In an article published Thursday in the journal ScienceNew York University researchers report that an artificial intelligence that takes into account only a small fraction of a child’s fragmented experiences can learn that there is something called a crib, a staircase, or a puzzle, and begin to discern the order in the pixels and match those words correctly. with pictures.
The tool used by the researchers is not an artificial intelligence that learns like a child. But research shows that artificial intelligence can pick up some basic elements of language from the sensory input of a single child’s experience, even if it has no pre-existing knowledge of grammar or other social abilities. It’s part of a larger quest to eventually create an artificial intelligence that mimics a baby’s mind, a sacred cognitive science that could help researchers understand our own development and lead to AI that can teach humans new skills more intuitively.
Chatbots, also known as “large language models,” have demonstrated that AI trained on large volumes of text can produce a communicative conversational partner with a dazzling mastery of the language. But many cognitive scientists argue that this verbal prowess is far from true human thinking.
Babies are the opposite of chatbots, they learn words not by rapidly digesting all the text in the world, but by being in the world itself, through sensory input and play.
“We estimate that it would take 100,000 years of listening to a child speak before the training sets for chatbots reach word count,” said Brenden Lake, a computational cognitive scientist at NYU who led the study. “I was also suspicious of them [chatbot] models would shed much light on human learning and development.”
There are linguists, philosophers, cognitive scientists, and increasingly AI developers how people learn language was confusing to everyone.
For years, scientists have been trying to understand how children’s minds are formed through carefully controlled experiments. There are many toys or mannequins that allow researchers to study different cognitive skills when they come online. They showed it 16 month old babies can apply statistical reasoning to determine whether a noisemaker is impaired and 5-year-old infants. moons know that an object still exists even if they can’t see it, this is called primary development object persistence.
In addition, some individual infants were followed closely over time. In 2005, Massachusetts Institute of Technology researcher Deb Roy installed cameras in every room of her home and recorded her son’s language development. large database which records the acquisition and evolution of words. This study suggested that it was not the number of times a word was repeated that predicted whether Roy’s son learned it early, but that it was said at home in an unusual place, at a surprising time, or in a different language context.
The innovative use of head cameras has allowed researchers to take a closer look at early childhood.
Several families have contributed since 2013 SAYCam database, a collection of audiovisual recordings of individual infants and toddlers during the critical period of cognitive development between 6 and 32 months. The families of the babies, known only by name, spend about two hours a week wearing taped cameras on their children’s heads.
Scholars can apply for access to data that provides a unique window into each child’s world over time and is intended to be a resource for researchers in a variety of fields.
Sam, whose identity is withheld, is now 11 years old. But recordings of his early life in Australia provided Lake and his colleagues with 600,000 video frames combined with 37,500 transcribed words of training data for the AI project.
They trained relatively simple neural networks on data collected from when Sam was between 6 months and 2 years old. They found that the AI learned to match key nouns and images with similar accuracy to AI trained on 400 million images from the internet with captions.
The results add to, but do not resolve, a long-running debate in science about the basic cognitive skills people must deploy in their brains to learn language.
There are various theories about how people learn language. High-profile linguist Noam Chomsky proposed the idea of an internal, innate language ability. Other experts believe that we need social or inductive reasoning skills to develop language.
New research suggests that some language learning can occur in the absence of specific cognitive mechanisms. Relatively simple associative learning – looking at a ball, hearing “ball” – can teach an AI to make a match when it comes to simple names and pictures.
“There’s nothing that gives the network clues about language or how language is structured,” said study co-author Wai Keen Wong, a research scientist at NYU.
The researchers did not have comparable data on how a 2-year-old child would perform on tasks the AI faced, but they said the AI’s abilities fell short of those of a young child. For example, they were able to track where the AI focused when prompted with different words, and found that while it was spot on for some words, such as “car” or “ball,” “when prompted, it looked in the wrong area. cat.”
“I want to learn the minimum ingredients needed to create a model that can learn more like a child — that’s one step,” Lake said.
Basics of language
The AI gathered a vocabulary of objects from exposure to 1 percent of Sam’s waking hours — 61 hours of footage collected over a year and a half. Of interest to outside scientists about the study was both how far AI has progressed based on it, and how far it still has to go to replicate human learning.
“It’s really important and novel to apply these methods to this kind of data source, which is data from both visual and auditory experience of a child,” said Joshua Tenenbaum, who is not involved in computational cognitive science at MIT. work
“I would add that there are still some things that are difficult to deduce from paper—it’s less clear to us how children actually learn words.”
Michael Tomasello, a developmental and comparative psychologist at Duke University, said the artificial intelligence model could mirror how a dog or parrot might learn words. Experiments show that some dogs can learn more than 100 words for common objects or stuffed animals.
But, he noted, it’s not clear how this AI can receive sensory input and collect verbs, prepositions or social expressions.
“It can learn that a repeating visual pattern is a ‘doll.’ But how does he learn that the same object is also a “toy”? How does he learn ‘this’ or ‘that’ or ‘that’ or ‘thing’? Tomasello wrote in an email.
He noted that the artificial intelligence model, which was developed based on the child’s experience, was able to identify what can be seen, and this is only a small part of the language that children hear and learn. He proposed an alternative model instead of a simple one an AI that associates images with sounds must infer communicative intent in order to learn a language.
Lake is starting to train its AI models on video instead of still footage to see if they can successfully expand their vocabulary to verbs and abstract words. There will soon be an additional stream of data to work with as Lake gathers information from her young daughter.
But he acknowledged that AI learning techniques go beyond children’s learning, even for simple words. The AI was really great at learning to identify sand, for example, but had trouble with the hands, which means its progress probably doesn’t reflect most children’s perception of their environment.
“‘Sand’ was too easy, ‘hand’ too hard,” Lake said. “And the model doesn’t know that milk and pears taste good.”