How CNNs Build a Feature Hierarchy
See how CNN layers combine edges and textures into meaningful objects.
How CNNs Build a Feature Hierarchy explained visually
Trace simple bird features through early, intermediate, and deeper layers.
What you will learn
Follow filters, pooling, flattening, and classification through complete CNN examples.
EARLY LAYERS
Early layers respond to simple local patterns near the original pixels.
EDGES AND TEXTURES
Intermediate layers combine simple responses into textures, curves, and recognizable parts.
INTERMEDIATE LAYERS
Deeper layers represent larger structures because their effective receptive fields are wider.
OBJECT PARTS
Training guides the complete hierarchy toward features that improve the final prediction.
Quick check: test your understanding
What do early CNN layers detect?
Simple patterns such as edges and textures.
What do intermediate layers create?
More recognizable parts from simpler features.
What happens to receptive fields deeper in the network?
They cover increasingly large regions of the original image.
Accessible referenceLesson 43 transcript315 words · Matches the published narration and captionsOpen transcript
Imagine identifying a bird in a photograph. You do not recognize the bird from one pixel. You combine simple visual clues into increasingly meaningful patterns. A CNN follows a similar process by learning a hierarchy of features. The early layers operate close to the original image.
Their filters usually respond to simple local patterns such as edges, brightness changes, colors, and basic textures. An early filter does not understand the complete bird. It may only detect a line separating a dark object from a bright sky. The next layers receive the feature maps produced by those early filters.
These intermediate layers combine simple patterns into more recognizable structures. Several edges may form a curve. Curves and textures may form a feather, an eye, or part of a beak. Deeper layers combine those parts into higher-level representations. A late layer may respond to the overall shape of a bird or to the arrangement of a head, wings, and body.
This progression from simple patterns to complete concepts is called a feature hierarchy. The receptive field also grows as information moves deeper into the network. A receptive field describes the region of the original image that can influence a particular activation. Early neurons examine small regions, so they capture fine details.
Deeper neurons indirectly receive information from larger parts of the image, allowing them to understand broader structure. The network learns this hierarchy during training. We do not manually assign one filter to edges and another to feathers. The loss function and backpropagation guide the filters toward patterns that improve the predictions.
Different networks may learn different internal features, even when they solve the same task. The important principle is consistent. Early layers detect local details. Intermediate layers combine them into parts. Deeper layers represent larger structures and complete objects. In the next lesson, we will explain how local connections and shared weights make this process efficient.