If children have everything to learn about life, it is well known that they teach adults, particularly their parents, many lessons. As proven by a study published in Nature on February 1, they are, in a way, also capable of training artificial intelligences. We’ll explain it to you.
Machines, through machine learning, are trained on large quantities of data which they analyze and through which they learn to make predictions. In the case of the study carried out by American researchers from New York University, the neural network used relied on videos of the daily life of a child who wore a helmet equipped with cameras, to learn to recognize objects.
I will tell you
source
Contrastive learning
Sam, the little Australian who took part in the experiment between the ages of 6 months and 2 years, recorded two hours of his life per week while he participated in various activities such as playing, reading and eating. In total, the scientists exposed their model to 250,000 images extracted from 61 hours of recording as well as the transcription of the words Sam spoke corresponding to each image. The AI was thus able, through contrastive learning, to learn which images and words went together to accumulate information allowing them to recognize objects such as a bowl or a ball.
When asked to associate a word with one of the four images presented to it, the neural network succeeded in 62% of cases. A good score that could have been achieved by a similar AI model trained on 400 million image-text pairs from another database. The scientists observed that their AI was better at spotting objects that varied little in appearance, but had difficulty learning words that could refer to various objects like “toy.” The experiment was not intended to encourage the generalization of training AI models from toddlers’ lives. Rather, it was about enlightening us on language learning to advance research in cognitive sciences.
Selected for you
2024-02-26 06:30:00
#York #University #trains #artificial #intelligence #Sam #2yearold #boy
