Average pooling downsamples by taking mean values, preserving more information than max pooling but potentially losing strong activations. Often used as final layer in classification networks.