A synthetic intelligence (AI) language model modeled after the means of children learning language has emerged. To do that, the researchers installed cameras and microphones on the heads of kids aged 6 to 25 months and recorded videos. And it was concluded that learning a language doesn’t require as much data learning as a typical large language model (LLM).

Science recently introduced that researchers at Latest York University developed an AI model that matches images and words using 61 hours of video and sound captured from the attitude of a toddler named Sam between 2013 and 2014. The name of this model is ‘Child’s View for Contrastive Learning (CVCL)’.

Most kids can understand 300 words by age 2, and this increases to over 1,000 words by age 4. Nonetheless, the power of humans to amass words isn’t fully understood, and a few scholars hypothesize that humans are born with the power essential to amass language.

Nonetheless, this experiment showed that determining word meaning from minimal data doesn’t require preprogrammed assumptions. It also suggested that the tactic of language acquisition is easier than expected, and the evaluation showed that easy information reflected in a toddler's eyes is sufficient for pattern recognition and word understanding.

“This study shows that it is feasible for machines to learn in a way much like humans,” said Wei Kinh Fong, a computational cognition researcher at Latest York University and lead writer of the paper. “We train with data, but humans live with much less information,” he said. “With the suitable variety of data, the gap between machine learning and human learning could be greatly reduced.”

“Today’s models don’t need as much input as we get to make meaningful generalizations,” said Brandon Lake, an associate professor of psychology and data science at Latest York University and lead writer of the study. “We learn words through the eyes and ears of a single child. “That is the primary time we have now shown that an AI model could be trained to do that.”

The researchers constructed a multimodal model consisting of a vision encoder and a text encoder to investigate the information containing what Sam sees and says. Through this, what Sam saw and what he said matched, however the sound of the speech was inconsistent and was mixed with the voices of his parents and surrounding people, so it seemed somewhat random.

Nevertheless, the AI ​​model accurately matched the word to the video after multiple tests. As well as, in consequence of conducting a benchmark with an AI model learned with way more data, it was revealed that it was near the accuracy of other models.

Moreover, it showed an accuracy of 62% in a multiple-choice test to find out whether a picture contained a particular object.

Particularly, it was revealed that Sam's video showed a 'zero-shot generalization' ability to discover many objects that weren’t included within the video and generalize what was learned. Regarding this, the researchers said, “We were very surprised.”

Although there have been similar studies before, this result using the vision-language model is taken into account to have opened a latest breakthrough. Jessica Sullivan, associate professor of psychology at Skidmore College, said, “I used to be one in every of the internalist theorists who said humans are born with the power to talk, but this study modified my mind. It is a very beautiful study.”

Nonetheless, this study also has clear limitations. Apart from easy pattern recognition, researchers haven’t identified other aspects that influence how children learn language. For instance, although the model recognized dozens of words, there have been still words it didn’t understand and it tended to get easily confused with certain words.

Moreover, this study focused only on recognizing nouns of physical objects. Nonetheless, it’s identified that human language acquisition is way more complex than that. It has also not been proven that AI can learn verbs or abstract concepts.

Nonetheless, this can be a step toward a deeper understanding of our minds and will ultimately help improve human education.

“AI research doesn’t should just be geared toward maximizing chatbot capabilities and company profits,” said Eva Fortrans, a computational linguistics researcher on the Millar-Quebec AI Institute. “It could also make clear long-standing questions on ourselves. “We are able to use these models in good ways in which profit science and society,” he said.

Reporter Lim Da-jun ydj@aitimes.com