Process of language learning in babies is still a mystery among linguists. Some think that babies’ minds are blank and they learn through their experiences, like hearing, seeing and playing. Other scientists believe that experience is not enough and some skills are hard-wired into the babies’ brains.
AI models are mostly tuned through scrolling millions of pages, which is not anything like the process in which babies learn. Recently, a group of scientists in New York University, tuned an AI model through the viewpoint of a baby. Between the age of six to twenty-five months a toddler named Sam, wore a helmet with a camera one hour per week the audio was also recorded. The camera recorded everything Sam saw and heard it from playing with toys, playing at the park and interacting with a pet. The video stream then was used to train an AI model.
The scientists then tested the model by asking it to identify objects Sam had seen before. Choosing from 4 different options, the model chose the correct answer 62 percent of the time, which is far more than 25 percent random correct answers. The model was also able to identify the kinds of chairs and tables Sam had not seen before. The AI model has learned at least 40 different words. Of course, it was much lower than the capabilities that Sam had developed. But considering the fact that the video streams only covered 1 percent of his awake hours, the progress is outstanding.
Through this paper researchers argue that experience is the only ingredient for learning a language. This experience develops through matching words with objects. Skeptics however, doubt that AI can learn abstract words and phenomena.
Source: Economist
