The DeepMind artificial intelligence research laboratory created the PLATO (Physics Learning through Auto-encoding and Tracking Objects) artificial intelligence system. The authors of the development say that she learns like a normal child.
Existing AI systems show high results in solving a number of problems, but in most aspects they are inferior to people. For example, even children instinctively understand that one object that is briefly behind another does not disappear. However, this rule was not as intuitive to the AI.
PLATO was taught through a series of coded videos designed to represent the same basic knowledge that babies learn in the first few months of life.
“Fortunately for us, developmental psychologists have spent decades studying what babies know about the physical world and cataloging the various concepts that go into physical understanding,” says neuroscientist Luis Piloto of the DeepMind Research Lab. “By extending their work, we have created and opened up a data set of physical concepts. This synthetic video dataset draws inspiration from original design experiments to evaluate physics concepts in our models.”
There are three key things that we all understand from a very young age: persistence (objects don't suddenly disappear); hardness (solid objects cannot pass through each other); and continuity (objects move sequentially in space and time).
The dataset the researchers created covered these three concepts, as well as two additional ones: immutability (an object's properties, such as shape, do not change); and directional inertia (objects move according to the principles of inertia).
This understanding was taught to PLATO through videos of balls falling to the ground, bouncing off each other, disappearing behind other objects and then reappearing, and so on.
The AI was then shown videos of "impossible" scenarios that contradicted the physics they had studied, and PLATO "expressed surprise".
This also happened after relatively short training periods, which in some cases took as little as 28 hours. From a technical standpoint, as in infant studies, scientists were looking for evidence of unrealistic expectations. They showed that the AI understood the concepts it was taught.
“Our object-oriented model exhibited robust out-of-expectation effects across all five concepts we studied, despite being trained on video data in which specific test events did not occur,” the researchers say.
The team ran additional tests, this time using different objects from those in the training data. Again, PLATO has shown a clear understanding of what should not happen.
However, the system does not even quite correspond to the level of understanding of a three-month-old child. So, the AI was less "surprised" when it was shown scenarios without any objects or when the test models were similar to training ones.
What's more, the videos PLATO was trained on included additional data to help it recognize objects and their movement in three dimensions.
"Our simulation work represents evidence that at least some of the core concepts of intuitive physics can be learned through visual learning," the researchers note. According to them, "the data show that in humans, intuitive physical knowledge emerges at an early age, but it can be influenced by visual experience."
bbabo.Net