One of the plainer ideas behind the current wave of machine learning is also one of the oldest ideas about learning in general. A system tries something, is told how well it did, adjusts, and tries again. Repeat that enough times and the behaviour improves, without anyone having written down the rule that was eventually arrived at.
Put that way it sounds unremarkable, because it is exactly how a person learns to read a room, cook a dish properly or tell a good picture from a merely competent one. Nicole Brachetti Peretti finds the resemblance worth sitting with, not because the two are the same, but because the machine version makes the human one easier to describe. What follows is not a method. It is a short list of the things that have kept turning out to be true.
- Progress is invisible from the inside. The tenth attempt feels almost exactly like the ninth. It is only across a few hundred of them that anything shows, which is why the middle of learning anything feels so much like standing still.
- The feedback matters more than the effort. Repetition without any signal about how it went produces a very well practised version of the original mistake. Something has to come back, even if it is only the faint sense that a sentence does not sound right yet.
- Most of the attempts are wasted, and none of them are. The ones that go nowhere are not overhead on the process. They are the process, and the successful attempt is simply the one that happened to come last.
- Speed is not the variable. A great many quick repetitions and a few slow ones are different activities producing different results. The reading notes keep finding that the slow kind is what survives the year.
- The habit outlasts the enthusiasm. Enthusiasm is what starts the thing. It reliably runs out somewhere around the fourth week, and what carries on afterwards is whatever had already become ordinary.
The comparison with machines should not be pushed too far. A system running through millions of attempts is doing something no person does, and it is not bored, discouraged or curious while it does it. Boredom and curiosity are not inefficiencies in the human version. They are most of what decides which things get practised at all.
What the resemblance does offer is a small reassurance. The slow, unglamorous middle of learning anything is not a sign that it is going badly. It is what the process looks like from inside, and it looks that way for everyone and everything that has ever learned by trying.
The companion note on looking closely when machines look too takes the same question into the gallery, the art pages keep the longer argument, and the journal collects the rest.