Deep learning represents the world as a hierarchy of concepts, each built from simpler ones

Goodfellow, Bengio & Courville define deep learning as an approach in which a computer learns to understand the world as a nested hierarchy of concepts, each concept defined through its relation to simpler concepts. Drawing a graph of how these concepts build on one another yields a deep graph with many layers — hence ‘deep’ learning. This is what lets a system learn a complicated concept by composing it from simpler learned ones (e.g. edges → corners → object parts → objects), rather than requiring a human to specify every concept directly. Depth here refers to the layered composition of learned representations, not merely a large model.

Examples

An image classifier learns pixels → edges → corners/contours → object parts → objects across successive layers; each layer’s concepts are defined in terms of the layer below it, so the concept graph is deep.

Assessment

Explain why the word ‘deep’ in deep learning refers to a hierarchy of concepts, and give a concrete layer-by-layer example of simple concepts composing into a complex one.