NYU discovered that babies are born with language skills and may even be better at learning than ChatGPT

Human babies are even better at learning than the best large language models. In order to be able to write competent English, ChatGPT must be trained on massive datasets containing millions, if not trillions, of words.

And children are only exposed to a small part of this data, but by the age of three they are able to communicate in quite complex ways.

A team of researchers at New York University wondered if AI could learn like a baby. What could an AI model do if it were given a much smaller data set, using the sights and sounds experienced by a child who learns to speak?

As it turns out, there is a lot that can be done. The AI model successfully matches the words with the objects they represent.

Brenden Lake, a computational cognitive scientist at New York University and author of the study, said: "Even in this brief experience of childhood, there is enough data to suggest that it can do real word learning." ”

The paper, published in Science, not only provides insights into how babies learn, but may also lead to better AI models.

(来源:WAI KEEN VONG)

For the experiment, the researchers used a 61-hour video shot from a helmet camera worn by a child living near Adelaide, Australia.

The child, Sam, wore the camera on and off for a year and a half from the age of six months until his second birthday.

The camera captured what Sam saw and noticed during about 1% of his waking hours. It chronicles Sam's two cats, his parents, his crib and toys, his house, his meals, and more.

"This dataset is completely unique," Lake said. "This is our best window into the information available to a single child. ”

To train the model, Lake and his colleagues used 600,000 video frames and paired them with phrases spoken by Sam's parents or others in the room while the images were taken, for a total of 37,500 "utterances."

Sometimes, words and objects match, and sometimes they don't. For example, in one still, Sam looks at a shape classifier and one parent says, "You like ropes." ”

In another photo, the adult's hands are covered with some bricks, and the parent says, "You want these blocks too." ”

The team gave the model two clues. When objects and words appear together, it indicates that they may be related. But when an object and a word don't appear at the same time, it's a sign that they may not match.

Wai Keen Vong, a computational cognitive scientist at New York University and author of the study, said, "So we have this sometimes matched, sometimes not-matched situation in the model. ”

"Hopefully, there are enough examples in the data that when parents say the word 'ball,' the child sees the ball," he said. ”

Matching words to the objects they represent may seem like a simple task, but it's not. To give you an idea of the difficulty of the problem, imagine the living room of a family with young children.

The living room has all the usual furniture, but also the children's clutter. The floor was littered with toys. Crayons are scattered on the coffee table. Snack cups are placed on the windowsill, and laundry is placed on the chairs.

If a toddler hears the word "ball," it probably refers to the ball. But it can also refer to any other toy, or a couch, or a pair of pants, or the shape, or color, or time of day. "Any word has an infinite number of possible meanings," Lake said. ”

The question is so tricky that some developmental psychologists believe that children must be born with an understanding of how language works in order to learn it so quickly.

Jess Sullivan, a developmental psychologist at Skidmore University in the United States, was part of the team that collected data from Sam's helmet camera but was not involved in the new study.

Still, she said the study showed that even without this innate ability, it is possible to learn parts of language from a very small set of experiences. "For me, it's really upended my worldview," she said. ”

But Sullivan points out that being able to match words to the objects they represent, while a difficult learning problem, is only part of what makes up language. There are also rules that govern the combination of words.

Your dog may know the words "ball" or "walk", but that doesn't mean it can understand English. Therefore, the baby's innate language ability may not be limited to vocabulary.

This can affect the way they move around the world, including the things they focus on, or how they react to language. "I don't think this research would have been successful if the babies hadn't created a dataset for neural network learning," she said. ”

The next step for Lake and his colleagues is to work to figure out ways to bring model learning closer to early childhood language learning. "There's more work to be done to get the model fully capable of a two-year-old," he said. ”

This may mean providing more data. Lake's child is now 18 months old, and she wears a helmet camera for a few hours a week, while she is also part of the next batch of children to provide this data.

Alternatively, the model needs to pay attention to the gaze of the parents, or to have a certain perception of the solidity of the object, which the child can intuitively grasp. Creating models that learn more like children will help researchers better understand human learning and development.

If an AI model can grasp some of the ways humans learn language, the learning efficiency may be much higher.

They may behave more like humans than "a clumsy statistical engine for pattern matching" as large language models like ChatGPT, as linguist Noam Chomsky and his colleagues once described.

Howard Shrrobe manages the project at the U.S. government's Defense Advanced Research Projects Agency, which helped fund Lake's team. "AI systems are still fragile and lack common sense," he said. ”

However, an AI that learns like a child may be able to understand meaning, respond to new situations, and learn from new experiences. Our goal is to bring AI one step closer to human intelligence.

Support: Righteousness

Operation/Typesetting: He Chenlong