My nine-month-old daughter, Ellie, is a statistics genius. This may sound like typical new mom bragging, but it’s not; it’s scientific fact. According to research from Professor Fei Xu’s Infant Cognition and Language Lab in the Department of Psychology at UC Berkeley, the average six-month-old is pretty good at making basic estimates of probability, and by the time they learn to walk—around a year old—most babies are experts. Children are also masters of language acquisition, pattern recognition, and inductive reasoning. In fact, in almost every arena that’s been investigated, babies and children are remarkably adept at learning. But while they may be excellent at figuring out the world around them, it’s still unclear exactly how much they know, and when, and what mechanisms are in place to allow this rapid learning.
Armed with colored ping pong balls, light-up lollipops, stuffed animals, and invented words, researchers in the Xu lab are making strides toward answering these questions. The answers they find may have applications in fields from parenting to computer programming.
What’s in a name?
It’s the oldest debate in developmental research: do we learn to learn, or are we simply biologically programmed to soak up information from our surroundings and experiences? The obvious answer, of course, is that it’s probably a little bit of both, but the precise location of the boundary between nature and nurture is a matter of intense debate.
Postdoctoral researcher Sylvia Yuan is investigating this boundary by studying word learning in toddlers. Previous research has suggested that by the age of two or so, children have a number of cognitive biases that help them solve the nearly impossible logic problem of what words mean. “Even if we explicitly label something, like, ‘this is a Lego’,” Yuan says, holding up a yellow block, “there are lots of logical possibilities as to what the word could be reffering to. It could be referring to the color yellow, it could be referring to something hard, it could be referring to something on my hand.”
But kids don’t run through all of these possibilities every time they learn a new word. Instead, they have a number of biases that help them narrow in on the right definition fairly quickly. For example, children tend to assume that a word refers to an entire object, rather than just part of it—“car” describes the whole vehicle, not just its hood. Similarly, they assume that labels apply to the shape of the object, rather than another characteristic, like color or texture. This is a useful assumption, because objects are likely to have a stereotypical shape (a “ball” is usually ball-shaped, a “cup” is usually cup-shaped) but may not always come in the same color or size.
These biases have long been thought to be innate, since they arise so early in development and are so universal. Intriguingly, though, children seem to weigh information differently depending on the type of object that’s being defined (color, for example, is more important when learning the names of foods, while texture becomes important when learning the names of animals), suggesting that experience might play a role in bias formation.
This process of bias-building is referred to as overhypothesis formation. “As they’re learning about each word,” Yuan explains, “they might be testing in their head: is it the texture, is it the color, is it the shape? And when they see another example, they might be thinking, ‘okay, it doesn’t seem like it’s the texture, it seems like it’s more the shape.’” As children form a hypothesis about what each object is called, they’re also forming a more abstract rule, or overhypothesis, that defines how object names are assigned in general.
Yuan and others in the lab are trying to determine what factors affect overhypothesis formation when children are acquiring new vocabulary. Instead of examining established biases, like shape, the lab introduces artificial categories so they can study the overhypothesis formation process as it happens. In one typical experiment, preschoolers must figure out that markings on the tail and left foot of otherwise identical stuffed animals determine their identity—for example, ones with a question mark on the tail and an exclamation point on the left foot might be “daxes,” while animals with a different set of marks on these two appendages might be “blickets.” Here, the usual biases aren’t helpful, so children have to learn not only what each animal is called (the hypothesis), but at the same time figure out a weird new rule that governs how these animals are named (the overhypothesis).
By manipulating this basic experimental setup, researchers can ask what variables affect how children form overhypotheses. One key finding has been that the number of categories presented seems to be more important than the number of examples per category. Preschoolers shown eight animals are more easily able to classify them if there are four categories with two animals each (two daxes, two blickets, two faps, two zoogs) than if there are two categories with four animals each (four daxes and four blickets). This suggests that each new category a child sees either strengthens or changes her overhypothesis about how categories are defined in general, indicating that it’s a dynamic process.
Yuan plans to use these initial studies as a launching point to investigate how other factors, like adding noise by varying the sizes or shapes of the items, or introducing exceptions to the rules, affect overhypothesis formation. Increasing the number of noninformative marks, for example, could go either way—it might help children focus in on the actually useful information more quickly, or it might just confuse them. “We’re trying to figure out what the environmental inputs are that make it easier or harder for them to achieve an overhypothesis,” Yuan says.
Behind door number three…
This sort of “smart” mechanism that allows children to draw up broad, organizing principles based on a small number of examples is crucial for learning. Without an efficient way to generalize the knowledge gained from one experience and apply it to another, it would simply take too long for kids to figure out how the world works. (And as anyone who has ever watched a child repeatedly test gravity with the food items on her high chair tray will tell you, it takes long enough as it is.) But until recently, it hasn’t been clear whether overhypothesis formation is limited to word learning, or when this skill first arises.
“Our working hypothesis is that there is a set of learning mechanisms in children that support rapid learning,” says principal investigator Fei Xu. She and others have predicted that even babies less than one year old might be able to form overhypotheses, but probing infant psychology can be difficult. Simply working with babies can be a challenge in and of itself. After all, there aren’t many fields in which papers routinely include lines like, “An additional four subjects were tested but excluded due to fussiness.” Entertainingly, researchers report that fussiness isn’t as much of a problem as bodily functions. Stephanie Denison, a graduate student in the lab, puts it delicately: “Occasionally they get distracted by… digestion during the trial.” Yuan elaborates, “We would have observers write down, for example, ‘face is all red and squinty’… the kids sort of stop looking at what’s going on on the stage and in the trial.” Distractability can also be problematic. “One little one just pulled off her socks in the middle of it. There’s a foot flying over there, a foot flying over here,” lab manager Christie Reed recounts. And the occasional baby will fall asleep during a study, too.
It’s also tough to find experimental methodologies that can truly illuminate infant cognition. “Smart as infants are, it is hard to work with them, since they do not yet talk or follow instructions,” Xu says. Researchers can’t just ask very young babies what they’re thinking—they have to figure it out in some other way. “We often capitalize on the fact that infants, just like older children and adults, are very curious,” says Xu. “They pay more attention to things that are new, interesting, and unexpected.” This is used to researchers’ advantage in the classic “looking time/violation of expectation” measure, a well-established test for determining what babies are able to predict. Because babies spend a longer time looking at things that are novel or surprising, an infant’s looking behavior can be measured to provide a metric of whether he finds an event expected or unexpected.
Measuring looking time was crucial for the Xu lab’s studies of overhypothesis formation in infants (as opposed to the toddlers in the object naming study). In these experiments, nine-month-olds watched while a researcher removed objects from various boxes. The first few boxes contained objects of the same shape, but of different colors and sizes. Then, surprise! The final box contained, say, a star and a circle. If the babies had formed an overhypothesis based on their previous experience—“boxes contain items with the same shape”—they should have looked longer at this unexpected event. And, indeed, this was the case. Importantly, babies formed overhypotheses equally well when the items in each box were all of the same color but different shapes, showing that this learning mechanism is general and not, say, the manifestation of an innate shape bias.
These experiments make it clear that infants can recognize patterns very quickly and use them to make generalizations at a very early age. According to Xu, that suggests the presence of a powerful learning mechanism that might underlie many different biases that were previously thought to be innate. Of course, it remains to be seen whether this mechanism is itself learned—are there over-overhypotheses to be discovered? In the future, comparisons between overhypothesis formation in infants and toddlers may also help illuminate how this process changes with age. If pattern recognition is something that improves with practice, it’s possible that young babies will have a harder time with confusing cases than more experienced toddlers and children; on the other hand, it’s also possible that the younger subjects may actually have an easier time because they haven’t yet learned to privilege certain kinds of information over others.
Masters of probability
Early in January, I had a firsthand look at studies investigating whether infants are able to use statistical reasoning to predict the likelihood of an event when my six-month-old daughter, Ellie, participated in an experiment in the Xu lab. After getting a basic rundown of the protocol from lab manager and researcher Christie Reed and signing some consent forms, we strapped Ellie into a high chair facing what looked like a puppet show stage in the experiment room. I was allowed to stay, but had to turn my back to the experimental setup. Babies pay close attention to cues from their parents, so any subtle shift in my behavior could have skewed Ellie’s responses and invalidated the results. On the other hand, babies are prone to meltdowns when left alone in a strange place. So: parents stay, but face away from the stage.
While Ellie watched from her high chair, Reed showed her a box containing a 4:1 ratio of pink to yellow ping pong balls. (Other versions of this study have used red, white or green balls—colors selected “entirely based on ping pong ball availability,” says Denison, one of the lead researchers on this project). After this demonstration, Reed took out different samples of ping pong balls, and filmed Ellie’s reaction when each sample was revealed. Was she surprised when the sample contained four yellow balls and one pink ball, instead of the opposite?
We weren’t told her looking time results (although, like most overbearing and/or intellectually curious parents, I did ask), but according to Reed, odds are pretty good that she was surprised and her reaction reflected it. “The four-month-olds aren’t doing all that well,” she says, “but so far the six-month-olds do have a grasp on it.”
Amazingly, older infants can even adjust their expectations based on other sources of information, from both the social and physical realms. For example, if, prior to the trial, the experimenter demonstrates a preference for white balls, 11-month-old babies will usually look longer at a sample that doesn’t match the researcher’s preference, even when it matches the contents of the box. More impressively, if the researcher is blindfolded the 11-month olds know to disregard the researcher’s preference and expect a representative sample. “This tells us that they understand something about some of the sampling processes, like visual access being important, random sampling versus non-random sampling, those kinds of things,” Denison says. Eight-month-olds, however, don’t adjust their expectations when the experimenter shows that they prefer a particular color. This suggests either that infants start to figure out other minds at some point between eight and 11 months of age, or that it takes a little while for them to apply that filter to the probabilitic intuitions they have already mastered.
Another permutation of this experiment—in which Ellie also participated this February—looks at how babies are able to “recalculate” expected probabilities. Babies were shown boxes containing three colors of ping pong balls, one of which was immobilized with Velcro. “We teach them that the ones with the Velcro don’t move, we obviously don’t think they know anything about Velcro,” explains Denison. The 11-month-olds were able to integrate the new information, and expected to see a sample that reflected only the remaining, mobile balls, showing again that babies’ probability estimates can be adjusted based on their knowledge about the physical world.
Now, Denison and others in the lab are investigating how babies deal with a slightly more sophisticated scheme, where some, but not all, balls of a particular color are immobilized. This effectively requires the babies to multiply two probabilities together, which should make it harder for them to predict what a representative sample would look like.
Determining how well babies can estimate expected probabilities under many kinds of conditions allows the researchers to probe more deeply into how infants arrive at these estimates. Humans are notorious for failing to evaluate probabilities accurately, depending on a variety of external factors (the 10 percent of American homeowners who are underwater on their mortgages can tell you about the perils of optimism bias), but not much is known about why we make the mistakes we do. Studying whether babies are susceptible to the same kinds of errors as adults may help us solve this cognitive puzzle.
Come on over, baby
While looking time is a well-respected and frequently used experimental measure, it can be tricky in practice. Looking behavior can be affected by many different factors, all of which need to be controlled. Even then, the difference between the reactions to an expected and unexpected condition may only be a few seconds. And of course, no matter how good the assay, it’s always nice to have a complementary experiment, particularly one that’s very different in approach.
With this in mind, the Xu lab has developed a novel, active, and frankly just darn cool measure of babies’ thought processes: crawling toward a hidden lollipop. For this assay, babies are first offered a black lollipop and a pink lollipop; whichever one they reach for or crawl to is established as the preferred color. Once a preference is determined, the babies are shown two boxes containing opposite ratios of pink to black, and the researcher removes one lollipop from each container in such a way that the baby can only see the stick. If this were a looking time experiment, the lollipop’s color would be revealed and the baby’s reaction would be monitored. In this new measure, however, the baby is allowed to crawl or walk to either cup to show that she knows which one is more likely to contain the preferred color. Eleven-month-olds pick the right cup about 70 to 80 percent of the time, showing that they have a reasonably firm grasp of single-event probability.
As with any research with babies, however, the crawling measure has its fair share of difficulties. First, it’s difficult to be sure that the baby has a true preference for one color. After all, babies can be fickle. The initial experimental design called for four preference trials, but the babies lost interest by the time the test trials rolled around, making the results difficult to interpret. “It’s always funny when as a researcher you think you’re doing this really intelligent, wonderful task, and the baby should be so engaged, and they’re like, ‘hmm, I think I’m just going to go see what’s over there on the door,’” Denison says. Short attention spans have also complicated the experiment in cuter ways. Some babies, when asked to select a lollipop, choose to hug the experimenter instead.
So, the experimenters try to make the single preference trial really count. After the selection is made, the researchers add some positive reinforcement, clapping and generally encouraging the baby to feel that she’s made a truly excellent choice. (This technique will also be familiar to anyone who has ever tried to convince a skeptical baby that she likes the new vegetable she just tried.) And at the end of the experiment, babies shown jars containing all pink or all black lollipops usually head for the preferred color, suggesting that the preference is consistent throughout the experiment.
The researchers also aren’t above using some tricks in an effort to achieve uniform color preference. “The pink one lights up now, which has made it much, much, much easier to get basically all the babies to prefer pink,” Denison says.
So far, this new method has been used to show that babies not only understand which bin is more likely to yield a pink pop, but that they can apply this understanding to guide their physical actions. Now, variations on this setup can be used to pick apart any number of cognitive processes, including overhypothesis formation. It’s also much easier to apply an active measure like this to non-human animals (in fact, Denison and Xu originally came up with the idea as something that could be used with rhesus macaques), and future comparative experiments are planned in monkeys and even squirrels. These comparisons may help us understand what makes human cognition so unique.
Lab to life
The Xu lab’s insights into baby cognition are fascinating in their own right, but there are also practical applications for this research. Increasingly, computer scientists are collaborating with developmental psychologists to create models of reasoning, learning, and language acquisition that inform artificial intelligence and natural language processing. There are also applications in clinical psychology. Infants’ performance in basic cognitive tasks like these is increasingly understood to be correlated with their abilities later in life, so more detailed knowledge of typical development may make it easier to identify atypical development at very early stages, when interventions would be the most effective.
Normally developing children can benefit from new insights, too. Knowing when specific cognitive skills are emerging can help parents and educators engage with these processes and give children richer learning environments. “I feel that if you’re aware of this sort of thing, that could make you interact with the baby differently, or maybe provide different kinds of stimulation,” Yuan says. Personal experience bears this out—now when I’m browsing at the toy store, I’m on the lookout for games that will challenge Ellie’s probabilistic reasoning skills. And when she approaches the age at which language acquisition explodes on the scene, I’ll be sure to rein in my use of expletives at just the right time.
Though major questions still remain, and the nature versus nurture debate is increasingly thought of as something of a straw man, work from the Xu lab and others in the field has certainly shown that babies are—as Denison puts it—“really, really smart.” As a doting mom, this just confirms what I already believed, but as a scientist? It’s nice to have some peer-reviewed citations to back me up.