Watch videos with subtitles in your language, upload your videos, create your own subtitles! Click here to learn more on "how to Dotsub"


0 (0 Likes / 0 Dislikes)
[AI Business School] [Introduction to AI technology for business leaders Introduction to machine learning and deep learning] >>Matt: Today we're going to dive into the basics of machine learning and deep learning. [Matt Winkler | Microsoft, Group Engineering Manager, Azure Customer Advisory Team] What I'd like to do is talk a little bit about how AI can be used across industries. We're seeing lots of different ways that we're optimizing businesses with AI. I just want to start with that. Then we'll go into the basics. We'll talk about what makes up machine learning. How does it work? How do you do it? And then we'll close with, why is this a topic now? We've been talking about artificial intelligence for 60 years. And why is it becoming a thing that we're talking about in every industry now? What I'd like to do first is start with some use cases across the financial services and retail industry. [Financial Services] This is just a small portion of all of the different ways we see people using machine learning inside of their businesses. But let's start with finance. And one of the best examples here is risk and fraud detection. Which is, how do banks, credit card companies know if a transaction is fraudulent? The way that we do that is we've got a lot of data about good transactions and bad transactions. We're going to use machine learning to see what we can figure out in order to predict if a transaction is fraudulent. Similarly, financial modeling, we have a ton of data about industries and securities, and how those are behaving. We want to take all of that data, and use that to create something which is going to make a prediction. Maybe it will be a prediction on how a stock will move. In the retail space, there's a couple of different places where machine learning is really, really important. The first is in managing the operations and managing your inventory and your supply chain. How can we optimize that? How can we learn to predict what inventory needs to be located where? How are we going to do that? We're going to have a lot of data about where have we had inventory, what is selling. We're going to want to pull in other data, like what's the weather supposed to be. And we're going to use all of that to predict what inventory should be where. And if we do that right, we should be able to improve the efficiency. The other thing that I'd love to be able to do is create recommendations for my customers. So that when a customer comes in or comes to my website, the things that I recommend to them are going to be things that are interesting to them, that they're going to be more likely to purchase. And if I do that right, then I'm going to increase my sales. Now, let's go into the basics of machine learning. If you go out, and you look up, "What is machine learning?", oftentimes, you'll see a lot of references to a model. And you'll see people talking about tuning models and training models, and tweaking them and deploying them. What is that? What's a model? [Building models] A model is just a function, just like any other function. The difference is what we're trying to do is learn the parameters of that function from all of the data that we have. And that's what machine learning is. It's taking a variety of different algorithms and techniques to take all of that data, and learn the parameters for that function. How do we do that? Typically, you'll be using Python or R, and frameworks like scikit-learn and a couple of others that we'll talk about in just a little bit. That's how you go, and that's actually how data scientists build models. What does a model do? What does this function do? Really, there's three different categories here. We can do regression classification, or clustering. Regression is the one that everybody remembers, which is "I've got a bunch of data. And I want to fit a line or a curve to that data." The way that we do that is we look at all those data points, we do some math, and we derive parameters m and b, for instance, in this equation where we're trying to draw a line. That's what machine learning is, is going through and finding what are the right parameters. The second machine learning problem is classification, which is, given an input, what class does it belong to? One really common example here is image classification. Which is given this image, is this a dog, a cat, or a tiger? This requires me to have a lot of pictures of dogs, cats, and tigers. Then I do some math on that, and create a function, which can take any picture, including those I've never seen before, and make a prediction. We don't have to just do classification on images, we can do classification on customers, such as, are they likely to churn? The final type of work that a model will do is clustering, which is trying to understand and segment all of my customers, and discover what are the different clusters of those customers. Similarly, search is a clustering problem. Which is, given a keyword, or given an article, how do I find other articles that are related to this topic? How do we teach a machine to go and do that? That's really what it takes to build a model. That's what a model does. That's how we can go and use it to do a regression, to do a classification or a clustering. But that's not the only part of the problem. You'll note, every time I've talked about a model, or talked about machine learning, I've talked about data. There's no machine learning without data. So a huge part of the machine learning problem is finding the right data, shaping it, cleaning it, and preparing it. Then I go and build and train the model. And then I have to actually go and do something with the model. That's where I deploy the model. And I might deploy it out to a machine in my factory, in order to shut it down in case it starts to exhibit anomalous behavior. I may deploy that into my web application to surface recommendations or detect fraudulent transaction. Those are the three stages of machine learning. Which is we find our data, we prep it, we then build and train the model, then we deploy. The last thing is once I've deployed the model, oftentimes I'm not done because I want to observe how that model works. I want to see, does it improve my customers' experience? Do I surface better recommendations? So there's a really nice life cycle that happens once I get started down this path. The last thing to touch on is, what is deep learning? There's a lot of buzz about deep learning. And deep learning refers to ways to build models, which use something called "neural networks." And one of the really interesting properties of neural networks is as you create deep networks... Those are lots of layers, you can actually learn really, really complex functions. And these things perform incredibly well when operating on unstructured data, like images or text or sound. The way that we go and do deep learning often uses frameworks, like Tensorflow, PyTorch, Chainer, and there's a host of others. And one of the really interesting things is that all of the math that's behind deep learning can actually be accelerated on specialized hardware like GPUs and FPGAs. And what this allows you to do is deploy models that are more complex, and thus, more accurate, into lots of different places. And it also lets us train on even more data. Which, again, helps us create a more accurate model. The last thing to touch on is why are we talking about this now? Why is this suddenly the hottest technology to learn, and one of the most important things that's impacting every business, and almost every customer we work with? There's really three things. The first is there's more data. We're able to capture and store more data about how our machine perform, about how our customers interact with the business. And I'm able to store that and process that. The more data I have, the more accurate a model I'm able to go and build. The second thing is that I've got a ton more compute available to me, and lots of different options about compute. One of the amazing things about the cloud is that I can very rapidly go and deploy a large number of machines. And this changes the economics of deploying that workload. I don't have to buy 200 servers with GPUs. I can just spin them up, and pay for them for a couple of hours. And then finally, there's been a ton of innovation in algorithms, tools, and frameworks over the last five to ten years. This is happening because the next wave of making better applications, making more personalized services and experiences is getting powered by insights gained from machine learning. This has gone and pushed the state-of-the-art forward. And what that means is that we all have available to us a ton of great work in order to start processing all of our data to build machine learning models to help our businesses.

Video Details

Duration: 9 minutes and 25 seconds
Language: English
License: Dotsub - Standard License
Genre: None
Views: 11
Posted by: csintl on Jun 18, 2019


Caption and Translate

    Sign In/Register for Dotsub above to caption this video.