Neural Network, Explained in Plain Language

A machine that learns by getting things wrong.

Dots and lines

sendsreceives

Dots hold a number. Lines carry it to the next dot. That's the whole machine.

In one end, out the other

pictureCATthe number travels left to right

Numbers go in one side. An answer comes out the other.

Every line has a dial

too littlejust goodtoo much

Turn a dial up and that line matters more. A big network has millions of dials.

At first, it's wrong

it saysDOGanswer wasCATthe gap between these two is the error

All the dials start at random. So the first guess is nonsense.

So nudge the dials

the mistake walks backwards

Every dial that pushed toward "dog" gets turned down a hair. A hair. Not a lot.

Now do that a million times

Signals flow forward through the network in teal, then the correction flows backward in red, then forward again, then backward again, forever. The dots' dials wobble as they get nudged. At first the network answers DOG in red; after a million nudges it answers CAT in teal.nudge #1nudge #1,000,000DOGCAT

Forward, then backward. Forward, then backward. Do that a million times and the dials settle somewhere that works. That's learning.