Skip to content
ML Academy

Introduced on day 36

Feature map

The output of one filter at every input position, showing how strongly its pattern appears where.

In detail

A conv layer with 8 filters produces 8 feature maps stacked as channels. Maps stay aligned with the input positions (so a shifted pattern gives a shifted map), which is the equivariance that pooling then turns into approximate invariance. Feature-map size, not weight count, grows with image size and drives activation memory.