The AI model was trained on frames from the video and words spoken to the baby, transcribed from the recording. It used a technique called contrastive learning to learn which images and text tend to go together. The model was able to correctly identify objects 62% of the time, much better than the 25% expected by chance. However, the study's reliance on data from a single child raises questions about the generalizability of its findings. Despite this, the findings challenge theories that language acquisition requires special mechanisms, suggesting instead that a lot can be learned through forming associations between different sensory sources.
Key takeaways:
- An artificial intelligence (AI) model has been trained to recognize words like 'crib' and 'ball' by studying headcam recordings of a baby's life, suggesting that AI can help us understand how humans learn.
- The AI learned by building associations between the images and words it saw together, challenging some cognitive-science theories that babies need some innate knowledge about how language works.
- The AI was tested by matching a word with one of four candidate images, successfully classifying the object 62% of the time, which is much better than the 25% expected by chance.
- The study's findings challenge scientists who claim that language acquisition cannot happen through general learning processes, and suggest potential for further refinements to make the model more aligned with the complexities of human learning.