Convolution: Why Blurring a Photo Is Secretly a Multiplication

One operation runs photo filters, sensor smoothing, and the first layer of every neural network that recognizes faces.

Open a photo app and soften an image; the app runs a small grid of numbers across every pixel. Ask a phone camera to trace the outline of a face, and it runs a different small grid across the same pixels. Smooth a jittery temperature log from a cheap sensor, step a heat-diffusion simulation forward one tick, or push an image through the first layer of a neural network trained to recognize cats — and in every one of those cases the machine underneath is not merely similar. It is identical. One operation, called convolution, is doing all of it, and the only thing that changes from job to job is a short list of weights.

That alone would be a decent curiosity: a blur, an edge detector, a differential-equation solver, and a neural-network layer turn out to be the same arithmetic with the dial set differently. But convolution has a second, stranger property, and it is the one that makes the whole story hang together. The operation looks, on the page, like a laborious sliding-and-summing over an entire signal. In a different coordinate system it collapses into something a schoolchild could do: pointwise multiplication…