There is a way of teaching a machine that has almost nothing to do with instruction and quite a lot to do with practice.

It tries something, gets a rough signal back about how that went, and tries again. Nicole Junkerman on what that has in common with any ordinary repeated thing, and where the resemblance runs out.

Learning by trying

The familiar way to teach a computer something is to describe it. Here are the rules, here is what counts as correct, now apply them. There is another approach in which nothing much is described at all. The machine attempts something, receives a small signal about whether that attempt went better or worse, and adjusts. Then it does it again, a great many times, until the attempts settle into something that works.

Nobody hands it the answer. It is simply allowed to try, and given a rough sense of better or worse afterwards. That is close to the whole of the idea, and the plainness of it is the part worth sitting with.

Why the repetition is the point

Nothing dramatic happens on any single attempt

Anyone who has learned to swim, or to play an instrument, or to cook the same dish until it stopped needing a recipe, will recognise the shape of this. No single repetition is where the learning happened. The tenth attempt is not visibly better than the ninth. What changes is cumulative and largely invisible, and it only becomes obvious somewhere around the four hundredth time, when the thing has quietly become easy.

The awkward part, for machines and for people, is that most of the attempts are unremarkable. A process built out of unremarkable attempts is hard to feel enthusiastic about while it is going on, which is exactly why enthusiasm makes such a poor foundation for one.

What the comparison is good for

  • It makes practice look ordinary rather than heroic, which is much closer to how practice actually feels.
  • It separates progress from the feeling of progress, since the two rarely turn up on the same day.
  • It explains why a small repeated thing tends to outlast a large intermittent one.
  • It puts the emphasis on returning rather than on any single attempt going well.

The comparison has limits worth keeping in view. A machine that learns this way has no interest in what it is doing and nothing at stake in it. It does not get bored, it does not lose heart on a wet Thursday, and it is not fitting the attempts around work and a family. This is a loose analogy rather than a description of anyone.

A gentler way to think about starting again

Something useful does survive the comparison, though. In this way of learning there is no such thing as a wasted attempt, only an attempt that carried a slightly different signal than the last one. Nothing about a poor effort disqualifies the next one. The process assumes interruption and absorbs it, which is a kinder assumption than most plans are built on.

Earlier writing here on rest and attention circles the same ground from another direction, the general collection sits under meditation, and the full journal archive holds everything else.

It is a plain idea and it stays plain, which is most of its appeal. Try, notice, return. Very little about that has changed because machines now do a version of it too.

Frequently asked questions

What is reinforcement learning, in one sentence?
It is a way of learning in which a system tries something, receives a rough signal about how the attempt went, and adjusts before trying again.
Is the comparison with human practice exact?
No, and it is better treated as a loose analogy. A machine learning this way has nothing at stake and never loses heart, which is a large part of what makes ordinary practice hard.
What does the idea actually offer?
Mainly a change of emphasis. It puts the weight on returning to something rather than on any single attempt going well, and it treats interruption as ordinary rather than as failure.

This site shares personal reflections and general writing on meditation, mindfulness, and mental well-being. It does not provide medical advice, diagnosis, therapy, counselling, or crisis support.