Pooling and Downsampling
Module 4 · CNN Downsampling

Pooling and
Downsampling

See how a small window slides over a feature map, summarizes each local region, and produces a smaller output map. This is the part of CNNs that keeps the strongest local information while shrinking spatial size.

Pool type Max or average
Window Local region size
Stride How far the window moves
Goal Compress spatial detail
Key concepts
What to watch for
Local summary Pooling looks at a small patch at a time and replaces it with one summary number.
Max vs. average Max pooling keeps the strongest activation. Average pooling keeps the average response.
Downsampling Larger stride or window size gives a smaller output map and more compression.
Interactive Playground

Controls

Keep your original pooling logic, but present it with the same visual language as your overfitting / underfitting page.

Quick presets
Fast ways to jump between the most useful teaching cases.
Pool type
Max pooling
Change the local summary rule without changing the sliding process.
Window size
2×2
The size of each local patch processed at one step.
Stride
2
How far the pooling window moves before the next output cell.
Pooling walkthrough
Input feature map → pooling window → output map
Input feature map
Downsample
→
step 1
Output feature map
Current operation
Output(1,1) = max(4, 2, 2, 3) = 4
Why this matters
Max pooling keeps the strongest local activation. This is useful when you want the model to preserve whether an important feature appears anywhere inside the small region.
Pooling · Downsampling · Local summary · Spatial compression · CNN intuition