called BRS labs a eyesight which is using a I to do you know, video surveillance on and also sort of mothering for abnormal behavior and so on and so forth. Another A couple of companies are digital signal and shield ai, and you know they're doing things like using computer vision for crowd surveillance, facial recognition and so on. She'll die if you kind of see by this pretty cool video they have on their website, is doing a lot of really amazing things with AI and autonomous vehicles in terms of, you know, homeland Security, defense, that sort of thing and the same sort of thing about monitoring situations, doing visual surveillance, insulin and so forth. Another really amazing example is Amazon. Go for some of you that may have heard this are a So this is this whole new concept of having stores that basically have nobody working in them more or less. Um, and there's no lies. There's no check out. You just kind of go in grab which need and leak. And so Amazon's actually piloting this technology right now as we speak. Um, and you know, at some point when they got it all worked out, it could be released too many different markets and cities and areas and what's really interesting here is they're using lots and lots of different artificial intelligence and machine learning techniques. One of them is the computer vision, but there's other techniques as well. They're being used here, so another really interesting area is recognition. And as I mentioned earlier, you know you have video recognition, image, text, speech and so on and so forth. So let's talk about some companies. It's real world examples of that, and obviously one of the ones you guys maybe immediately familiar with this Facebook and how they do facial an image recognition so that you could tag your friends almost automatically in post that you make on Facebook and so forth. Another one that may have you are probably familiar with. This is an which is doing some pretty incredible stuff for doing sound and audio recognition in order to identify different songs and music to really interesting cos Aziz well are delectable and vino. So here's Delectables website and Savino there. These are companies that allow you to take pictures of wine bottle labels. Let's say when you're at the grocery store and then give you information about that particular bottle and year and vineyard for that specific wine and including things like ratings as well. And so it's really nice if you're trying to make a decision on which wanted to buy the store. Both of these after pretty good at helping you up. Another couple of pre interesting applications of artificial intelligence and recognition are Google has an A P. I called the Cloud Speech a p I, and basically this is an A P I that lets you speak to it and provides actual recorded speech. And then it's able to convert that speech to text. So that's pretty cool. Speech recognition and one of the cool things that Amazon's doing, which is sort of the reverse of this comes via their Amazon Polly a. P I. And basically it's sort of the opposite where they take text and they turn it into a lifelike speech. Ah, and as you can see here, they point out that they use deep learning for that. And so a good example of that is, you know, like Amazon. Alexa, for example, talks to you and things like that and tries to do sort of that lifelike speech. Um and so anyway, pretty interesting applications there, all right, so we talked a bit about text and speech recognition. But another couple of examples worth noting as well are, and most of you are probably familiar with this. Your bank probably allows you to deposit checks using their mobile app, and you do so by taking pictures of the front and back of the checks. And then the system is able to extract whatever relevant relevant information it needs from those images. You know, things like the amount of the deposit. Another example is extracting address information or postal addresses from envelopes or other sources, and basically using recognition to not only recognize and extract those addresses, but even individual components of the address like the city, the zip code and so on. And then there's a few other examples will go through the finish out this section. So there's a company called Clarify that's doing some pretty amazing things with image and video recognition, and they offer it through an A p I, uh, another company who you're all more than familiar with his Google. That's got a platform called Cloud Video Intelligence on. As you could see at the time of this recording, it's in beta mode and basically allows you to take your media content, video content and essentially search and discover information from that video content. And they do that through metadata extraction, so that's pretty amazing. Another thing is image recognition as we talked about and analytics from images, and Google offers another AP. I called their cloud vision E P I. That does this, and this really focuses heavily on object and facial recognition within images. So, as you can see in the pictures there, sort of being able to taken image and recognize automatically that there is a sailboat in it, or the sun or whatever, so pretty cool and then Amazon also has an A p I. That's pretty amazing. Called recognition uses deep learning, and it basically allows you to also do image analytics. It's all pretty cool stuff, and that concludes our section on recognition. 11. Clustering and Anomaly Detection Applications: All right, So now let's talk a little bit about clustering. An anomaly detection, and you'll remember. I already mentioned those two fields when I was talking about unsupervised learning earlier in the class. So let's take a look at some of the examples of how clustering an anomaly detection are being used in the real world. So we've talked about clustering anomaly detection, and there's a lot of different applications, as I mentioned. And a lot of them are found in things like manufacturing, for example, trying to find defects which could be definitely concerned out letters or anomalies. But also data security or network traffic. Unusual network traffic. Ah, cybersecurity, even personal security, like things like personal screen or security screenings at airports or stadiums and other venues, but also law enforcement and application performance. Let's say you have software, some cloud based solution, and you're kind of monitoring it and trying to look for any sort of weird performance issues or no unusual source slowdowns of the application and that sort of thing, and then also fraud detection, like credit card fraud, detection and so forth. So here's a company called Silence, and they're using artificial intelligence to prevent cyber attacks. And they are doing that through prediction and also some of that anomaly detection that we're talking about and then giving ways to help with responding to those types of cyberattacks. Another company doing similar things is dark trace, and they're also looking for those strange patterns or anomalies and so on that are associated with cyber crimes and cybersecurity. 12. Natural Language (NLP, NLG, NLU) Applications: All right, let's move on to a really interesting area of artificial intelligence, knowing his natural language on some of the sub fields of that are actual language processing or NLP Natural Language Generation or an LG and natural Language Understanding? Or in L. U. And let's go over different examples of what of companies using these incredibly interesting techniques. But also just to understand that you know, we as humans, we use natural language all the time. We were able to speak languages and say words and sentences and, you know, talk about sort of concepts and abstractions and create meaning through words and all that sort of thing. And other humans can hear those things and automatically sort of absorbed them and understand them. You know, while the brain processes all sort of that sensory import, input of sound or speech. In the case of natural language, well, artificial intelligence, you know, also is trying to do things, is able to do things with natural language, as I just mentioned in those three sort of sub fields. And so it's all about taking, you know, written or spoken speech or text, and converting that into sort of that same kind of processing and understanding or ability to create meaning or words from natural language, but by a machine. And so that's what started this fields all about. And so let's look at a couple of real world examples in this space, all right, so we've talked a bit about natural language, and I mentioned the fields of natural language processing generation and understanding. And we'll talk more about that as we go through this section. Now, another part of natural language and an application is known as personal assistance, and you all are probably very familiar with the three that I'm about to point out. The most popular three are Amazon's Alexa, Google's assistant or the Google assistant and Apple's Siri. And what are these used for? And why are they useful and going back to the customer goals of the beginning? What jobs do they get done for people? Well, you know, for all three of these, you can do Internet searches and answer questions. You could set reminders and calendar events, integrate with your calendar, make appointments. You could receive various kinds of news and sports and updates of the stock market. You could create lists like to do lists and shopping lists. You can, in some cases, order items online. In the case of Alexa, you can use other services like maybe ordering an uber. You could play music, play games, and some have smart home integration as well. So all pretty amazing and all able to help with many day to day tasks and goals and jobs that people want to get done. So let's talk a little bit about natural language processing and what that means. Essentially and natural language processing, by the way, is kind of a very in depth aspect of artificial intelligence. And so there's a lot you can do with it more than we'll discuss here. But some of the main things you could do our extract information from text or natural language things about like as Google points, I hear people, places, events and all that. But you can also extract information about things like adjectives and now owns and adverbs and so on from text, which is pretty incredible and kind of see what the content of the text is in terms of all those things and what the density the varying amounts are or what kind of words occur in what frequencies or what combinations together. Another thing you could do is try and extract what's known as sentiment. And sometimes you'll hear about something called sentiment analysis, which essentially tries to take some sort of natural language via tax, let's say, or speech and translate that into a measure of sentiment. Meaning, How positive is it or how negative is it, or how neutral is it? So that's all very interesting, and Google has this AP. I called their cloud natural language a p I, which allows you to do many of those things pretty cool. Another company out there is called Text CEO, and they're doing something really unique where they're using artificial intelligence, to analyze lots of lots of data around job descriptions and job listings and so forth and then using that data. And that that analysis to be able to create automatically highly effective job listings on job descriptions. So that's pretty cool. And these are just a couple of really interesting examples of natural language processing. So now let's talk a little bit about natural language generation, and that's basically using artificial intelligence to automatically generate language on that language. might come in the form of reports or news or summer organizations of some sort, like summarizing a document, telling a story. Maybe some recaps of sports and so on and so forth, and one of the most cutting edge and sort of big players in this space. Out there is narrative science based in Chicago, and so they're definitely leading the way in terms of natural language generation. All right, let's finish out this section by talking about natural language understanding. And some of that includes something you may have heard of called chatbots. But basically there's a lot of companies doing some pretty cool things. A natural language understanding is sort of like, uh, bit of what's known as a hard ai problem in that it's pretty challenging involves things like speech recognition on natural language processing. But it's sort of more than that's not just processing the natural language, but it's also understanding it and really developing a sort of comprehension of it and then using that to do pretty interesting things like, you know, smart messaging app, sor conversational interfaces. Ah, and so on and so forth. And so that's what this final section is really about. And then once you have that kind of an interface in the case of this cock sort of conversational, natural language, understanding driven interface like Amazon's Alexa, for example, you can then carry out tasks with that, whether it's, you know, have it help you do some shopping or to carry out day to day tasks or track certain things on so on and so forth. So some of the companies out there doing that one that's pretty interesting is x dot ai, and they've created artificial intelligence based personal assistant that allows you to schedule meetings. Another company that's using natural language techniques to build conversational user interfaces is mine Mill, who was acquired by Cisco. And as you can see here, they're building really amazing voice and chat assistance on and things with voice. You can do things like commands eso, for example, playing music or watching certain TV shows and things like that. But then you could also have sort text or messaging base chat assistance as well. Another interesting example and we mentioned this before was G e mails. Smart reply on here you can kind of see it in the mobile form again, where you know, basically, they use natural language processing and understanding to try and come up with a really short, succinct but yet effective and relevant responses to e mails. Another example is Amazon's Lex, a p I. And this is also a conversational interface on Daz, they note here it's powered by the same deep learning technologies as Alexa. So pretty cool stuff there and then finally, Google alot And this is there Ai Bay sort of smart messaging application. 13. Hybrid and Miscellaneous Applications: all right, so we've covered a lot of this point. And this is the last video in the section of going over real world applications of artificial intelligence and machine learning on We've already talked about recommend, er systems prediction and classification, natural language, computer vision and recognition and so forth. So finally, we're gonna talk about a lot of other and quite amazing fields of artificial intelligence and machine learning things like autonomous vehicles which are becoming all the rage these days and even Google search and Home Security, for example, in robotics and things like that. And given that let's examine some real world applications and companies, they're doing some of these really amazing AI and machine learning techniques. One of the really exciting applications of artificial intelligence and machine learning that's becoming very talked about these days and hot in the press as it were, is autonomous vehicles and is only a get more so and I can't really wait to see what's coming out of that space here in the very near future. So going back to begin this class where we wanted to talk about goals and we really made a point to emphasize the wise. Why do we do these things and why do we use these tools and techniques? So in the case of autonomous vehicles, there's actually many, but one of the main ones is so that you can reduce accidents and related injuries and death . Another one is efficient. Ridesharing onda also improved traffic because of that improved fuel efficiency as well. Reduce carbon emissions, faster commutes and travel times, and also getting back your time in the vehicle that you'd rather spend doing something other than driving the car itself. So here are some of the few companies there really paving the way and leading the charge in the autonomous vehicle space one is UC's dot com or Dukes. Another one is notto or an auto, and another one is new autonomy. And again, these three are sort of the leaders in the space. This is really exciting stuff, and it combines a lot of what we've been talking about before from recognition task, computer vision, just all kinds of different elements of machine learning, artificial intelligence to solve many, many problems and achieve some pretty significant goals. So again, very exciting. It'll be amazing to see what comes out of this, you know, suit. All right, let's finish up this section by looking at some other really amazing companies and applications of artificial intelligence and machine learning. And then from there, the last video will be the conclusion and final steps for this class. So one of the biggest ones is Google Search, and it's easy to take it for granted and kind of forget about it in the sense that we use it every day. And so naturally use Googling. Something is even a verb that we use. But let's not forget that Google Search is actually one of the most impressive, comprehensive, an amazing applications of machine learning artificial intelligence basically out there. It's really truly incredible to think that they've got ahead and essentially index the entire World Wide Web and are able to, you know, present results to people. They're looking for something that are highly relevant, highly targeted and ranked in a very specific way. And, of course, Google's search does a lot more than just that. But God, it's definitely worth mentioning it. So we discuss that here. Another interesting application is a company called Flare that's using our official intelligence for home security and things like facial recognition, sound recognition and so on. Pretty cool stuff. Uh, N Y used doing something really amazing in which they're using some pretty involved and sophisticated artificial intelligence techniques to create photo realistic images just from text. So a lot. A lot of cool stuff going on with that. Another company called Jukt Deck is using a I to create highly specialized and highly targeted music for certain applications. A company called Digital Genius is in the customer service space and kind, trying to find all sorts of innovative ways, using artificial intelligence to really power, the next generation and the future of customer service and contact centers and all that sort of thing. Another interesting application is what's called decoder, and this is using artificial intelligence. It's actually right coat. So rather than a software engineer writing code or programmer finding ways to use algorithms to create code to do certain things, that's pretty fascinating. There's a company out there called typing D N A, and they're working on sort of a new wave New Age way of doing authentication instead of just the typical sort of user name and password based authentication in this case, they're doing something using keystroke dynamics. They point out here in order to do authentication. It's basically like saying that you know, the way people type is almost like a fingerprint in the sense that it could be fairly you need to individuals and therefore you could authenticate people that way. Another interesting cos benevolent a. I. And they're doing artificial intelligence to drive pharmaceutical creation and basically trying to bring high quality, highly relevant in sort of targeted pharmaceuticals to market as quickly as possible and as effectively as possible. And then finally, one that you may be familiar with already is. Zillow has their Z estimate, which kind of gives a value of homes. And as you know, when you have Realtors, let's say you're selling your house. Often you get compass and you know that's done very manually. And you know, people are realtors kind of look around and try and figure out the value of your home based on other homes nearby. And this sort of thing Z estimate from Zillow is a similar sort of thing, but it's driven largely by you know, the kinds of techniques and algorithms we're talking about here and taste into account. You know, a tremendous amount of different kinds of data in order to do that. So all very interesting. And I think at this point, you know, hopefully, as we've gone through all of the real world applications and talked about different companies involved in those, you've gotten a much better sense not only what artificial intelligence and machine learning are and what they're used for, but what's being done today by riel companies out there and what real world problems are being solved and what kind of goals are being achieved. And also, I hope you've kept in the mind in your mind the whole time that why that we're talking about. So why Why do they want to do these things? Obviously, companies do it because you know there why often, as we said the beginning with the business goals, things like making money and increasing profits and revenues and so on. But, you know, a lot of companies also do really amazing things that really do improve people's lives and things like that, too. So on then, of course, you know you as a customer, you've got your own goals in a lot of these products and services help you with that and also help you get your jobs to be done, as we discussed earlier in this class. So with that, let's conclude the Real World application section and move onto the last video of the class . 14. Summary and Next Steps: welcome to our final video and thanks for taking the time to take this class. We've covered a lot in a short period of time, and hopefully you'll walk away with many great takeaways. Remember, there's a class project as well, and I highly encourage you to give it a go. You can find all the project details on the projects with that thank you again and best of luck in all your technical and not so technical endeavors.