Module 4 · CNN Representation

CNN Feature Map
Visualization

A feature map is the output produced after a filter looks at the input or at a previous layer. Shallow layers usually respond to edges and simple texture, while deeper layers combine those signals into larger parts and more abstract attention regions.

Input One image
Layer choice Shallow to deep
Feature maps One per channel
Goal See what the network attends to
Key concepts
What to watch for
Each filter creates one map The same image can produce very different outputs because each filter searches for a different pattern.
Shallow layers are local Early feature maps often light up around edges, strokes, corners, and simple textures.
Deeper layers are more abstract Later maps combine earlier responses, so they can highlight larger parts or a more semantic region.
Interactive Playground

Controls

Pick an input image, move across shallow → middle → deeper layers, and click any feature map to see what region it highlights on the original image.

Input presets
Different shapes make different filters activate. Try switching presets before changing the layer.
Layer depth
Layer 1
Earlier layers respond to simple patterns. Later layers summarize larger combinations.
Animation
Layer / map playback
Animation helps you see how the highlighted region changes as the representation gets deeper.
Current map
Vertical edge
Feature map idea
map = filter ⨂ input → activation
This topic is about what the outputs look like, not about doing long convolution arithmetic by hand.
Input image and feature maps
Left: original image · Right: one layer's channels
Input image + selected attention
Click any feature map on the right. The colored overlay shows where that map is strongest on the original image.
→
Feature maps in the current layer
Current operation
Layer 1: simple filters respond to edges, contrast, and small local structure.
Why this matters
A feature map is useful because it converts the raw image into a representation where important patterns become brighter and easier for later layers to reuse.
Feature maps · Filters · Layer depth · Attention regions · CNN intuition