Pooling

Also: max pooling, average pooling

A parameter-free layer, common in CNNs, that slides a window over each channel and keeps one value per window (the max or the mean). It shrinks spatial size, cuts computation and makes features tolerant to small shifts.

Example

2×2 max pooling with stride 2 turns the patch [[1, 3], [2, 0]] into 3, and halves a 32×32×16 feature map to 16×16×16.

See it in the lesson: Convolutional Neural Networks →

Related terms

More in Neural Architectures