From the course: Introduction to Artificial Intelligence

Step through the network

From the course: Introduction to Artificial Intelligence

Step through the network

“

- So what does it take to build an AI system? To think about this, let's go back to our challenge of finding dogs in images. The first step data scientists need to do, is figure out what they want from the data. In this case, they're not asking the AI system to cluster together its own groups. Instead, they're asking the system to classify data into two categories. One category will have images with dogs and the others will be not dogs. This is a classic binary classification challenge. Remember, that's when the neural network just has two possible classifications. This means they'll be doing supervised machine learning. Remember that supervised machine learning starts out with labeled data. Here, the system will be trained with hundreds of thousands of images known to contain dogs. The next step for data scientists is to figure out if they want to use standard machine learning algorithms, or if they'd like to use an artificial neural network. Remember that this is a classification problem, so if they go with machine learning algorithms, they'll probably either use K nearest neighbor or naive bayes. You've seen that the system will break down each image into pixels. That means that this is going to be a complex challenge with a lot of data, so they're going to use an artificial neural network. They'll create the input layer, hidden layers and output layer. Now, remember that since this is a binary classification challenge, there are only two options, dog or not dog, so they'll need just two nodes for the output layer. Next, the neural network will be initialized. The system will assign random numbers to all the weights of the connections. Then, the system will set the bias on all the nodes to zero. This is almost like shaking an etch a sketch to give itself a clean slate. Now, the training set needs to be fed into the neural network. The first few images will probably be not that much better than random guesses. The neural network will say something like, there's a 62% chance that the image contains a dog, or a 55% chance it's not. Then, the network will compare its answer to the label on the data. If it misidentifies the dog image, then it will look at the gradient descent to determine how much to change the weights and biases. The neural network will go through all the data in your training set to fine tune its results. Remember that the network will tune itself by using backpropagation to change the weights and bias to lower the cost function. In a sense, it will go backward through the network and twist all the dials to increase its accuracy. Once the artificial neural network has gone through the training set, then data will be added from the test set. The test set will not be labeled. It could be hundreds of thousands of images of anything. Then you'll see how well your neural network performed when identifying dog photos. Sometimes, the neural network will do very well with the training set, but not so well with the test set. When this happens, it usually means that you are overfitting the data. Remember, that's when the system's really good at identifying the smaller training set but doesn't have enough complexity to deal with the new data in the test set. Now, it's unlikely that your team will classify dog, or not dog, anytime soon. However, this approach to binary classification is a powerful way to get insights from your data. Think about your customer data, your sales data, or even data from your IOT devices and what you can learn from this approach.

Contents