The day the recycling line taught a computer to see
The conveyor belt at the recycling centre rattled and shook, pushing a messy mix under one big sorter. It kept choking. The manager tapped the rail and said, “Stop forcing everything through one machine. Split it.”
That same jam used to happen when people built photo-spotting software. They’d just make one giant checker, bigger and bigger, and shove every picture through it. It got slow and spent effort where it didn’t need to.
Then someone copied the recycling line. At one spot in the picture, the system runs a few checks side by side: one looks for tiny clues, one for medium ones, one for big shapes, plus a quick smoothing pass. Then it bundles the results together. Takeaway: check several sizes at once, and you miss less.
But even a split line can bog down if each station is costly. So the trick was a fast pre-sort first, like tossing items into a few rough tubs before the fussy machines. In the picture system, a tiny scan does that job, so the heavier checks have far less to chew on.
They stacked loads of these multi-lane stages, and the whole thing stayed workable instead of turning into a money-pit. While it was being trained, they also added small side judges partway along, like temporary inspectors, so the early parts didn’t get ignored. Later, those side judges were taken off.
In 2014, this design, called GoogLeNet, did brilliantly in a big image-recognition contest without needing as many stored numbers as many older, heavyweight systems. The shift was simple: not “make it all bigger”, but “spend effort where it pays”, like that recycling line that finally stops backing up.