This is the fourth article in the series entitled “What is AI?” If you have not already read the previous articles, we recommend that you do so. To access them click here.

Let us also illustrate supervised learning with an example. Imagine a child learning to read. Often there is an adult pointing to a letter of the alphabet and then making the relevant sound. With the adult’s help, the child learns to link the letter to a particular sound. In the same way, the algorithm too needs to be properly trained. During this training process, the algorithm is supplied with data linked to a label that reflects a type of correct or desired value for that particular data. After the algorithm has been trained, the model is verified against data that contains labels but which the algorithm has not previously encountered. See the image below for a graphic illustration.

A somewhat technical video showing how teaching of an Artificial Neural Network (ANN) takes place can be seen here, where input data (from the left) consists of three figures with an attached label that is also a figure (on the right). This is followed by quantification of the discrepancy between the label and what the network believed, and this is then used to correct the network. As this process is continuously repeated the network gradually becomes better, resulting in fewer discrepancies. The network can quickly become very large and complex, depending on the size of the data for which the algorithm is being trained. An illustration showing how different networks can be built up can be seen in this video. When a black-and-white image of a hand-written figure between 0 and 9 is fed in, the network can determine what that figure is.

The application area is very broad and this approach is ideal for problems that are complex and that contain a lot of data. One example of a model is face recognition on Facebook, which automatically tags people in photographs. Previously, users themselves tagged where in the photo each person was, and his or her name. With the help of these labels for a particular image, it is possible to train a model that automatically suggests the location and name of each person in a photo.

Another somewhat unusual example comes from a farm in Japan that grows cucumbers. The problem here is that the cucumbers have to be classified into nine different types depending on shape and quality. In the following article you can read more about how a machine was built to automate this classification process.