Artificial intelligence (AI) already outperforms us humans in certain areas, such as playing Go or processing large data sets. However, there are other aspects where the AI is reluctant to overtake us, even if we only have a few months to live. A good example of this is how babies instinctively know that when one object passes behind another it should not disappear. If that happens, they are surprised.
This simple rule of continuity along with other physical laws is, however, not so instinctive for AI. Now, PLATO (Physics Learning through Auto-encoding and Tracking Objects) AI has been inspired by the study of how babies learn to think like them.
PLATO has been trained using videos of balls that are specifically designed to represent the basic knowledge that babies have in their first months of life. “Luckily for us, developmental psychologists have spent decades studying what infants know about the physical world and cataloging the different ingredients or concepts that go into physical understanding,” says neuroscientist Luis Piloto of the AI Research Lab. DeepMind in the UK. “Expanding their work, we have created an open source Physical Concepts Dataset. This synthetic video dataset draws inspiration from the original development experiments to evaluate the physical concepts in our models.”
When we are babies there are three concepts that we understand perfectly: permanence, solidity and continuity . Thanks to permanence, objects do not suddenly disappear. Solidity is the quality that makes solid objects unable to pass through each other. Continuity refers to the consistency with which objects move in space and time. In addition to these three concepts, the researchers have added two others to the dataset they have built: immutability , whereby object properties such as shape do not change, and directional inertia , which causes objects to move coherently with the principles of inertia.
All these concepts were transmitted to PLATO through videos of balls that staged them falling to the ground, bouncing off each other, disappearing behind other objects, reappearing, etc. After training the AI with these contents, it was put to the test.
What happened was that when PLATO was shown videos with impossible scenarios that defied the physics it had learned, the AI was shocked . He recognized that something strange that broke the laws of physics had happened. This happened after relatively short training sessions, some as short as 28 hours.
The researchers were looking for evidence of violation of expectations (VoE) signals , which would show that the AI understood the concepts it had been taught and which are also used in infant studies. “Our object-based model showed robust VoE effects across all five concepts we studied, despite being trained on video data in which the specific polling events did not occur,” the researchers write in their paper published in Nature . Human Behaviour .
The researchers ran more tests, this time with different objects than the training data. Again, PLATO showed a solid understanding of what should and should not happen , showing that he could learn and extend his basic training knowledge. However, the AI is not yet at the level of a three-month-old baby, and his surprise when shown scenarios that did not include any objects, or when the test and training models were similar, was less. The videos with which the AI was trained included additional data that helped it recognize objects and their movement in three dimensions.
The research could serve to better understand the mind of human beings , as well as to build a better representation of AI. “Our modeling work provides proof of concept demonstrating that at least some core concepts of intuitive physics can be acquired through visual learning,” the researchers write in the paper.
“Although research in some precocial species (born at an advanced stage: they can see, stand and perform other adult functions) suggests that certain basic physical concepts may be present from birth, in humans the data suggest that Intuitive physical knowledge emerges early in life, but can be impacted by visual experience.”
Referencia: Piloto, L.S., Weinstein, A., Battaglia, P. et al. 2022. Intuitive physics learning in a deep-learning model inspired by developmental psychology. Nature Human Behaviour. DOI: https://doi.org/10.1038/s41562-022-01394-8
