Thank you for visiting nature. com. Your browser version has limited support for CSS, so we recommend using a more up-to-date browser or disabling compatibility mode in Internet Explorer for the best browsing experience. Currently, we have disabled styles and JavaScript to ensure ongoing support. A groundbreaking study has revealed that a neural network, trained on the recorded experiences of a single infant, has successfully learned to recognize objects, shedding light on human learning. The AI model studied headcam recordings of a baby's life, learning to identify words like 'crib' and 'ball'. This research provides valuable insights into how humans acquire knowledge, according to Wai Keen Vong, co-author of the study and an AI researcher at New York University. Previous language-learning models, such as ChatGPT, were trained on massive amounts of data, whereas this study focused on the real-world experiences of an infant, stating, "We don't get given the internet when we're born. " Heather Bortfeld, a cognitive scientist at the University of California, Merced, describes the study as a fascinating approach to understanding early language development in children. To collect data from the infant's perspective, researchers used 61 hours of recordings from a camera mounted on a helmet worn by a baby boy named Sam. Sam, residing near Adelaide in Australia, wore the camera for one hour twice a week, capturing around 1% of his waking hours, from six months to approximately two years of age. By training their neural network— an AI model inspired by the brain structure— on frames from the video and transcribed words spoken to Sam, the researchers exposed the model to 250, 000 words and corresponding images during various activities such as play, reading, and eating. Through the use of contrastive learning, the model learned to associate images and text, enabling it to predict which images certain words referred to, like 'ball' and 'bowl'.

The AI's performance was evaluated by having it match a word with one of four images, a common test in assessing language skills in children. Impressively, it classified the object correctly 62% of the time, outperforming the 25% chance level and performing similarly to another AI model trained on 400 million image-text pairs outside of this dataset. Additionally, the model was capable of correctly identifying previously unseen examples of certain words such as 'apple' and 'dog', a skill typically mastered by humans. On average, it successfully accomplished this task 35% of the time. The AI excelled at identifying objects that appeared frequently in the training data and had less variation in appearance, while it struggled more with words like 'toy' that can refer to a variety of items. While the study's reliance on data from a single child raises questions about generalizability due to the various experiences and environments children have, Bortfeld emphasizes that associations between different sensory inputs can be formed even in the earliest days of an infant's life. The findings challenge experts like Noam Chomsky, who argue that language acquisition requires special mechanisms due to its complexity and the scarcity of information input. Bortfeld states, "These are among the strongest data I've seen showing that such 'special' mechanisms are not necessary. " However, the AI's experience is limited compared to real-world language learning as it only trained on still images and written text, lacking the interactions inherent to a baby's life. Vong points out that the AI struggled to learn the word 'hand', a concept infants usually learn early on, emphasizing the need for additional components in their model. Anirudh Goyal, a machine learning scientist at the University of Montreal, sees immense potential for refining the model to align it with the complexities of human learning, opening exciting pathways for advancements in cognitive sciences.