Max Pooling

Max pooling is a compact way for a network to keep the strongest signal in each small neighbourhood while shrinking the feature map. Think of it as asking, “Did this useful pattern appear anywhere in this patch?” rather than preserving its exact pixel-by-pixel position.

How the operation works
A max-pooling layer slides a window across each channel of an activation map and outputs the largest value in every window. With a 2×2 window and stride 2, each four-value region becomes one value, halving height and width. For example, the patch [1, 0; 3, 2] produces 3. Channels are pooled independently: a feature detector for one pattern does not compete with a detector in another channel.

What learning sees
During backpropagation, the gradient from each pooled output flows only to the input value that won the maximum. Frameworks such as PyTorch’s MaxPool2d record these winning positions, sometimes called argmax indices, so they can route gradients correctly. This makes max pooling selective: a strongly activated feature receives reinforcement, while its weaker neighbours receive no direct gradient from that pooling window.

Why it helps—and what it sacrifices
Max pooling reduces spatial size, which cuts the compute and memory needed by later layers. It also gives limited tolerance to small shifts: if an edge detector fires one position to the left but remains within the same pooling window, the output can stay unchanged. A classic convolutional network might use convolution, ReLU, then 2×2 max pooling repeatedly to build broader receptive fields.

• Benefit: fewer activations and stronger emphasis on detected features.
• Cost: exact location and fine detail are discarded permanently; this hurts tasks needing precise boundaries or alignment.
• Design choice: modern architectures such as ResNets frequently use stride-2 convolutions instead of repeated pooling, while global average pooling commonly replaces a final large pooling stage.