Can artificial intelligence (AI) learn language like a baby? According to a recent study conducted by researchers at New York University (NYU), the answer is yes. The team equipped a baby with a head-mounted camera and recorded videos from the age of six months to their second birthday. The researchers then used this footage to train an AI system or neural network, which is a computational model capable of learning patterns from input data. Despite only capturing around one percent of the child’s waking hours, this limited data was enough for the AI system to learn language.
Published in the journal Science, the study revealed that the neural network trained on the child’s developmental input was able to link words to their visual counterparts. This breakthrough demonstrates how recent algorithmic advancements, combined with a single child’s naturalistic experiences, have the potential to reshape our understanding of early language acquisition.
Typically, top-tier AI systems are trained using massive text datasets containing trillions of words, whereas children are exposed to only millions of words annually. By using AI models to study language learning, researchers hope to shed light on the components necessary for word acquisition in children, such as language-specific biases, innate knowledge, or associative learning.
The researchers had approximately 60 hours of footage capturing the child’s daily activities, with roughly 250,000 words being communicated. These words were associated with video frames that depicted what the child saw when those words were spoken during activities like mealtimes, reading books, or playtime.
To train the model, the researchers utilized two modules: one for video frames and another for transcribed speech directed at the child. These modules were combined and trained using contrastive learning, a type of machine learning that helps the model understand the associations between visual and linguistic cues.
The next step for the researchers was to test the model, named the Child’s View for Contrastive Learning model (CVCL), using the same methods employed to measure babies’ word learning. They showed the model a word and four pictures, asking it to choose the picture that matched the word. The results showed that the CVCL model successfully learned many words from the child’s daily life. Furthermore, it could also apply some words to different pictures that it hadn’t seen during training, mirroring the way children learn to associate words with various objects or scenarios.
These findings suggest that this aspect of word learning is feasible from the kind of naturalistic data that children receive while using relatively generic learning mechanisms such as those found in neural networks, stated Brenden Lake, an assistant professor at NYU and the senior author of the paper.
This groundbreaking research could have far-reaching implications for language acquisition studies and AI development. By understanding how AI models can learn language using limited, naturally occurring data, scientists can gain insights into the learning mechanisms employed by children. This research has the potential to reshape our understanding of early language and concept acquisition.
While this study represents a significant step forward, there is still much work to be done to fully comprehend the complexities of language learning in both infants and AI systems. However, with continued advancements in AI algorithms and data collection methods, researchers are optimistic about the potential for further breakthroughs in the future.
