SoylentNews is people
SoylentNews
Kids store 1.5 megabytes of information to master their native language
[...] research from UC Berkeley suggests that language acquisition between birth and 18 is a remarkable feat of cognition, rather than something humans are just hardwired to do.
Researchers calculated that, from infancy to young adulthood, learners absorb approximately 12.5 million bits of information about language — about two bits per minute — to fully acquire linguistic knowledge. If converted into binary code, the data would fill a 1.5 MB floppy disk, the study found.
The findings, published today in the Royal Society Open Science journal, challenge assumptions that human language acquisition happens effortlessly, and that robots would have an easy time mastering it.
"Ours is the first study to put a number on the amount you have to learn to acquire language," said study senior author Steven Piantadosi, an assistant professor of psychology at UC Berkeley. "It highlights that children and teens are remarkable learners, absorbing upwards of 1,000 bits of information each day."
For example, when presented with the word "turkey," a young learner typically gathers bits of information by asking, "Is a turkey a bird? Yes, or no? Does a turkey fly? Yes, or no?" and so on, until grasping the full meaning of the word "turkey."
Humans store about 1.5 megabytes of information during language acquisition (open, DOI: 10.1098/rsos.181393) (DX)
(Score: 2) by AthanasiusKircher on Friday March 29 2019, @12:38AM (3 children)
1.5 MB to contain all human knowledge about language? Seriously?
I'm reasonably certain our current attempts at natural language AI systems take up a LOT more than 1.5 MB, and yet they utterly fail at reasonable conversation that requires any understanding of content. Of course, that doesn't prove that one couldn't store this information much more efficiently somehow, but I doubt we'll ever create a fluently conversant AI with 1.5 MB of data.
Perhaps if they were just referring to basic grammar, syntax, morphological structure, phonetics, etc., without any actual datapoints for meaning of linguistic utterances, 1.5 MB could be enough. (And maybe assuming some sort of wacky Chomskian preprocessing system ready for language acquisition and "deep structure" grammar that there's basically no empirical evidence for.) But I skimmed TFA and it seems to imply this estimate is also supposed to take semantics into account.
I don't know -- once we have functional AI that can carry on normal conversation with actual understanding, maybe we could make an estimate for how much data needs to be stored. But I can't help thinking this estimate in TFA is really low.
(Score: 2) by Mykl on Friday March 29 2019, @02:58AM (1 child)
Agree - I think this figure is total bullshit.
While it may be possible to stuff a dictionary and set of grammar rules into a floppy, I doubt you could also fit:
That's all I can think of for now. I'm sure there are more. Now stick that in your floppy!
(Score: 0) by Anonymous Coward on Friday March 29 2019, @05:05AM
Yeah, how is learning 1000 bits/day remarkable? That's a weak-ass data rate, and anyway, bits and bytes are probably not the right way to measure human knowledge.
(Score: 2) by maxwell demon on Friday March 29 2019, @07:09AM
But that AI cannot use a large non-language related set of experiences to figure out what that language means. Our general world knowledge allows us to fill in lots of missing bits when we can't infer them from the actual language; indeed, that mechanism even sometimes leads to us hearing something different than what is actually said because of our expectations.
Or in short, the language-specific information is only a tiny fraction of the information we use when understanding language.
The Tao of math: The numbers you can count are not the real numbers.