There are multiple types of deep learning algorithms, the one we generally use in computer vision is called a Convolutional Neural Network (CNN). CNNs break down images into small groups of pixels and then run calculations on them by applying filters. The result is then compared against pixel matrices they already know about. This helps CNNs to come up with a probability for the image belonging to one of the known classes.

In the first few layers, the CNN identifies shapes, such as curves and rough edges, but after several convolutions, they are able to recognize objects such as animals, cars, and humans.

When the CNN is first built for the available data, the filter values of the network are randomly initialized and so the predictions it produce are mostly false. But then it keeps comparing its own predictions on labeled datasets to the actual ones, updating the filter values and improving performance of the CNN with each iteration.