Meta Platforms, Inc.
Differentiating physical and non-physical events

Abstract:

To differentiate physical and non-physical events, a discrimination system based on unsupervised machining learning is used to predict a plausibility of objects' behaviors between a starting and ending time point. The discrimination system receives a set of initial, or "starting" content frames, each depicting a state of objects at a starting time point and an arrangement or "behavior" of those objects at the starting time. To train the discrimination system, the first model uses the starting content frame to generate a subsequent content frame, while the second model generates a subsequent content frame without using the starting content frame. A discriminator model may thus be trained without supervision by treating the subsequent content frame generated from the first model as a possible behavior of the starting content frame, and the subsequent content frame generated from the second model as an impossible behavior of the starting content frame.