More Thoughts On Critical Windows
On the fetish post, I discussed people who had some early sexual experience - like seeing a sexy cartoon character - and reacted in some profound way, like becoming a furry. Sometimes people have described this as a “critical window” for sexuality (similar to the “critical period” in language learning?) where young children “imprint” on sexual experiences - and then can’t un-imprint on them later, even when they see many examples of sex that don’t involve cartoon animals.
One of my distant cousins won't eat tomatoes. His parents say when he was very young, he bit into a cherry tomato and it exploded into goo in his mouth, and he was so upset he wouldn't eat tomatoes from then on. Now he’s in his 30s and still hates them. Is this fairly described as a “critical window” for food preferences?
Both of these sound like trapped priors: a very strong early event coloring experience so hard that it becomes impossible to update away from it. This can happen whenever: trauma is an age-neutral example. So why are there so many examples in childhood?
In a typical AI training run, the AI starts at a very high learning rate (also called “temperature”) and gradually “cools” down over time. Think of this as a solution to a hill-climbing algorithm: in the early part of the run, you’re trying to explore the surface of the earth and find where the high mountains are; in the middle part, you’re trying to explore the Himalayas to find Everest; near the end, you’re trying to see if going a few feet east or west on Everest will bring you closer to the summit. So at the beginning, you might start in a random place and want to see if 500 miles away is more mountainous. But once your hard work has brought you 1000 feet from Everest’s summit, you don’t want to take a 500 mile jump and end up in New Delhi and have to start all over again.
Plausibly, early childhood plasticity exists for the same reason; children should have higher learning rates; each update should be larger. So maybe incidents like the tomato explosion are more likely to have long-term consequences because children’s very large updates are more likely to land them in trapped priors. Not only is an adult’s ability to update away from the tomatoes-are-bad attractor limited by the prior affecting their future judgments, they can’t make updates of that size anymore.
Many critical periods are more specific and better-defined than these: animals that can’t see out of one eye for the first few months of life permanently lose the ability to process that eye. So learning rate decay is probably better seen as a general principle which is implemented differently (and at different timescales, and with different levels of severity) in different systems.
This still doesn’t explain the unpredictable nature of preference-changing events. My cousin had a bad experience with a tomato in childhood, but was that really his worst childhood experience? I had a really scary time at a beach once during childhood, where I thought I would be swept away and drowned - but I got over it and now I like beaches just as much as everyone else. Was my cousin’s experience with the tomato worse (along some axis) than my experience with almost drowning? What about all of us who see cartoon animals but heroically resist becoming furries?
I think it’s helpful to notice these kinds of events seem more common in childhood - but even with a theory to explain them, they’re still pretty mysterious.