Table of Contents
Recommendation
Artificial intelligence (AI) systems struggle to quantify human beings’ intuitive understanding of the world around them, even though such perceptions are readily developed by infants. Computer scientist Luis Piloto and his team modelled software that mimics infant cognition of basic objects. It allows AI, at the least, to express surprise at inexplicable changes in the physical realm.
Take-Aways
- A new software, PLATO, allows AI to discern how objects behave, and express surprise at aberrations.
- PLATO could represent a first step for testing infant learning behaviors through AI.
Summary
A new software, PLATO, allows AI to discern how objects behave, and express surprise at aberrations.
Computer scientists developed a software program for artificial intelligence systems in response to research about how babies learn. Psychologists test an infant’s understanding of object movements by tracking its eyes. For example, when a ball suddenly disappears, the youngster appears surprised. Researchers measure that emotion by the length of the infant’s gaze toward the ball’s former placement.
Computer scientist Luis Piloto, working at London’s Google-owned DeepMind, worked with a team to develop a similar AI test. They worked with a neural network software that spots patterns in large data sets, training it to analyze animated videos of objects like balls and cubes.
The software, “Physics Learning through Auto-encoding and Tracking Objects (PLATO),” analyzed both raw images and videos highlighting the objects. Scientists programmed PLATO to track the objects’ movements and velocities. PLATO watched “tens of hours” of scenes such as balls bouncing or rolling down slopes, learning to predict how the balls would behave in various situations. Parameters such as continuity and solidity were incorporated.
“At every step of a movie, it [PLATO] makes a prediction about what will happen next. As it gets further into the movie, the prediction becomes more accurate.” (computer scientist Luis Pilato)
When an object “magically” disappeared, PLATO measured the unexpected event using its data set predictions, producing a “measure of surprise.”
PLATO could represent a first step for testing infant learning behaviors through AI.
Jeff Clune, a University of British Columbia computer scientist, called the study “an important research direction,” but pointed out that “…the paper does hand-design much of the prior knowledge that gives these AI models their advantage.”
“We’re hoping this [PLATO] can eventually be used by cognitive scientists to seriously model the behavior of infants.” (Luis Pilato)
Other researchers, including Clune, continue to work on programs that develop algorithms for interpreting the physical world.
About the Author
Davide Castelvecchi, senior physical sciences reporter for Nature, covers physics, astronomy, math and technology subjects.