PLATO was taught to recognise objects and their interactions by DeepMind AI researchers using simulated videos of objects moving as we would expect. To improve PLATO’s performance, they used video training data to show videos of many simple scenes. It employs objects at all stages of processing, including representing visual inputs as a set of objects, reasoning about object interactions, and producing outputs. According to Piloto, the findings indicate that an object-centric view of the world could provide an AI with a more generalised and adaptable set of abilities. The findings could open up new avenues for AI research and even provide insights into human vision and development. However, PLATO implementation is not externally viable. The findings suggest that visual animations can explain some intuitive physics learning, but not enough to explain what we see in infants.
— DeepMind (@DeepMind) July 11, 2022 “In comparison to even very young children, current artificial intelligence systems pale in their understanding of intuitive physics,” the study authors wrote in their paper. “We address this gap between humans and machines here by drawing on developmental psychology.”
