There is no competition between babies and AI – 02/09/2024 – Tech

Facing a screaming baby and throwing food on the floor, stressed parents may be surprised to hear that their beloved creation is probably the smartest learner in the universe.

But some computer scientists have long recognized this reality and are now trying to mimic babies’ extraordinary processing powers to develop artificial intelligence (AI) models.

In some ways, our latest technological creations seem almost magical in their capabilities, but when it comes to autonomous learning, they are as dumb as a diaper.

Could these technologies be trained to learn like infant processing units, exploiting partial, messy, real-world data?

A team of researchers from NYU (New York University) has been trying to do just that and this month published their findings in the journal Science.

The experiment collected data from a lightweight camera attached to the head of an Adelaide baby called Sam, recording 61 hours of his life from 6 to 25 months.

The video stream, including scrambled images and sounds of parents, cats, play and toys, was then processed into 600,000 video frames and 37,500 “utterances” transcribed and fed into a neural network.

The challenge was to combine what Sam saw during approximately 1% of his waking hours with the sounds he heard, to create a multimodal AI model.

But how does a baby understand that the word “ball” relates to very different round, bouncy, multicolored objects?

Cognitive scientists are divided on the explanation, but they all agree that babies are incredibly skilled learners, generalizing from limited input.

Between six and nine months, babies begin to connect words to pictures. Before the age of two, they have already learned an average of 300 words, mainly nouns.

Until now, attempts to build multimodal AI models that can combine text, images, audio and video have mainly relied on applying enormous computational power to vast amounts of curated data.

But the NYU researchers found that their model could successfully associate images and sounds with substantially less data coming from a baby’s video. Their model had a 61.6% accuracy rate when classifying 22 “visual concepts.”

“We were very surprised that the model could exhibit a pretty remarkable degree of learning given the limited data it had,” said Wai Keen Vong, lead author of the NYU paper, in a video interview.

These findings are an encouraging boost for the development of future AI models. But, as Vong notes, they also highlight the phenomenal learning abilities of babies themselves, who can respond to visual cues and develop their own learning hypotheses.

Part of the reason for their precocity is that human babies spend a long period of time actively exploring the world before having to fend for themselves.

“Children are the R&D department [pesquisa e desenvolvimento] of the human species — the blue sky guys, the creators of ideas. Adults are production and marketing,” as Alison Gopnik memorably wrote in her book “The Philosophical Baby.”

According to Gopnik, a psychology professor at the University of California at Berkeley, babies have three key abilities that AI systems do not.

First, babies excel at imaginative model building, creating a conceptual framework to explain the world. They are also curious, adventure-loving, and embodied learners, actively exploring new environments rather than being passively involved in lines of code.

And babies are social animals, learning from everyone they interact with, helping to develop empathy, altruism and a moral sensitivity.

In an email, Gopnik says the NYU study is “fascinating and very clever” and shows that AI models can extract linguistic information from the data babies deal with.

But as the paper’s authors themselves acknowledge, babies also use different data, which they obtain through active exploration and social interaction.

“The success of models, which is still very limited, may take advantage of babies’ exploration and social learning capabilities, but this does not mean that they themselves have these skills,” writes Gopnik.

Much more research will be needed to computationally replicate what babies learn naturally. In particular, how can we build machines that demonstrate common sense and social reasoning?

AI models may be able to learn nouns associated with physical objects, but they still have difficulty with abstract concepts and verbs. Despite impressive advances in AI, we still have a lot to learn from the small set of biological “hardware” that is a baby’s brain.