Here’s what you need to know when researchers trained this AI to ‘think’ like a baby.
In a world full of opposing viewpoints, let us focus on 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. We can all agree that it still exists and is most likely the same shape and color it was before it went behind my back. This is simply common sense. These physical laws of common sense are universally understood by humans. Even two-month-old infants understand this. However, some aspects of how we achieve this fundamental understanding continue to perplex scientists. And we have yet to create a computer that can compete with a typically developing infant’s common-sense abilities.
This gap is being filled by new research by Luis Piloto and colleagues at Princeton University. The researchers developed a deep-learning artificial intelligence (AI) system that learned about some common-sense physical laws. The findings will help to improve computer models that simulate the human mind by approaching a task with the same assumptions that an infant would have.
Typically, AI models begin with a blank slate and are trained on data containing a wide range of examples, from which the model constructs knowledge. However, research on infants indicates that this is not what babies do. Instead of starting from scratch, infants begin with some basic expectations about objects. For example, they believe that if they attend to an object that is then hidden behind another object, the first object will remain. This is a fundamental assumption that gets them started in the right direction. With time and experience, their knowledge becomes more refined. Piloto and colleagues discovered that a deep-learning AI system modeled on what babies do outperforms a system that starts with a blank slate and tries to learn solely through experience.
Balls and cube slides into walls
Both approaches were compared by the researchers. The AI model was given several visual animations of objects in the blank-slate version. A cube might slide down a ramp in some examples. In others, a ball collided with a wall. The model detected patterns in the animations and was then tested for its ability to predict outcomes using new visual animations of objects. This performance was compared to a model that had “principled expectations” built-in before any visual animations were applied.
These principles are based on infants’ expectations of how objects behave and interact. Infants, for example, believe that two objects should not pass through each other. If you show an infant a magic trick that violates this expectation, they will be able to detect the magic. They reveal this knowledge by examining events with unexpected, or “magic,” outcomes for a significantly longer period than events with expected outcomes.
Children also believe that an object should not be able to blink in and out of existence. They can also detect when this expectation is broken. Piloto and colleagues discovered that while the deep-learning model that started with a blank slate performed well, the model based on object-centered coding inspired by infant cognition performed significantly better. The latter model could predict how an object would move more accurately, was more successful at applying expectations to new animations, and learned from a smaller set of examples (for example, it managed this after the equivalent of 28 hours of video).
An innate comprehension
Learning through time and experience is undeniably important, but it isn’t the whole story. Piloto and colleagues’ research sheds light on the age-old question of what is innate in humans and what can be learned. Beyond that, it is defining new boundaries for the role of perceptual data in the acquisition of knowledge by artificial systems. It also demonstrates how research on babies can help to improve AI systems that simulate the human mind.