From the course: Introduction to Artificial Intelligence

Artificial neural networks

From the course: Introduction to Artificial Intelligence

Artificial neural networks

“

- Machine learning has gotten a big boost from artificial neural networks. An artificial neural network is an AI system that mimics the structure of the human brain and it's currently one of the most popular approaches to machine learning. Think of it this way. When I was a kid, we used to play a game called animal, vegetable, or mineral. The point of the game was simple. Someone would think of an item. Then all the other kids would ask questions about what the item was. We'd ask things like, does it have fur? Is it bigger than a house? Then in another round, we'd ask if it's alive, does it make noise? After each round, we would start to zoom in to identify the item. Then at the end, we started making guesses. Is it a horse? Is it a cat? Is it a dog? An artificial neural network uses a very similar approach, except instead of asking questions, the neural network uses hundreds or even millions of numerical dials. Also, the network makes much more specific guesses. It would say that there's a 64% chance that it's a cat or a 32% chance that it's a dog. To build the network, there's a row of neurons on the left called the input layer. Then there's a row of neurons on the right called the output layer. All the neurons in between are called the hidden layers. They're hidden because it's not the input or output. To start, we'd feed the input layer of the network a picture of a dog. We know it's a picture of a dog because we've labeled it but the neural network doesn't look at these labels until it makes a guess. Just like the animal, vegetable, mineral game. The network would then look at each dot in the image as it passes through the hidden layers. Then it makes a guess for the output layer. Let's say for the first guess, it says there's a 10% chance the image contains a dog. Then the network compares its guess to the label on the image. This is called training the neural network. Then the network goes backwards and adjusts the neuron dials so they can see patterns in dog photos. Then it takes another dog image and sees if it gets closer. The neural network then tunes itself by looking at hundreds of thousands of pictures of dogs. Then it will adjust its own dials until it consistently guesses correctly. Now, it's important to keep in mind that the neural network isn't seeing the dog the same way you do. It doesn't think about panting, barking, or fur. Instead, the system only sees the dog as a recognizable pattern of dots in an image. When there's dots in a dog-like pattern, then the system makes a guess. Now, remember that these neural networks are still machine learning systems, so the network needs access to huge amounts of data. If you don't have hundreds of thousands of pictures of dogs, then there's no way for that artificial neural network to learn. In the end, this is the key benefit of an artificial neural network. It can train itself to understand patterns and recognize that input when looking through massive amounts of data.

Contents