Princeton researchers train an AI model to learn language the way children do

The process by which children go from unspeaking newborns to babbling toddlers who can speak in complete sentences in just a few years is largely opaque to scientists. In the first study of its kind, researchers have followed the development of 17 babies during their first three years, recording every sound they heard and spoke in their homes. In a paper being published in the journal Science on November 5, Princeton researchers identify key tenets of how children learn to speak. The scientists also detail how they trained an AI model to learn language the way children do, opening the door to better understanding learning impairments and child development at the individual level. 

"For more than a century, psychology has largely studied development from the outside. We observe behavior and try to infer the mechanisms that produced it,” said Uri Hasson, professor of psychology and neuroscience at Princeton and principal investigator on the study. "We wanted to turn the usual logic of developmental research around. Instead of observing a child for an hour in the laboratory, we asked whether we could build a model capable of learning and immerse it in the child's actual experience, day after day.” 

The work relied on access to 15 NJ households with gracious and curious families, representing a wide variety of income levels and sociocultural norms. There were 17 babies total and this work analyzed the learning for 8 of the children. Throughout the roughly 1,000 days, the researchers collected and sorted through 1.2 million hours of data from recording devices in the homes. Next, they used the data to build an artificial intelligence model that mirrors how children absorb words and sounds, feeding it a specific amount of language at a time and adjusting the ‘replay,’ the equivalent to sleep for a baby, when the model is again exposed to data from previous days. After tinkering with certain parameters, like the amount of repetitions during replay and the processing rate, the result were neural networks with word comprehension trajectories that matched each child, accounting for individual differences - something the fields of psychology have largely not been able to do, instead relying on averages. 

 ‘Children grow up in very different language environments. The sounds they hear and the amount of speech they encounter can vary significantly from one household to another. Yet children still learn to understand language. We wanted to see whether our model could do the same. When we trained it separately on language input from different households, we found that it learned to comprehend language in each case, despite these differences in exposure,’ said co-author Hadas Raviv, a computational cognitive scientist at Princeton who led the analysis infrastructure development and computational modeling. 

The team’s language-learning model ultimately needed what a child gets: a whole day's worth of speech, day after day, and a night to replay it, not mass amounts of written information and shuffled learning, as typical neural networks are developed.

The researchers hope to use the model to study the effects of various circumstances on language development, for example, the effect of more or less sleep, or exposure to different speech types, such as speech from the television.The researchers see the work as opening the door to be able to better diagnose learning impairments by understanding the way a ‘typical’ brain functions and learns language. 

Raviv said, “Unlike much of modern AI, which aims for superhuman performance, we’re trying to understand how humans learn and how that learning gives rise to intelligence. These models allow us to ask what it takes to learn language at multiple levels from everyday experience.”

The researchers believe the work will be useful to developmental psychologists, clinicians, the AI and machine learning community.