From the course: Introduction to Artificial Intelligence

Weighing the connections

From the course: Introduction to Artificial Intelligence

Weighing the connections

“

- As human beings, we add weights to our data all the time. We look at the features of the data to better predict our output. Let's say that you're looking at a photograph of a beautiful grassy open space. Then you see a little blurry object in the photo. What do you think the odds are that that blurry object is a dog? Now, imagine you're looking at an image of a dry desert. This picture also has a little blurry object. What do you think the odds are that this object is a dog? If you're like most people, you'd guess that a dog is much more likely to be in a grassy field, so your human neural connections added a positive weight to the grassy field and a negative weight to the arid desert. Artificial neural networks do the same thing. Like us, these networks need to work in a world of probabilities. It's possible that there's a dog in the middle of the desert, but if you're an artificial neuron, you're going to be very skeptical about activating. An artificial neural network is structured in a way so they can better tune itself to understand your data. It's almost like a self-tuning musical instrument. To tune an instrument like a guitar, you need knobs to twist as you strum the note. With artificial neural networks, these knobs change the weights of your connections between your neurons. An artificial neural network adds weights to the connections between neurons in each layer. Each neuron in the hidden layer feeds forward into every other neuron in the next layer. So if there are 100 neurons in every hidden layer, each neuron in that layer will have 100 connections going out. That's a lot of connectivity. But where it gets really powerful is that each one of these connections will have a weight. That's why if you've ever seen a sketch of a neural network, you'll see that each one of the connection lines has a W with a number. So in this case, you would have a W1, W2, W3, all the way up to W100. You'd see this for each one of the hidden layers. Now, the weights in each one of these connections is a key part of how an artificial neural network tunes itself. Keep in mind, an artificial neural network is just a form of supervised machine learning. So, data scientists use the same technique that they've used to train the network. Remember that supervised machine learning starts out with a training set. Then once the algorithm is tuned to make accurate predictions, you can then move on to the larger test data set. The same thing happens with your artificial neural network. When you first initialize your neural network, the systems will randomly assign numbers to these thousands of weights. Then you'll feed your training data into the network and let the system adjust the weights based on whether you're getting correct output. The network will repeat this over and over again until it's accurately identifying the patterns for the output. It'll tune itself over time to zero in on the best predictions.

Contents