Can AI Learn Like a Baby? Scientists Turn to a Single Toddler to Find Out

Saturday - February 3 2024, 01:03 UTC - 9 months ago

A team of researchers at NYU used a small data set from a toddler's helmet camera to train an AI model to match words to the objects they represent. The model was surprisingly accurate, challenging the belief that massive data sets are necessary for AI to learn. The study has implications for future AI development and has received funding from multiple organizations.

In the world of language models, ChatGPT reigns supreme. But even this top-performing AI has its limits when it comes to learning. For example, it took massive data sets containing billions of words for ChatGPT to be able to write in passable English. On the other hand, human babies are able to communicate in complex ways by the age of three with access to only a small fraction of that data. This has led scientists to ask: can AI learn like a baby? .

A research team at New York University sought to answer this question by giving an AI model a much smaller data set—just 61 hours of video recorded by a helmet camera worn by a child named Sam. The camera captured the sights and sounds that Sam experienced from the ages of six months to two years, providing a unique window into a single child's learning process.

Through analysis of 600,000 video frames paired with the corresponding spoken phrases, the AI model was trained to match words to the objects they represent. This involved making connections between words and objects that appeared together, and pulling apart words and objects that did not match. While this may seem like a simple task, the complexity lies in the fact that a single word can refer to multiple objects in a toddler's cluttered environment.

Surprisingly, the AI model was able to accurately match words to objects, even with the small data set available. This challenges the belief that AI models require massive amounts of data to be successful, and sheds light on the amazing learning abilities of human babies. The study, published in the journal Science, not only provides insights into how children learn, but also has practical applications for the development of future AI models. With funding from multiple organizations, the research team hopes to continue their groundbreaking work in this field.