Is a method to reduce the computational complexity and control overfitting in Convolutional Neural Networks. This is done by reducing the dimensions of the feature maps, known as downsampling, by introducing a pooling layer.
A pooling layer works by summarising a set of adjacent neurons from the preceding layer by a single value. Pooling can thus be seen as a convolution layer that is fixed contrary to the other convolutions that are learned.
Two common ways of doing the summarising are average-pooling and max-pooling.
- Average pooling computes the average value of its inputs, with being a chosen kernel size, which corresponds to using a uniform kernel. If the stride is set to the resolution of the input image is down-sampled by a factor of .
- Max-pooling computes the maximum value of the inputs.