Teaching AI to Learn Language Like Babies
- Chase Noteware
- Apr 22, 2024
- 2 min read
My oldest sister went to MIT and she recently sent me this article from MIT’s Technology Review magazine:
It's about a group of researchers at NYU who were curious about how ChatGPT learned English compared to infants and small children (ages 0-3). They noted that ChatGPT had to be trained using datasets containing millions–if not trillions–of words, but that developing children can learn words and communicate in sophisticated ways using only a fraction of the number of words. This got the scientists thinking about the underlying mechanics of human learning and how it could be applied to AI.
The research team gathered 61 hours of video recordings from a helmet camera worn by an Australian child named Sam. Sam wore the camera on and off for a year and a half, from the time he was 6 months old until just after he turned 2. The recordings provided data on his daily interactions with things like toys, family members, and pets, and provided a window into his experiential learning processes.
The researchers then used 600,000 images and 37,500 spoken phrases from Sam’s helmet camera recordings to train an AI model to associate words with objects, mimicking the way small children learn language through exposure to their environment. The study revealed that even with a limited linguistic dataset, the AI could successfully learn words by matching them to graphic images.
Based on this outcome, the team at NYU believes that AI models that learn like children may be able to understand meaning, respond to new situations, and learn from new experiences in a more efficient and human way. They plan to refine their model to mimic early language learning in children more closely. This includes considering factors such as the parents' gaze and the physical properties of objects, which play a crucial role in how children learn about their environment.
Comentarios