Deep learning in visual perception aims to extract features from the input data at different levels of complexity, from simple edges to complex shapes and objects, ultimately enabling the system to identify the most important parts of an image.