An AI just learned its first words with the help of a camera-strapped baby

Artificial intelligence researchers used a baby to build its early language dataset, per a report.
Artificial intelligence researchers used a baby to build its early language dataset, per a report. | Adobe Stock

Language acquisition is a topic that many have tried to understand. How does a baby learn its first words? What goes on in their minds?

According to a new study from Science, children start learning their first words starting at six months by connecting what they see with words. The study wanted to find out if children learned words with specific objects or a visual concept.

Why a baby wore a camera helmet

According to the MIT Technology Review, the research team behind the study wondered if AI had the capabilities to learn language like a baby. They just had to create a dataset the AI could use.

That’s where baby Sam comes in. Using a helmet camera, scientists recorded Sam’s everyday life from the time he was 6 months old until he was 24 months old, giving them 61 hours of footage to use.

Now it was time to train the AI. According to Nature, scientists used a neural network, which is a form of AI that is meant to replicate the brain’s structure and processes, and exposed it to frames from Sam’s videos. These frames included 250,000 words and images that Sam had recorded.

To train the AI, scientists used a technique called contrastive learning, per Nature, where the AI must learn which images and words went together and which of them didn’t, just like how Sam was learning.

According to MIT Technology Review, some psychologists believe that children are born with an innate ability to understand language as they grow. But this study suggests that there’s a probability that children can learn a language by the little experiences they have in their daily lives.

How well did AI do?

Nature says that AI exceeded scientists’ expectations, successfully connecting images with their correct words 62% of the time, much higher than the 25% success rate scientists were expecting it to have by only chance.

The researchers behind the study believe this will help others figure out more of the processes behind language acquisition. Psychologist Jess Sullivan says that a part of language learning involves matching the right images with the right words, per MIT Technology Review.

This study portrays the possibilities of a new future for AI. According to Popular Science, this study can help AI learn more like a human when it comes to word acquisitions, instead of pulling words from a massive dataset on hand.

Current AI systems are still considered “brittle and lack common sense,” says Howard Shrobe in an article of MIT Technology Review, who works at the U.S. Defense Advanced Research Projects Agency.