R·S = M

Why this quietly runs your life

Every time a game character turns to face you, or a robot arm swings into place, something is stacking transforms exactly like this: rotate the shoulder, then the elbow, then the wrist — three machines, one motion. Graphics engines, CGI rigs, even your phone's AR filter weld a pile of rotates and stretches into a single matrix before touching a single pixel, because doing that math once per frame beats doing it three times.

And the same trick, stacked six times instead of two, is the skeleton of "deep" in deep learning — with one twist you'll meet in the ReLU lesson: a tiny nonlinear squish sits between the machines, and it's the only thing that stops the whole row from welding back into a single matrix.

Two machines, one arrow

A matrix is a machine: arrow goes in, transformed arrow comes out. One matrix, one machine, one pass. Fine.

But what if you send your arrow through a machine, then another machine? First Rotate, then Stretch. That sounds like two trips. Is it?

Try it: send one arrow through two machines

Click the buttons in order. Keep an eye on the little matrix box — count how many grids of numbers it ever shows you.

The shortcut: weld them into one

Notice the matrix box never showed you two grids stacked up — after both clicks it still holds exactly one 2×2 grid. That's because multiplying two matrices builds one new matrix that does both jobs:

Feed your arrow into that single grid once, and you land in the exact same spot as the two-step trip. Matrix multiply is just "do this machine, then that machine" welded into one machine.

Order is not a suggestion

Reset the demo and picture doing it backwards — Stretch first, Rotate second. You'd land somewhere different. Rotate-then-Stretch and Stretch-then-Rotate are not the same trip.

Same reason socks-then-shoes beats shoes-then-socks: identical two steps, swap the order, embarrassingly different result. Matrix multiplication remembers order — A·B ≠ B·A, in general.

Do it six times and they call it "deep"

Here's the de-fang: a "deep" neural network is this trick, repeated — with one crucial extra. Transform the input, transform that result, transform that — a long hallway of matrix-machines (the real thing you'll meet later stacks six of them). But notice: if the hallway were only matrices, today's welding trick would collapse it all into one machine. So between every two machines sits a tiny nonlinear "squish" (ReLU — a rung coming up) that can't be welded through. Matrices do the transforming; the squishes keep the depth real. Not mystical — a hallway of machines with a doorman between each pair.