11 Jan 2026
Mateo Lafalce - Blog
In the architecture of Convolutional Neural Networks, the choice between Max Pooling and Min Pooling is rarely a matter of preference; it is dictated by how these networks represent information.
To understand why Max Pooling is the standard, one must first recognize that the pixel values in a feature map are not just arbitrary numbers, but indicators of confidence. When a filter passes over an image, a high value signifies a strong activation, meaning the network has detected a specific feature like an edge, a curve, or a texture in that location. A low value, or a zero, represents the absence of that feature.
Max Pooling is designed to capture and preserve these strong signals. By selecting the highest value within a given window, the operation effectively summarizes the region by reporting the most prominent feature found. It tells the subsequent layers that a feature exists somewhere within that patch, providing a necessary degree of translation invariance.
This ensures that the strongest evidence of a visual pattern is carried forward through the network, allowing the model to build complex hierarchies of understanding based on the most relevant data points available.
In contrast, Min Pooling performs the exact opposite function, which is generally detrimental to the learning process. By selecting the lowest number in a patch, Min Pooling discards the detected features, the high values, and instead preserves the background noise or the "silence" of the image.
In a network using ReLU, where the absence of a feature is often represented by a zero, Min Pooling would consistently pass zeroes to the next layer. This effectively erases the features the convolutional layers worked to detect, dampening the signal and preventing the network from learning meaningful patterns.
While Min Pooling can theoretically be used when features are represented by low intensity values, such as dark spots on a light background, standard practice relies on Max Pooling to ensure that the "loudest" and most important signals are the ones that survive.
This blog is open source. See an error? Go ahead and propose a change.