Signal City Archive Desk · after Frank Rosenblatt, 1958 · back to the desk
Like the 1958 demo: cards marked on the left or the right. Draw a mark on the 20 by 20 eye (that's the Mark I's 400 photocells), then tell the machine what it is. It only changes its weights when it guesses wrong. Or press Run the 1958 demo and watch it see fifty cards.
Two switches. The light should be on when exactly one switch is on. Not both, not neither. Minsky and Papert proved in 1969 that a single layer can't learn this. Watch it try. Then give it a second layer.
| switch A | switch B | should be | it says |
|---|
A perceptron adds up its inputs, each multiplied by a weight, and fires if the total crosses a line. When it's wrong, the weights get nudged toward the right answer. That's all, here and in 1958. Today's AI, including every voice on Signal City, is stacked layers of that same idea, trained a different way. Still a machine for getting less wrong. Hear the episode: History No. 3, The Machine That Learned From Being Wrong. Everything runs in your browser.