The entertainment company have developed a new method - which is known as factorized variational autoencoders (FVAEs) - that allows the company to detect how people are reacting to a movie and what emotions it makes them feel.
They wrote in the paper: "We focus on learning such a representation for a large data set of facial expressions extracted from movie-watching audiences. Intuitively, we expect audience members to have correlated reactions, since each movie has been specifically crafted to elicit a desired response.
"Thus, one can view audience analysis as a form of collaborative filtering, which has been popularized for modeling recommender systems (e.g., the Netflix challenge). For instance, we can assume that there are underlying exemplar facial expressions which form a basis to reconstruct the observed reactions of each audience member. The most common approaches for collaborative filtering are factorization techniques, which we build upon in our work."
The paper was presented at the IEEE Conference on Computer Vision and Pattern Recognition and the company reported that they had already had a lot of success with it compared to more conventional methods.
They added: "Our goal is to learn a compact and expressive representation of audience reactions that is semantically meaningful, so that we can identify patterns in the data and succinctly summarize observed behaviour ...
"Finally, we demonstrate the predictive power of FVAEs in the challenging task of anticipating the facial reactions of audience members for an entire movie, using only observations from the first few seconds/minutes. Our suite of results suggests that FVAEs can capture significantly more expressive representations than conventional baselines."