What An AI Learns From A Baby’s-Eye View Of The World

JOHN DANKOSKY: This is Science Friday. I’m John Dankosky.

KATHLEEN DAVIS: And I’m Kathleen Davis.

There’s a lot to learn in the first couple of years of a child’s life, not the least of which is how to talk. But little kids don’t sit down and just study a vocabulary book. They soak up language from their daily experiences, which are often filled with parents and caregivers asking questions like, Do you want some water? or saying things like, Oh, look, a kitty cat.

Scientists wondered whether an AI model could learn about language the same way, not by being fed a curated set of pictures and words or the entire content of the internet, but by eavesdropping on the day-to-day activities of a two-year-old. Dr. Wai Keen Vong is a research scientist in the Center for Data Science at New York University and one of the authors of a report on that effort recently published in the journal Science.

Welcome to Science Friday, Dr. Vong.

WAI KEEN VONG: Great to speak to you, Kathleen.

KATHLEEN DAVIS: So you trained your AI model on about 60 hours of video. Where did all that video come from?

WAI KEEN VONG: Yeah. So this video was really the efforts of a number of pioneering developmental psychologists and cognitive scientists, who realized that we really needed data that better matched what children really experienced to formulate better theories. And so about a decade ago, a number of researchers got together and started coming up with how to gather this data. How could we get something that resembled the real experiences of a child and produce this mini-camera that they were able to attach to a number of baby’s heads and record them over the course of their developing years?

KATHLEEN DAVIS: OK. So I’m trying to imagine this for myself. This may be the cutest study ever done. I mean, what would it look like if I was looking at this baby with a camera?

WAI KEEN VONG: So the initial set of babies wore the camera on a headband or a strap. But a second set of kids, whose data is now being collected, involves a more sturdy and firm helmet.

It looks like a home video. You are seeing these kids playing with their toys. Their parents are reading books to them. They’re having mealtimes or they’re out in the playground on a slide or a swing. You’re really getting a real first-person glimpse at what the baby is interested in, like on a second-to-second level, really. It’s incredible.

KATHLEEN DAVIS: Wow. So you fed 60 hours of this, what I would imagine is a stream of consciousness data, to this AI model, as you said, which could include things like going to the park or talking to a caregiver. What did this model actually learn?

WAI KEEN VONG: Right. So the way the model works is it’s trained to learn language by associating what it’s seeing with what it’s hearing. So in cognitive science, this style of learning is what’s known as associative learning because you’re learning to associate certain words with certain images or certain objects out there. And we’re really the first to apply these kinds of theories to this representative subset of what children are seeing and hearing and seeing what it could learn.

So what it could learn is basically a number of the earliest object words, or nouns, that children also display when learning a language. So those are often the first words that children learn to speak. And we wanted to see if we could capture that aspect of learning from this data and this model.

KATHLEEN DAVIS: So what kind of words? I’m imagining like “dada” or “mama” or “cat” or “dog,” things like that.

WAI KEEN VONG: Exactly. Yeah, words like “bowl,” words like “crib,” or “car” or “chair.” Yeah, these objects pop up almost in every video that you see. And so they’re the ones that are relatively straightforward to test for as well.

KATHLEEN DAVIS: Were there any words that were surprisingly hard for this AI to learn?

WAI KEEN VONG: Yeah. I think one of the obvious failure modes we saw was that the model just couldn’t really acquire the word “hand” no matter how we tested it.

KATHLEEN DAVIS: “Hand,” like your hand?

WAI KEEN VONG: Exactly. Exactly. And one reason is that hands appear all over the video data set because the kid is manipulating things or playing with things. They always show up in the video no matter what frame or what words are being spoken. And so that is a bit of a challenge with how our model operates. And so we didn’t find a consistent learning with that word.

KATHLEEN DAVIS: Did the AI think that “hand” meant something else, or vice versa?

WAI KEEN VONG: Yeah, exactly. So if you look through the data, which I did quite a lot while working on this study, the word “hand” really only gets spoken when the child is playing at the beach or at the sandpit. And so because of this association between words and objects that is built into our model, the model mistakenly learns to associate the word “hand” with bits of sand instead.

KATHLEEN DAVIS: [LAUGHS]. That is very interesting. There are all sorts of television shows that we hear are supposed to be educational for young kids. Do you think that the model may learn more from watching Sesame Street or Peppa Pig than it would from an actual child?

WAI KEEN VONG: That’s a fantastic question. I don’t know. I do know a number of other researchers from the Netherlands ran a similar study, training a similar kind of model, but on episodes from Peppa Pig, and showed that it was able to learn a similar amount of words as our model.

Obviously, children start watching television a little bit later in life. We’re primarily interested in language at the very conception, at the very beginning of life, and showing how that emerges from this blooming, buzzing confusion that every newborn is experiencing.

KATHLEEN DAVIS: I mean, to be clear, we aren’t saying that this is necessarily how kids’ brains actually work, but just that this approach is enough to start to train this AI model. Is that right?

WAI KEEN VONG: That’s right. This work speaks to this larger debate in cognitive science around nature versus nurture. What needs to be built in for humans to learn versus what can we acquire from experience? So our model, our work, is really focused on that latter aspect, where we actually have very few elements, or innate components to our model, and much of the learning is really driven by the video data that’s being fed into our model.

So I would say it’s something like 5% built-in, or innate, elements and 95% experience. And I couldn’t give you a range on people’s opinions in the field of what the right split is, but I hope our work shifts people’s minds to realize that experience is really valuable and we don’t really need that much built in on top of that.

KATHLEEN DAVIS: There are parts of language that aren’t just the words, but things like tone. Do you think that you could train the AI to understand that it’s different for a calm voice to say “Watch out for the dog,” or a scared voice saying, “Watch out for that dog”?

WAI KEEN VONG: Yeah, absolutely. I think that’s something in the future we’d definitely like to try. Like I mentioned before, most children’s words are these common objects or these concrete nouns. But children also learned words like “uh-oh” quite early on in life, right? And the tone in which that is spoken indicates they’re recognizing something about the situation is surprising or a mistake has been made. So I think tone is a huge factor.

Another thing that many other researchers have looked into is this notion of child-directed speech, and recognizing that the way that we often talk to babies is different from how we talk to adults. We often accentuate certain words or certain phonemes. And other researchers have shown that that also might be a pedagogical strategy to help with learning.

KATHLEEN DAVIS: If you gave this AI model more hours of video to work with, do you think that it would just get better at working with this set of nouns that it learned, or would it possibly have a bigger vocabulary?

WAI KEEN VONG: I think the answer is probably both to those. We see that in children as well. We see that children, a little bit after the age of one, really show this large comprehension boost in their word recognition abilities. And we’d hope that our model displays a similar qualitative effect if we were able to feed in more data.

Now, that’s something we’re actually already working on. So stay tuned for that.

KATHLEEN DAVIS: OK. And so far, as you said, you’ve been pretty focused on these concrete nouns. But I mean, what’s next? Do you do color or verbs? How do you expand this language understanding to the next level?

WAI KEEN VONG: Yeah, absolutely. I think going to verbs, to color words, to adjectives, to prepositions– I mean, we really want, in some shape or form, to be able to explain all of the different elements of language that children eventually acquire.

In this work, we really focused on this multi-modal aspect of linking language to vision because that is really where children get started. But later on, it seems like children might have other mechanisms available, perhaps similar to the next token prediction objective in language models, that might enable acquiring some of these more complex words.

I’ve looked into this question a little bit for verbs. It’s quite funny because you think, if a child is learning a word like “run,” if you look in the videos when the word “run” is spoken, it’s because the kid is running, and the camera is really just bobbing up and down. So feeding that into the model, it’s going to have a very different notion of what running is, I think, to most people.

KATHLEEN DAVIS: Right. Well, that’s all the time that we have for today. Dr. Wai Keen Vong is a research scientist in the Center for Data Science at New York University. Thank you so much for taking the time to talk with us today.

WAI KEEN VONG: Thank you for all of your lovely questions, Kathleen. It was great to speak to you. 

Copyright © 2023 Science Friday Initiative. All rights reserved. Science Friday transcripts are produced on a tight deadline by 3Play Media. Fidelity to the original aired/published audio or video file might vary, and text might be updated or amended in the future. For the authoritative record of Science Friday’s programming, please visit the original aired/published recording. For terms of use and more information, visit our policies pages at http://www.sciencefriday.com/about/policies/