How Does a Computer Learn?

Step 1 of 5

Some rules cannot be written down.

Try writing exact instructions for telling a cat from a dog. Every rule you write has exceptions. People recognise cats without knowing how, which means we cannot type it out.

Step 2 of 5

So it starts by guessing badly.

The machine begins with random numbers inside and guesses at random. Shown a cat, it says dog. Being wrong is the starting point, not a failure.

Step 3 of 5

Each mistake nudges the numbers.

Every wrong answer says which internal numbers pushed it that way, and each is shifted very slightly the other way. One nudge changes almost nothing; millions of them add up.

Step 4 of 5

It has to be tested on things it has never seen.

A machine can memorise its examples perfectly and be useless on anything new. So some examples are always held back, and only those count as a real test.

Step 5 of 5

It learns what was in the examples, faults included.

A machine shown only sunny photos will fail in the rain. A machine shown biased decisions learns the bias precisely. It has no way to know that the examples were unrepresentative.

This is why these systems need such vast numbers of examples. There is no understanding to fall back on — only what it has been shown.