Researchers explore how children learn language to enhance AI models, with potential benefits for both fields and applications in education and therapy.
Language models, the neural networks driving generative AI, are trained extensively on data but often function as advanced autocomplete systems, predicting words with high accuracy. In contrast, children learn language by absorbing information from their environment, developing a comprehensive understanding of the world.
Brenden Lake, a psychologist at New York State University, believes studying children’s language acquisition could enhance AI models and aid children with learning difficulties. Lake’s research involves training an AI model using stimuli similar to those experienced by children learning their first words. This includes using cameras to gather video and audio data, including from Lake’s own daughter.
The model aims to associate video footage from a child’s viewpoint with words spoken by caregivers, learning to match words to observed objects, even when the child isn’t directly looking at them. This experiment investigates whether a model can learn object identification with the limited data available to a child.
Lake’s team previously trained a neural network on 61 hours of child video footage, enabling it to connect words and phrases with experiences captured in the videos. The model could recall images and generalize object names in unseen images, albeit with varying accuracy.
Future research aims to enhance the model’s ability to learn verbs, action words, and abstract concepts using advanced video modeling techniques. Success could lead to more efficient AI models and insights into learning and development, with potential applications in speech therapy and understanding developmental disorders.