convolution
/ kon-vuh-LOO-shun /
Convolution sounds intimidating but the action is simple: slide a small stencil over a big picture and, at every stop, see how well the patch underneath matches the stencil's pattern. Imagine a tiny see-through card with a slanting line drawn on it. You glide it across a photograph; wherever the photo also has a slanting line, the card lights up; everywhere else it stays dim. Do this across the whole image and you get a new picture that shows where that one pattern appears.
More precisely, the little stencil is called a filter (or kernel) — just a small grid of numbers, often 3×3 or 5×5. At each position you multiply the filter's numbers by the pixel values beneath it and add the results into a single number. Slide one step over and repeat. The grid of all those output numbers is the result of the convolution, and it highlights one feature — an edge, a blur, a particular texture — everywhere it occurs. In a deep network the filter's numbers are not chosen by a human; they are learned during training, so the network figures out for itself which patterns are worth detecting.
Two ideas make convolution powerful for images. First, the same filter is reused at every position, so a pattern is recognized no matter where it sits — a feature called translation equivariance, and it means far fewer numbers to learn than connecting every pixel to everything. Second, each filter looks only at a small neighborhood, matching the fact that nearby pixels are what form edges and shapes. The honest limit: convolution bakes in an assumption that local, position-independent patterns matter most. That assumption is gold for photos and sound, but a poor fit for data with no such grid-like neighborhood structure.
A 3×3 vertical-edge filter has a column of +1s, a middle column of 0s, and a column of −1s. Slid over a photo, it gives a big number wherever the left side is dark and the right side is bright — marking every vertical edge.
A small grid of numbers, slid everywhere, hunting for one pattern.
Reusing one small filter everywhere is the trick that makes convolution efficient and lets it spot a pattern regardless of position — far cheaper than wiring every pixel to every neuron, as a plain dense network would.