In a world filled with opposing views, let’s call attention to something we can all agree on: if I show you my pen and then hide it behind my back, my pen still exists – even if you can’t see it anymore. . We can all agree that it still exists, and it probably has the same shape and color it had before it passed my back. This is just common sense.
These common sense laws of the physical world are universally understood by humans. Even two-month-old babies share this understanding. But scientists are still puzzled by some aspects of how we arrived at this fundamental understanding. And we have yet to build a computer that can rival the common-sense abilities of a typically developing child.
New research by Luis Piloto and colleagues at Princeton University — which I’m reviewing for a paper in Nature Human Behavior — takes a step toward filling that gap. The researchers created a deep learning artificial intelligence (AI) system that gained an understanding of some common sense laws of the physical world.
The findings will help build better computer models that simulate the human mind, approaching a task with the same assumptions as a child.
Typically, AI models start with a blank slate and are trained on data with many different examples from which the model builds knowledge. But baby research suggests that’s not what babies do. Instead of building knowledge from scratch, babies start with some principled expectations about objects.
For example, they expect that if they serve an object that is hidden behind another object, the first object will continue to exist. This is a core assumption that starts them in the right direction. Your knowledge then becomes more refined with time and experience.
The exciting discovery by Piloto and colleagues is that a deep learning AI system modeled on what babies do outperforms a system that starts with a blank slate and tries to learn from experience alone.
Read more: Artificial intelligence can deepen social inequality. Here are 5 ways to help prevent it.
Cube slides and ball on walls
The researchers compared the two approaches. In the blank version, the AI model received various visual object animations. In some examples, a cube would slide down a ramp. In others, a ball bounced off a wall.
The model detected patterns from various animations and was tested on its ability to predict results with new visual object animations. This performance was compared to a model that had “principle expectations” built in before trying out any visual animations.
These principles were based on the expectations babies have about how objects behave and interact. For example, babies expect two objects not to pass each other.
If you show a child a magic trick where you violate this expectation, he or she can detect the magic. They reveal this knowledge by looking significantly longer at events with unexpected or “magical” outcomes, compared to events where outcomes are expected.
Babies also expect an object to not be able to simply flicker in and out of existence. They can also detect when that expectation is violated.
Piloto and his colleagues found that the deep learning model that started with a blank slate did a good job, but the model based on object-centric coding inspired by children’s cognition was significantly better.
The latter model could more accurately predict how an object would move, was more successful in applying expectations to new animations, and learned from a smaller set of examples (for example, it achieved this after the equivalent of 28 hours of video).
An innate understanding?
Of course, learning from time and experience is important, but it’s not the whole story. This research by Piloto and colleagues is contributing to the age-old question of what can be innate in humans and what can be learned.
Furthermore, it is setting new limits on the role that perceptual data can play when it comes to artificial systems that acquire knowledge. And it also shows how studies on babies can contribute to building better AI systems that simulate the human mind.