This Artificial Intelligence System Gains Intuition Like a Human Baby

By definition, “human intuition” seems to denote a barrier between us and artificial intelligence. It’s why we have gut feelings and reflexive — sometimes impulsive — reactions inexplicable by logic, and therefore not simply transferrable to computers. I mean, we can hardly parse our own reasoning for instinctive behavior, so how could we develop algorithms to encode it?

But if we’re to enter a world of lifelike AI, we’re going to need to bridge that gap. We’ll need to figure out how to give robotic systems the power of intuition. And on Monday in the journal Nature, scientists announced they’ve propelled the quest forward. 

In collaboration with AI research laboratory DeepMind in the UK, this team developed an artificial intelligence system that learned “intuitive physics,” that is, commonsense understanding of how our universe’s mechanics work, just like a human baby.

It’s named Physics Learning Through Auto-encoding and Tracking Objects, or PLATO — undoubtedly a nod to the Greek philosopher famous for his allegory of the cave, a thought experiment that probes the nuanced nature of knowledge and meaning.

“Current artificial intelligence systems pale in their understanding of intuitive physics, in comparison to even very young children,” the study authors wrote in their paper. “Here we address this gap between humans and machines by drawing on the field of developmental psychology.”

What’s intuitive physics?

If you were to show a baby a red ball, then block it with a large book, the child might be a little shocked at first. She might wonder, “Uh, did that red ball just…disappear?” But if she sees this situation happen enough times, she’ll eventually realize, “Oh, it’s still there even though I can’t see it. Stuff doesn’t just disappear randomly. We have physics!” 

This is called object permanence, and between birth and age two, it starts to meld into what we consider our intuition. 

Fast-forward to adulthood, when an item is blocked from our view, we don’t ever deliberate the fact it’s still there. We just know. And the new study’s team wanted to help PLATO get to the point where it just knows physical stuff like that. Intuitive physics.

Here’s how everything went down.

Basically, the study team first perused decades of developmental psychology research about how babies learn intuitive physics. Slowly, after reading through that literature, a shared theme began to emerge — “the idea that physical understanding is supported by breaking the world into a discrete set of objects,” Luis Piloto from DeepMind said in a press conference Monday.

In other words, babies seem to learn intuitive physics by observing objects move around, fall down, interact, appear and disappear. Gotta see it to believe it, you might say. Zeroing in on that principle, the researchers developed a deep learning model, which is a system based on massive datasets that can sort of gain skills over time and therefore adjust its own code. This is PLATO. 

Then, the team showed PLATO 28 hours of animated videos about simple physics that involved lots of objects. 

For instance, PLATO watched a ball falling to the ground or rolling behind other objects — and even “impossible” scenarios that defied the laws of physics. Things like objects moving through each other. Scenarios you might find in a magician’s handbook.

At last, a question remained: Could PLATO eventually pick up on intuitive physics like you and I did when we were babies?

Nature vs. Nurture

Well, after those 28 hours of training, the researchers found that it did. 

According to Piloto, PLATO passed all intuitive physics learning tests — with regard to benchmarks put forth by the team — and understood lots of patterns the way human babies begin to as they grow older. It digested continuity, or how an object must follow a trajectory to get from point A to point B rather than teleport, and solidity, meaning two objects can’t be in the same physical space at the same time. And that’s just two such findings.

The way the team recognized that PLATO really learned these commonsense physics concepts was by measuring how accurate its predictions were about what would happen next in a video. As time passed, PLATO’s predictions started getting better and better. Plus, Piloto and colleagues could also measure a degree of “surprise,” which means PLATO’s guesses were super different from PLATO’s observations. 

And sure enough, PLATO exhibited “surprise” when it saw weird magic trick-like videos where things didn’t add up. It knew something was wrong when balls were intersecting and breaking the laws of physics. If it were a human, it’s jaw would drop the way ours does when we watch a magic show. 

But there’s more. 

When exposed to new phenomena that it hadn’t seen before, PLATO was still able to apply the intuitive physics it learned to accurately predict what’ll happen next.

“We leveraged a subset of another synthetic dataset developed by researchers at MIT,” Piloto said. “This dataset also probes physical knowledge, but it has different visual appearances, and importantly, a set of objects that PLATO has never seen.”

All in all, PLATO is pretty intelligent — and quite impressive, to say the least. But further down the line, not only could it help scientists develop more lifelike AI systems, but might inform human learning studies, too. 

As the study’s authors write, “we consider the implications of these results both for AI and for research on human cognition.” For example, PLATO learning so much about intuition from animated videos essentially proves visual demonstrations truly help someone – or something – gain knowledge. 

This then leads to the dilemma of nature versus nurture. 

“The data suggest that intuitive physics knowledge emerges early in life but can be impacted by visual experience,” the authors wrote. “Of course there is extensive debate and legitimate uncertainty about innateness.”