How CNNs See

A dense layer looks at every pixel at once — and must relearn 'edge' separately for every position. A convolutional filter is a tiny window of weights (often 3×3) that slides across the whole image, asking the same question everywhere: is my pattern here?

Same weights, every position — so a filter learned in the top-left works in the bottom-right for free. The output is a feature map: a heat map of where the pattern fired.