Step by Step
F
One feature map per filter
After each convolution layer, the network produces one feature map per filter, with each feature map highlighting where a particular pattern (edge, curve, texture) appears throughout the image.
Example: a convolutional layer with 64 filters producing 64 separate feature maps, each one highlighting a different specific pattern detected across the image.
E
Early layers — simple features
Feature maps from early layers reveal simple features like horizontal edges and vertical edges.
Example: an early-layer feature map lighting up brightly along all the horizontal edges present in an image, while staying dark everywhere else.
M
Middle layers — shapes
Feature maps from middle layers reveal more complex shapes, like curves and corners, built up from combinations of the simpler early-layer features.
Example: a middle-layer feature map lighting up specifically wherever a particular curved shape appears in the image.
D
Deep layers — complex parts
Feature maps from deep layers reveal complex, recognizable parts, like eyes, wheels, or entire faces, representing highly abstract, task-relevant visual concepts.
Example: a deep-layer feature map lighting up specifically wherever something resembling an eye appears in the image, having been built up through many layers from simpler earlier features.
Applied Walkthrough
1
A researcher wants to understand what a trained CNN has actually learned to "see" at different depths within the network.
2
Visualizing feature maps from an early layer reveals simple patterns lighting up: specific feature maps activating along horizontal edges, others along vertical edges.
3
Visualizing feature maps from a middle layer instead reveals more complex shape-based patterns, like curves and corners, built up from those earlier simple edge detections.
4
Visualizing feature maps from a deep layer reveals highly abstract, recognizable parts like eyes or wheels — illustrating the full edges-to-shapes-to-parts hierarchy directly through visualization, like giving the researcher X-ray goggles tuned to see exactly what the network has learned to detect at each depth.
Exam Application
Exams test whether you understand that feature maps are produced one per filter after each convolution layer, and whether you can correctly match feature map complexity to network depth (early = simple edges, middle = shapes, deep = complex parts/objects) — directly connecting to the broader CNN hierarchy concept.
⚠ Common Trap
The most common trap is assuming feature maps are some separate, additional network component rather than a direct, natural output of the convolution process itself. Every single convolutional filter automatically produces its own corresponding feature map as part of standard CNN operation — visualizing these feature maps is simply a technique for understanding what the network has already learned, not a special additional step.
✓ Quick Self-Check
1. How many feature maps does a convolutional layer produce, relative to its filters?
One feature map per filter.
Tap to reveal / hide
2. What does a feature map highlight?
Where a particular pattern (edge, curve, texture) appears throughout the image.
Tap to reveal / hide
3. What kind of features do early-layer feature maps typically reveal?
Simple features like horizontal and vertical edges.
Tap to reveal / hide
4. What kind of features do deep-layer feature maps typically reveal?
Complex, recognizable parts, like eyes, wheels, or entire faces.
Tap to reveal / hide
5. Are feature maps a special additional network component, or a natural byproduct of convolution?
A natural byproduct — every convolutional filter automatically produces its own corresponding feature map as part of standard CNN operation.
Tap to reveal / hide