Artificial intelligence (AI) systems are already well ahead of us in certain areas (playing Go, for example, or analyzing large data sets), but in other respects AI is still far behind humans. even just a few months after you were born.
For example, even young babies instinctively know that one object that passes briefly behind another should not disappear and reappear elsewhere. Presented with such a magical act, the babies act in surprise.
But such a simple rule of continuity, along with other basic physical laws, has not been so intuitive to AI. Now, a new study introduces an artificial intelligence called PLATO that has been inspired by research on how babies learn, and they can think like a human baby does.
PLATO stands for Physics Learning through Auto-encoding and Tracking Objects, and was trained through a series of coded videos designed to represent the same basic knowledge that babies have during their first months of life.
“Luckily for us, developmental psychologists have spent decades studying what children know about the physical world and cataloging the different ingredients or concepts that go into physical understanding,” says neuroscientist Luis Piloto of the research lab of AI DeepMind in the UK.
“Expanding their work, we created and opened a physical concept data set. This synthetic video data set is inspired by the original development experiments to evaluate the physical concepts of our models.”
There are three key concepts we all understand from a very young age: permanence (objects will not suddenly disappear); solidity (solid objects cannot pass between each other); and continuity (objects move coherently through space and time).
The data set constructed by the researchers covered these three concepts, plus two additional ones: immutability (the properties of the object, such as form, do not change); and directional inertia (objects move in a manner consistent with the principles of inertia).
These concepts were transmitted through clips of balls falling to the ground, bouncing on top of each other, disappearing behind other objects and then reappearing, and so on. Having formed PLATO in these videos, the next step was to try it out.
When videos of “impossible” scenarios challenging the physics he had learned were shown in AI, PLATO expressed his surprise (or the equivalent of AI): he was smart enough to acknowledge what had happened something strange that broke the laws of physics.
This also happened after relatively short training periods, only 28 hours in some cases. Technically speaking, as in early childhood studies, researchers were looking for evidence-of-expectation (VoE) evidence, demonstrating that AI understood the concepts that had been taught to it.
“Our object-based model showed robust VoE effects in the five concepts we studied, despite being trained in video data in which probe-specific events did not occur,” the researchers write in their article published.
The team did more testing, this time using objects other than the training data. Once again, PLATO showed a solid understanding of what should and should not happen, demonstrating that he could learn and expand his basic training knowledge.
However, PLATO has not yet reached the level of a three-month-old baby. There was less surprise from AI when scenarios involving no object were shown, or when test and training models were similar.
In addition, the videos in which PLATO trained included additional data to help him recognize objects and their movement in three dimensions.
It seems that some integrated knowledge is still needed to get a complete picture, and this “nature vs food” question is something that development scientists still ask in babies. Research could give us a better understanding of the human mind, as well as help us build a better representation of AI.
“Our modeling work provides a proof-of-concept demonstration that at least some central concepts of intuitive physics can be acquired through visual learning,” the researchers write.
“Although research in some early [born in an advanced state] The species suggests that certain basic physical concepts may be present from birth, in humans the data suggest that intuitive knowledge of physics arises early in life, but that it can be affected by visual experience. “
The research has been published in Nature Human Behavior.