How models learn
← The series

Chapter 01 · Not written yet

Learning without a rule

Nobody writes the recipe.

You have never written down how to recognise handwriting, and you do it anyway. Instead of a rule, a machine with knobs on it and a way of scoring what comes out.

  • examples
  • parameters
  • scoring

How this chapter will go

Say what a model is, and what training means, without using the word algorithm.

The image it is built onA cook who has lost the recipe. Taste, adjust, taste again. Nothing else is going on.

  1. Something that already works

    A machine reads handwriting and gets it right. It is shown working before anything is explained, because the question only becomes interesting once you have seen the answer exist.

  2. The thing you cannot write down

    You recognise a handwritten seven instantly, and you cannot say how. Every rule you try to write has an exception, and the exceptions have exceptions.

  3. So write a machine instead

    Rather than the rule, write something with knobs on it, and a way of scoring what comes out. The rule is what the knobs settle into.

    what goes intwo knobswhat comes out
    Data goes in one end, a guess comes out the other, and the two knobs decide what happens in between.

  4. Turning them by hand

    With two knobs you can find good settings yourself, by feel, in about a minute. Try it here.

    knob A+0.00knob B+0.00how close0%what comes out
    • the shape wanted
    • what your knobs make
    knobs
    2
    how close
    0%
    best so far
    0%
    adjustments
    0

    Turn the two dials until the solid line sits on the dashed one. There is no rule to follow; go by feel.

    Turn the two dials until your line sits on the dashed one. Nothing here tells you which way to turn; that is the point. Then ask for twelve knobs and try again.

  5. Where the hand stops

    A real model has thousands of knobs, and no one turns those by feel. Everything that follows is one method for turning them without looking.

The misreading to head off

A model does not store the examples it was shown. When training ends the examples are thrown away, and only the settings are kept.

There is no date, no list to join and nothing to subscribe to. The chapter simply appears here once the drawing works.