How do convolution layers learn hierarchical feature representations in CNNs?

CNNs learn hierarchical features by stacking convolution layers where early layers detect edges and deeper layers combine them into complex structures.

CNNs build hierarchical representations through successive convolution layers. The first layers learn low-level features like edges, corners, and gradients. Middle layers combine these into textures and patterns. Deeper layers detect object parts and full objects. This hierarchy emerges automatically through backpropagation, where filters are optimized to minimize classification loss.