The network builds features to link the input data with the output data. It has to create representations to understand what is in each screenshot, the HTML syntax, that it has predicted.