👶 Why Do Scientists Want to Give Artificial Intelligence a Childhood?
The GPT-4 model has been trained on approximately 13 trillion tokens (around 10 trillion words)—thousands of times more than any human could read or hear in a lifetime. Meanwhile, by the age of two, a child knows around 300 words. And for this learning process, children don't need datasets with hundreds of thousands of examples, months of continuous training, or millions of dollars—it happens naturally in everyday life.
Unlocking the mystery of how children perceive and learn about the world could help solve one of AI's biggest challenges: teaching models to truly understand physical reality.
🍼 How to See the World Through the Eyes of a Child?
Last year, scientists from New York University conducted an experiment where they trained an AI algorithm to learn language in the same way a typical child does. An infant named Sam helped them. Over 1.5 years (from six months to two years old), Sam wore a helmet with a camera several times a week. The camera recorded everything Sam saw, heard, or said. From the hundreds of hours of footage collected, researchers selected 61 hours of video. The camera captured around 250,000 different "words."
Next, they tasked a "naive" neural network with no prior knowledge of the real world to analyze the audio and video recordings to find connections between words and objects.
In 62% of cases, the AI model correctly identified an object in the video based on a word prompt. For example, when given the word "cat," it identified Sam's pet cat in the footage. This performance matched an algorithm trained on 400 million pairs of images and text. In 80% of cases, the model identified learned objects in images it had never seen before.
"We've shown for the first time that a neural network trained on realistic developmental input data from one child can learn to associate words with their visual counterparts," explains Dr. Wai Keen Wong, the study's author.
⛓️ How Can This Help Us?
The experiment demonstrates that AI models don't necessarily need massive datasets for initial training, as is currently the case with advanced algorithms.
Improving this associative approach—teaching AI not only to recognize objects but also actions (verbs) and intonations—could lead to a new type of AI algorithm capable of understanding the real world using just a camera and microphone. In the future, this could pave the way for empathetic robots based on such technology, like those seen in the animated film The Wild Robot.
More on the topic:
▶️ Why Are Scientists Teaching AI to Understand Emotions?
