Pooling Layers
How pooling layers shrink feature maps, reducing computation and granting spatial invariance.
Convolutional layers detect features at exact pixel locations. If an object shifts slightly, the exact match breaks for subsequent layers. The network becomes overly sensitive to precise positioning and struggles with massive, high-resolution feature maps that are computationally expensive to process.
The Pooling Window
A pooling layer slides a window over the feature map, extracting a single summary value per region. Max pooling takes the highest activation—representing the strongest presence of a feature—while discarding the weaker signals around it.
Spatial Invariance
By downsampling the grid, the exact position of the feature is discarded while its presence is preserved. A small shift in the input image still activates the same pooled output region, granting the network translation invariance.
Where It Breaks
Aggressive pooling destroys the precise spatial information of where features live. This fails completely in tasks like image segmentation, pose estimation, or object detection, where outputting the exact, pixel-perfect boundary or coordinate of the object is the entire point.
The Quick Version
- Convolutional maps are huge and spatially sensitive.
- Pooling slides a window to extract region summaries.
- Max pooling keeps only the strongest signal.
- The output shrinks, reducing computational cost.
- It provides spatial invariance to small shifts.
- It fails when exact pixel locations are required.