Pooling & Architecture

Pooling downsamples a feature map — typically taking the max of each 2×2 block and discarding the rest. It sounds destructive, and it is: you lose precise location.

That's the point. Once you know there's an eye somewhere in a patch, its exact position stops mattering. Max pooling keeps the strongest signal ('yes, an edge here') while buying you translation tolerance, smaller tensors, and — crucially — a wider receptive field: after several rounds, each neuron effectively 'sees' a large chunk of the original image.