For decades, linguists have argued about how children learn language. Some people believe that children are born as “blank slates” who learn language simply through experience – listening, seeing and playing with the world. Others argue that experience is not enough and that children’s brains must be programmed to facilitate language acquisition.
AI models like GPT-4 have done little to settle the debate. The way these models learn language – by sifting through piles of text data from millions of web pages – is very different from the experience of babbling babies.
A team of researchers from Modern York University investigated this issue by training an artificial intelligence model based on the experiences of a single infant. From ages 6 to 25 months, the toddler, named Sam, wore a head-mounted brace for an hour a week, which accounted for about 1% of his waking hours. The camera recorded everything he saw and heard as he played with toys, spent time at the park and interacted with his cats. The recordings and transcribed audio were fed into the artificial intelligence, which knew that images and words appearing at the same time were related, but was otherwise left to make sense of the mess of color and speech that Sam had experienced.
Despite narrow training data, the AI was able to select objects and learn matching words. The researchers tested the model by asking it to identify objects that Sam had seen before, such as a chair from his house or one of his toy balls. Given a list of four options, the model selected the correct word 62% of the time, well above the chance level of 25%. To the researchers’ surprise, the model was also able to identify chairs and balls that Sam had never seen before. The AI learned at least 40 different words, but by the end of the experiment it was far from matching Sam’s vocabulary and language skills.
Scientists, whose work was recently published in the journal Science, argue that learning from experience is enough to match words to objects. However, skeptics doubt whether artificial intelligence will be able to learn abstract nouns and verbs and question the similarity of the learning processes. The mystery of language acquisition still lives on.
© 2024, The Economist Newspaper Confined. All rights reserved.
From The Economist, published under license. Original content can be found at www.economist.com
Published: April 8, 2024, 6:00 PM EST
