Pooling and Receptive Fields

Summarize local evidence

Pooling reduces spatial size by summarizing a neighborhood. Max pooling keeps the largest activation, answering did this feature appear here?; average pooling keeps the mean. Downsampling reduces compute and makes exact small position changes matter less, but discards detail.

2x2 activations       max pool
[1 5]
[2 3]                 -> 5

A unit's receptive field is the region of the original image that can influence it. Stacking convolutions grows receptive fields: later units integrate evidence from broader regions even though each kernel remains small.

Scenario: Detecting a corner requires nearby edges; recognizing a face requires spatial arrangements across a much larger region. Layered receptive fields build from the first scale toward the second.

Warning: Aggressive pooling can erase tiny but crucial features, such as a hairline crack or small medical marker. Architecture must match the scale of relevant evidence.