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] - David Carmona: Hello everybody. In this section, we will see a quick of a view of the primary concepts of AI. But before we start, let me first introduce what is Artificial Intelligence. The most common definition refers to the ability of machines to show capabilities that are usually associated with human intelligence. That's a very very big definition, so let's focus on the three primary capabilities that you will usually see associated with AI. The first one, learning. It is the ability to learn over time. Traditionally, computer programs are created by a developer and they are deployed at some point, right? For the program to improve over time, the developer will need to update the code and deploy it again. With AI, the machine can learn over time without that direct human intervention based just on the experience. The second one is perception. Perception is the ability to understand the world around us. For example, recognizing an object in an image like a hand or understanding a human speaking like me speaking right now or being able to interpret a piece of text like for example, dislike. And, the last one is cognition. Cognition can be very broad, so it is the trickiest of the three. At a high level, it is the ability to reason over data. If perception is about interpreting data, cognition is about reasoning on top of that data. That could be for example identifying [INDISCERNIBLE] or making predictions of what is going to happen or identifying the best possible action for a given outcome. Now, how do we actually create AI? Doing these three things, learning, perception and cognition with traditional programming is extremely difficult. It was simplified a lot. So, the developers create an algorithm that turns an input into an output so the developer would write a program with the steps that combine with the input turns it into an output. Now, imagine using that approach to do for example perception like recognizing an object in an image. The developer would need to think of a program with things like, if, then, else that is able to generalize such a difficult problem. These has been done for simple cases by very smart developers but is not practical at all. Now, machine learning takes a very different approach. Instead of manually creating the algorithm, we use a lot of inputs with the corresponding outputs that we call training data. With that training data, the system will create an algorithm that maps the inputs and the outputs. So, instead of manually creating an algorithm, we train it with a lot of pairs of inputs and outputs. Once, we have the result which we call a model, we can apply to new inputs to get the outputs. For the previous example, the training data could be thousands of images, each of them with a corresponding object identified for example, a cat or a dog. Now, with that training data we can train the model that you can use for a new picture. The model will take that picture as an input and it will provide an output with the guess of whether that is a dog or a cat. Now, there are many many algorithms that have been used for decades in machine learning. The easiest example could be a linear regression. Linear regression algorithm calculates the output based on a linear function applied to the input. The training phase in that case will be the parameter for a linear function that can limit that transformation between the inputs and outputs. And, that can work for simple scenarios, but it can be challenging for others. In the example, we used before image recognition, the relationship between the inputs, images, and outputs where you send that image is way more complex than a linear function. For those cases, there is a very very special machine learning algorithm that I am sure you have heard a lot lately. It is based on the concept of artificial neural networks. An artificial neural network is just another machine learning algorithm. So, it is trained with pairs of inputs and outputs. But, in this case, the algorithm is loosely based on how a brain works with individual neurals that are connected in a network. Each artificial neural is extremely simple. It takes several inputs, it applies a weight on each of those inputs and then it applies as a simple activation function to create an output. When you have many neurals in every ledgers and many ledgers of those neurals, we can see amazing results for complex problems compared to more traditional machine learning algorithms. We call that Deep learning. Deep learning has been a primary area of research for the past years. We have learned new deep learning architectures that can be applied to different fields like vision speech or natural language processing. This one for example is called Residual Neural Network. Each box in this picture is a layer. This one in particular has 152 ledgers. You can see them all here. Resnet is used a lot in computer vision. But you have many other different possible configurations and even different architectures. That can make deep learning very daunting to use. So, it is important that you also know the technologies and tools that you can use to make it easier. [Microsoft]

Video Details

Duration: 6 minutes and 16 seconds
Language: English
License: Dotsub - Standard License
Genre: None
Views: 8
Posted by: csintl on Jun 18, 2019


Caption and Translate

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