What infant fMRI is revealing about the developing mind

A decade ago, Nick Turk-Browne unintentionally created a paradox. He demonstrated in a series of studies that statistical learning, or the ability to extract patterns from experiences, depends on the hippocampus, a brain region also involved in forming episodic memories. And yet he knew from past work that babies, who were thought to not form such memories, are excellent statistical learners—it’s one of the main ways they acquire language.

“The lore in the memory literature was that the hippocampus doesn’t really come online until age 4 or 5,” says Turk-Browne, professor of psychology and director of the Wu Tsai Institute at Yale University. So how, he wondered, could infants be such good statistical learners?

The only way to settle this conundrum, Turk-Browne reasoned, would be to measure activity in the hippocampus of awake infants, like he did for his statistical-learning studies in adults. But babies are terrible functional MRI (fMRI) participants: They wiggle, cry, fall asleep and cannot take instructions, which is why they are typically scanned while sleeping or sedated. On the other hand, the neuroimaging methods used in awake infants can’t record activity from the hippocampus and other deep brain regions.

Turk-Browne and his team “somewhat naively” endeavored to make fMRI a suitable method for awake infants, he says. “We had to try it, because it’s the only way of measuring brain activity in some of these really critical systems.”

It took a lot of troubleshooting, but the effort eventually produced data that untangled the contradiction: The hippocampus activates during statistical learning as early as 3 months of age, the team reported in 2021. By age 1, it is also active during memory encoding, meaning our lack of infant memories is likely due to a memory retrieval issue, according to findings Turk-Browne’s lab published today in Science.

Turk-Browne is part of a small but growing number of cognitive neuroscientists who have taken on the challenge of scanning awake infants. “It’s become a field in kind of an amazing way,” says Rebecca Saxe, professor of brain and cognitive sciences at the Massachusetts Institute of Technology. After a decade of technical fits and starts, the method is poised to answer questions about how and when the infant mind takes shape—questions that, until now, the field has been unable to tackle.

M

uch of what we know about the infant mind is based on where babies look and for how long. Behavioral tasks often measure “looking time” to infer what an infant pays attention to and remembers. “Very detailed cognitive models have been generated based purely on behavioral data,” says Richard Aslin, senior scientist at Haskins Laboratories at Yale University, who studies language development using primarily electroencephalography and functional near-infrared spectroscopy.

But looking time is an opaque measure, Aslin says. For example, in some situations an infant may prefer to look at a novel object, but in others they may gaze at a familiar object for a longer time. “Behavior alone is not sufficient,” Aslin says. “If you want to pinpoint the underlying mechanism, you pretty much have to study the brain.”

The first fMRI scans of sleeping or sedated infants took place in the late 1990s. They measured resting-state activity and revealed “how competent the developing brain already is,” says Petra Hüppi, professor of medicine at Geneva University, who was an early adopter of fMRI in infants. Resting-state work continues today, through collaborative efforts such as the Baby Connectome Project and the HEALthy Brain and Child Development Study.

Combining infant neuroimaging with behavioral data and adult cognitive neuroscience studies can help “triangulate the answer to questions about the developing brain and mind,” says Michael C. Frank, professor of human biology and psychology at Stanford University. “Putting those different sorts of information together, ideally with computational theories as the glue, then provides our best guess as to what’s going on.”

But scans in sleeping babies cannot unpack the neural mechanisms driving “high-level cognitive function,” says Ghislaine Dehaene-Lambertz, scientific director of the developmental neuroimaging lab at Neurospin. For that, the babies need to be awake.

The first attempts to scan awake infants were riddled with technical hurdles—and not-so-technical ones. In one early experiment, Dehaene-Lambertz played normal and reversed clips of speech to 2- and 3-month-old infants while they lay in the scanner. “The first baby was perfect,” she says, and her team captured a “beautiful” blood oxygen level dependent (BOLD) signal, a proxy for brain activity. But the subsequent 19 participants were a mixed bag: All but six babies fussed or fell asleep. In those six, though, a portion of the frontal cortex had a stronger BOLD signal in response to normal speech than reversed speech, the team reported in a 2002 paper.

Saxe picked up the baton when she opened her lab in 2006, but scanning infants proved more difficult than she expected. “The first paper in my lab on that came out in 2017,” she says. “That is maybe an honest signal of how hard it was to get this to happen, for all kinds of reasons.”

In addition to trying to keep the babies awake and engaged, Saxe says she worried about how to protect her tiny participants’ hearing and minimize their head motion. In adult studies, participants wear earplugs and covers and, unlike babies, can alert researchers if their hearing protection slips out. Adults can also wear a head coil that gently presses on their skull to reduce motion, but infants, who have soft skulls, cannot.

Saxe and her team tinkered with solutions until 2013 when a firm deadline appeared: Saxe became pregnant. “There is no way I’m missing the chance to scan this baby’s brain,” she recalls thinking.

Her son was born in September of that year; his first fMRI scan took place a month later. “I spent a lot of my maternity leave inside an MRI machine, singing ‘The Wheels on the Bus’ over and over and over again, trying to keep the baby calm and happy,” Saxe says. They used a quiet scanning sequence to lower the risk of hearing damage and a custom infant-sized head coil designed to improve the signal-to-noise ratio.

The first scan that actually worked took place in January 2014, when Saxe’s son was 4 months old. The team captured activity in his visual cortex as he watched short movie clips. Over the next year, they successfully scanned eight more babies between ages 4 and 6 months, three of whom were the children of lab members. A participant’s parent or another lab member lay in the scanner during every scan to check that the infant stayed awake. In 2017, they reported that the infant visual cortex is organized in roughly the same fashion as in adults and contains regions that respond to faces and scenes.

A

round the same time, but before Saxe’s data were published, Turk-Browne and Cameron Ellis, his graduate student at the time, started their own awake infant fMRI program. Instead of collecting data for a single experiment at a time, they decided to design several tasks and switch between them based on what an infant found most appealing. “Rather than putting all of our eggs in one question basket, let’s have multiple baskets and multiple questions and pursue them in parallel rather than sequentially,” says Ellis, who is now assistant professor of psychology at Stanford. They chose four baskets: statistical learning, attention, visual processing and infantile amnesia.

This approach was a “very bright, brilliant idea,” says Dehaene-Lambertz, who is not involved in their work. “Adults, you can make them do boring stuff,” but not so with babies—switching tasks increased the odds that the team could collect usable data for at least one experiment.

Tailoring each imaging session to an individual infant is something that Turk-Browne and his team have stuck with. A participant’s parent joins Turk-Browne in the scanner room—but not in the scanner itself—and lets him know when their child needs a break, snack or diaper change. Infants can also bring a favorite stuffed animal or blanket into the scanner with them. “I’ve held hands with probably 100 babies during these scans,” Turk-Browne says. During one scan in January, he sang songs and played peek-a-boo to soothe one fussy participant in between runs and twice lunged partway into the scanner to return a spit-out pacifier.

“In a way, it’s terrible science because it’s not repeatable. I don’t think I’ve ever had two scans be the same. Which is the opposite of what you normally want—you want to have a protocol that you follow to the letter for everybody so there’s no bias and no confounds,” Turk-Browne says. “But that’s the reality of working with this population: being adaptable and working with each individual kid and family.”

And it works: Turk-Browne’s group collects an average of one usable experiment per imaging session, they reported in a preprint last month. Even though not every scan yields data, “on average, we know that this research program will bear fruit,” Ellis says. In his own lab at Stanford University, Ellis says he also gleans the same average of one usable experiment per scanning session.

“There’s a lot of talk about how impossible it is. And I think we’re showing that it’s not completely impossible,” says Tristan Yates, a postdoctoral research scientist in Nim Tottenham’s lab at Columbia University and former graduate student in Turk-Browne’s lab.

A

fter Turk-Browne and Ellis discovered that—contrary to what was commonly thought—the hippocampus shows signs of activity as early as 3 months old, they started to wonder if other assumptions about the brain region were true.

One computational model suggests that the hippocampus contains separate pathways for episodic memory and statistical learning. And the learning pathway matures earlier than the memory one in nonhuman primates, a 2013 paper showed. This staggered development could explain why babies are good statistical learners yet don’t form memories that can be accessed in adulthood, Turk-Browne says. Without neural evidence, however, it was impossible to pinpoint which part of the infant memory process fails: encoding, storage, retrieval or something else entirely.

So, Yates and Turk-Browne designed an fMRI experiment to fill this gap. They presented a series of images for two seconds each—on top of a moving psychedelic pattern just to hold the infants’ attention—and then showed an old and new image together and tracked which one the infant looked at the most. The infants would likely spend more time looking at an image if they remembered it, the pair predicted.

During memory encoding, the infants had a stronger BOLD signal in their hippocampus while looking at the images they later recognized than for the ones they appeared to forget; the infants with the greatest preference for familiar images showed the strongest difference in signal, the team reported in today’s Science paper. This difference is more pronounced in infants between ages 12 and 24 months than in infants between ages 4 and 9 months.

“It seems that the hippocampus is capable of encoding new individual memories,” Yates says, so potentially something “further down the line” is going awry. This idea jibes with findings in rodents: Stimulating neurons tagged while a mouse encodes a memory in infancy reactivates it in adulthood. In other words, the memory is there; it’s just not typically accessible.

In his ongoing projects, Turk-Browne is exploring how the complexity of representations in the hippocampus changes over the course of development, and when and how retrieval of autobiographical memories encoded during infancy breaks down. “You can’t get that level of granularity of ‘What are the neural representations of an experience?’ with any other method in infants right now,” he says.

And fMRI in awake infants is poised to address other “really big questions that, at the same time, are kind of low-hanging fruit,” says Ellis, who is using the technique to study attention and language. “You can do a relatively simple task and answer something that’s been a mystery for a long time.”

Outside the memory domain, Saxe says she is analyzing data from projects on language lateralization in toddlers and scene-processing in infants; Rhodri Cusack, professor of cognitive neuroscience at Trinity College Dublin, says his team finished scanning 134 2-month-olds last year for a study on the development of the visual system.

A “first generation of young investigators” is also joining the charge, says Vlad Ayzenberg, incoming assistant professor of psychology and neuroscience at Temple University, referring not only to himself but also to Ellis and Yates. Yates created a toddler-scanning program in Tottenham’s lab, which previously imaged only children, and she plans to study the relationship between caregiver-infant bonds and cognition. When Ayzenberg opens his lab later this year, he plans to use fMRI in awake infants to study how cortical and subcortical areas contribute to cognitive abilities at different points in development.

A decade ago, it would have been too risky for a junior faculty member to take on an awake fMRI project, Ayzenberg says. “Now, I think there’s enough evidence that this can work, and enough people have done it successfully, that it doesn’t feel as insane of a prospect.” It’s possible, he says, but “it’ll still be really hard.”