As a results of a criticism related to the disappointing performance of its AI products, especially in such areas notification Summaries, Apple on Monday detailed How he tries to improve his AI models through a private analysis of user data using synthetic data.
Application of the approach called “differential privacy“The company said that it would first generate synthetic data after which surveys of user devices (provided that they decided to provide an analysis of Apple devices) with fragments of generated synthetic data so as to compare how accurate their models are, after which improve them.
“Synthetic data is created to imitate the format and important properties of the user’s data, but they do not contain any actual content generated by the user,” wrote the corporate within the blog post. “To select a representative set of synthetic E -Mail messages, we start by creating a large set of synthetic news on various topics […] Then we lead the representation, called the embedding of each synthetic message, which records some key dimensions of messages, such as language, topic and length. ”
The company said that these embedding are then sent to a small variety of user devices that decided to analyze the devices, after which compare them with the e -Mail sample to inform Apple which deposition are essentially the most accurate.
The company said that that is using this approach to improving its Genmoji models, and in the long run it would use synthetic data for the playground of the image, wand, creating and writing memories, in addition to visual tools. Apple said that there may even be user surveys who resolve to provide an analysis of devices with synthetic data to improve the summary of E -Mail messages.
