100 million adjustable numbers

In 2009, researchers trained a deep network with 100 million parameters. That means the learning process had 100 million adjustable numbers to manage.

To feel that scale, imagine touching one number every second without stopping. Reaching all 100 million would take a little more than three years.

A computer works much faster than that. But training does not merely touch each parameter once. It calculates how the parameters should change, adjusts them, then repeats the process again and again.