If You Think An AI Model is Intelligent, Meet a 1-Year-Old
If you think an artificial intelligence model running on thousands of state-of-the-art computer chips is intelligent, allow me to introduce you to the concept of a 1-year-old.
OK, babies may not be able to write computer programs, solve advanced math problems, or discuss philosophical ideas. But unlike today’s AI models, which use an ocean’s worth of training data and as much energy as a small country, babies learn to understand the world with astonishing efficiency. They identify new objects after seeing them once or twice and learn through fleeting observation and physical interaction.
When it comes to improving AI, babies—and the architecture of their brains—could provide crucial insights. Building a more baby-like version of AI could make frontier models more cost-effective and less energy-intensive, and it could also be valuable if AI-powered robots are to learn their surroundings in a more natural way.
The EgoBabyVLM Challenge
To explore this bold new frontier, researchers from Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure have developed a new test that highlights babies’ learning abilities and pushes AI researchers to develop algorithms that suit them.
The EgoBabyVLM Challenge assesses how well visual language models (VLMs), which learn from both text and images, can understand the world as a baby sees it. It requires a model to describe the world after recording about a thousand hours of video footage collected by cameras attached to the heads of infants and young children. (Yes, really.)
It turns out that the state-of-the-art models fail miserably when fed this realistic and chaotic footage, suggesting there may be something else about the design of the baby’s brain that allows it to learn so quickly from so little information.
Beyond Curated Data Sets
Instead of curated data sets, babies learn from a kaleidoscopic view of things: parents talk about objects that are no longer visible, point out things with their gaze or a gesture, or discuss events from the past or the future rather than what is currently happening. Babies learn not just through language but also through a rich multimodal and tactile experience, says Michael Frank, a cognitive scientist at Stanford University who specializes in language learning and helped develop EgoBabyVLM.
The test shows: “It is clear that there is more,” [than just language] “That’s necessary,” says Frank.
Language Learning
EgoBabyVLM is just the latest example of how scientists are using AI to study human intelligence. A challenge called BabyLM, launched in 2023, asked AI models to learn the syntax of language using about the same amount of data as a 10-year-old ingests—tens of millions of words, compared to trillions for AI models. Remarkably, it turns out that transformer-based AI models—which process language by paying attention to the relationship between words in different sentences—can do this quite well, a finding that challenges Noam Chomsky’s ideas about how syntax might be hard-wired into the human brain.
Ryan Cotterell, a linguist at ETH Zurich who first developed BabyLM, says the situation is different when it comes to understanding the physical world. “There will not be a large corpus of human interactions – there is no Internet of human interactions,” he says.
Joshua Tenenbaum, a cognitive scientist at the Massachusetts Institute of Technology, notes that BabyLM has shown that models do not develop “common sense” about the physical world, social dynamics, or theory of mind.
“Transformers are very good at finding patterns in data,” says Tenenbaum. “But it seems that pure pattern learning systems are not able to take the kind of data that a baby or a child receives and learn all the things that they do.”
For more detailed insights, you can read the full article Here.
“`
