Creating an AI system that can think like a human has been one of computer science’s major obstacles.
Now, researchers claim to have constructed an artificial intelligence that can think like a baby by teaching it basic physical world laws.
Their deep learning system is capable of understanding “intuitive physics” — the common-sense laws that govern how physical objects interact.
In experiments, the academics educated the new PLATO system with a series of animated slides depicting the movement of balls.
After being taught a modest number of visual animations, PLATO was able to display learning and even surprise when a ball moved inconceivably.
The new research was undertaken by scientists from Princeton University in New Jersey, University College London, and DeepMind, a company owned by Google, and published in Nature Human Behaviour.
According to them, their discoveries are crucial in the effort to create AI models with the same physical comprehension as adults.
Dr. Luis S. Piloto of DeepMind, who authored the study, remarked, “Understanding the physical environment is a crucial ability that the vast majority of people employ with ease.”
However, this remains a barrier for artificial intelligence; if we are to deploy beneficial systems in the real world, we need these models to share our intuitive understanding of physics.
In 1950, the famed British computer scientist Alan Turing recommended educating an artificial intelligence to have the intelligence of a child and then providing it with the right experiences to raise its intelligence to that of an adult.
Instead of trying to create a program that models the adult mind, why not try to create one that simulates the mind of a child? His major research work, Computing Machinery, and Intelligence were written by Turing.
If this were subsequently subjected to a proper educational program, an adult brain would result.
According to the authors of this new study, even very young children are aware of “intuitive physics” — the common-sense rules that govern how the world functions.
For instance, if a person dangled their keys in midair and announced they were about to let them go, everyone around them would recognize that unsupported things do not float in midair.
They would also understand that two items, such as the keys and the table underneath, cannot travel through one another. People would therefore anticipate that the keys will fall till they reach the table.
Even three-month-old infants possess this knowledge, and they react if they meet a magical circumstance that appears to contravene these expectations.
For instance, infants as young as five months are astonished when presented with a physically impossible scenario, such as a toy suddenly disappearing.
For their study, the researchers questioned whether AI models can learn a variety of physical notions, notably those that toddlers grasp, such as solidity (the idea that two objects cannot pass through one another) and continuity (the notion that items do not vanish into thin air).
They created the artificial intelligence system PLATO so it could represent visual inputs as a set of objects and reason about object interactions.
The authors trained PLATO by showing it movies of a variety of simple scenarios, including balls dropping to the ground, balls rolling behind other objects and reappearing, and balls bouncing off one another.
After training, PLATO was evaluated by giving it videos with implausible situations, such as balls vanishing and reappearing on the opposite side of the frame.
Similar to a young toddler, PLATO displayed surprise when presented with nonsensical information, such as objects flowing through each other without interacting.
According to Dr. Piloto, one interpretation of the term “surprise” is anticipating something and discovering something else.
“PLATO predicts the configuration of upcoming items it will observe. As the movie continues to play, the true configuration of the objects is observed.
The surprise is the difference between the expected configuration and the actual configuration in the next video frame.
These learning effects were observed after only 28 hours of video viewing.
The authors suggest that PLATO could be a useful instrument for investigating how humans acquire intuitive physics knowledge.
Results also indicate that infant-modeled deep learning systems outperform more conventional “learning from scratch” systems.
Susan Hespos and Apoorva Shivaram write in an accompanying News & Opinions post, “The results of this study suggest that Turing may have been correct.”
‘Common-sense In physics, progress enhances and refines existing knowledge without fundamentally altering it.
This means that studies of object knowledge in infants can provide insight into object knowledge in adults and potentially teach us how to create more realistic computer models of the human mind.
