Definition
Pooling is a technique that implements downsampling.
It consists of grouping adjacent pixels according to a specific rule, for example:
- the average (average pooling)
- the maximum (max pooling)
Example
Max Pooling in Action
Suppose we are halfway through a CNN and have a matrix, as shown in the figure.
If we want to obtain a matrix through downsampling with a factor of 2, one possible approach is to take the maximum within each non-overlapping region of size .
The result is the matrix shown in the figure: this is max pooling.
Important
Both max pooling and average pooling have no parameters to learn.
They are not learnable operations, but rather simple reduction functions.
Pros
- Fast
- Less sensitive to details: averaging suppresses noise (the mean is the enemy of noise)
- Max pooling is sensitive to the neuron’s strongest activation, i.e., the dominant pattern in the convolution output
Cons
- Not learnable:
it is only a calculation (max or average) and does not adapt to the data.- Positional information is lost:
once the maximum is taken, there is no record of where that value came from.This becomes a problem for tasks such as segmentation, where pixel-level precision is essential.
In segmentation, it is crucial to retain the exact position of each pixel throughout the network in order to map back to the original image and say, for example: “Here is the contour of the object we are segmenting.”
🌍 Global Pooling
If the region considered has the same size as the entire image, and only one maximum is computed, this is called global max pooling.
The same applies, mutatis mutandis, to the average with global average pooling.