Artificial intelligence is not smarter than a child, yet

If you think that artificial intelligence The model runs on thousands of high-end computers chips Smart, let me introduce you to the concept of a one-year-old.

Well, children may not be able to write computer programs, solve advanced mathematical problems, or discuss philosophical ideas. But unlike current AI models, which consume an ocean’s worth of training data and Same amount of energy as a small countryChildren learn to understand the world with amazing efficiency. They recognize new things after seeing them once or twice, and learn through casual observation and physical interaction.

When it comes to improving artificial intelligence, children – and the structure of their brains – may hold important insights. Building a more AI-like version could make parametric models less expensive and less power-hungry, and it might also be helpful for AI-powered robots to learn about their environments in a more natural way.

To explore this bold new frontier, researchers at Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure in France sophisticated A new test highlights children’s learning skills and prompts artificial intelligence researchers to design algorithms that suit them.

the EgoBabyVLM Challenge It judges the extent to which visual language models, or VLMs, which learn from text and images, understand the world as a child sees it. It requires a model to describe the world after it is taken up A thousand hours of video They were collected from cameras mounted on the heads of infants and young children. (Yes, really.)

It turns out that sophisticated models fail miserably when fed this messy, realistic snapshot, suggesting that there may be something different about the design of a child’s brain that enables them to learn so quickly from so little information.

Instead of curated sets of data, children learn from a changed view of things: parents talk about things that are no longer visible, point to things using their gaze or gestures, or discuss events from the past or future rather than whatever is happening at that time. Not only do babies learn from language, they also learn through a rich, multimodal, tactile experience, says Michael Frank, a cognitive scientist at Stanford University who specializes in language learning and who co-developed EgoBabyVLM.

The test shows that when it comes to artificial intelligence, “clearly more is needed[than just language],” Frank says.

Learn the language

EgoBabyVLM is just the latest example of how scientists are using artificial intelligence to explore human intelligence. Called challenge Beebe LMintroduced in 2023, tasked AI models with learning the syntax of a language using the same amount of data a 10-year-old would receive — tens of millions of words, compared to trillions of words for AI models. Remarkably, it turns out that transformer-based AI models – which process language by paying attention to the relationship between words across different sentences – can do this well, a finding that challenges Noam Chomsky’s ideas Regarding how syntax is installed in the human brain.

The situation is different when it comes to understanding the physical world, says Ryan Cottrell, a linguist at ETH Zurich who first developed BabyLM. “There won’t be a whole lot of human interactions, there’s no Internet of human interactions,” he says.

Joshua Tenenbaum, a cognitive scientist at MIT, points out that BabyLM has shown that models do not acquire “common sense” about the physical world, social dynamics, or theory of mind.

“Transformers are very good at finding patterns in data,” Tenenbaum says. “But it seems that purely modular learning systems are not able to take the kind of data that a child or child receives and learn all the things that he or she does.”