Apple research paper focuses on how to train facial recognition AI
Apple recently published its first research paper that describes the process of creating realistic fake images—mostly of humans—to train facial recognition artificial intelligence (AI), according to a report by Quartz.
The paper discusses the issue of how training a machine takes a massive amount of data, particularly when it involves faces and body language.
The ability to achieve high results with training data could ultimately allow Apple to develop AI software that understands human movement without relying on any user data.
The research focuses on the examples of basic image recognition issues, such as identifying hand gestures and detecting the direction where people are looking. Both examples could have several applications, including tracking user behavior and a wave-to-unlock iPhone feature.
In both scenarios, the researchers took established datasets of fake images and used a neural network previously trained on real images to improve the fake images to appear more realistic.
The system then compares the improved image to an authentic image, tries to determine which picture is real, and then updates itself based on what the system perceived as fake compared to the real image.
The paper describes the research achieved “state-of-the-art results without any labeled real data.”
In addition to covering user data security issues, the research reinforces an established trend in 2016 of using neural networks to create new data rather than simply identifying it.
The researchers write that they might explore the same technique for videos.