Researchers train an AI model to 'think' like a baby, and suddenly it becomes amazing
In a world full of opposing views, let’s draw attention to something we all agree on: if I show you my pen, and then hide it behind me, my pen is still there – even though you can’t see it anymore. We can all agree that it’s still there, and probably the same shape and color as before it went behind me. This is just common sense.
This common sense law of the physical world is universally understood by humans. Even a two month old baby shares this understanding. But scientists are still confused about some aspects of how we reach this fundamental understanding. And we haven’t built a computer that can compete with babies’ normally developing common sense abilities.
New research by Luis Piloto and his colleagues at Princeton University – which I reviewed for an article in Nature Human Behavior – takes steps to fill this gap. Researchers created a deep learning artificial intelligence (AI) system that gains 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 the task with the same assumptions as babies.
Childish behavior
Typically, an AI model starts with a blank whiteboard and is trained on data with many different examples, from which the model builds knowledge. But research on babies shows that this is not what babies do. Instead of building knowledge from scratch, babies start with some principled expectations about objects.
For example, they hope that if they notice an object that is then hidden behind another object, the first object will continue to exist. It’s the core assumptions that start them in the right direction. Their knowledge then becomes more refined with time and experience.
An interesting finding by Piloto and colleagues is that deep learning AI systems that mimic what babies do outperform systems that start with a blank whiteboard and try to learn by experience alone.
Read more: Artificial intelligence can deepen social inequality. Here are 5 ways to help prevent this
Slide cube and ball to wall
The researchers compared the two approaches. In the blank-slate version, the AI model is rendered several visual animations of objects. In some instances, a cube will slide down a slope. Elsewhere, the ball bounces off the wall.
The model detects patterns from various animations, and is then tested for its ability to predict outcomes with new visual animations of objects. This performance is compared to a model that has “principled expectations” built into it prior to experiencing any visual animation.
These principles are based on the expectations babies have about how objects behave and interact. For example, babies expect two objects not to pass through each other.
If you show babies a magic trick where you violate these expectations, they can detect the magic. They expressed this knowledge by looking significantly longer at events with an unexpected, or “magical” outcome, compared to events in which an outcome was expected.
Babies also expect an object to not just flash in and out of existence. They can detect when these expectations are violated as well.
Shutterstock
Piloto and colleagues found a deep learning model that started with a blank whiteboard did a good job, but a model based on object-centred coding inspired by infant cognition was significantly better.
The latter model can predict more accurately how an object will move, more successfully apply expectations to new animations, and learn from a smaller set of examples (for example, this model succeeds after the equivalent of 28 hours of video).
Innate understanding?
Obviously learning through time and experience is important, but that’s not the whole story. This research by Piloto and colleagues contributes insight into the age-old question of what may be innate in humans, and what can be learned.
Beyond that, it defines a new frontier for what role perceptual data can play when it comes to artificial systems that acquire knowledge. And it also shows how research in infants can contribute to building better AI systems that simulate the human mind.
#Researchers #train #model #baby #suddenly #amazing
Comments
Post a Comment