Apple has recently faced criticism over the lackluster performance of its AI products, particularly in areas such as notification summaries. In response to this feedback, the tech giant unveiled its strategy to enhance its AI models through the analysis of user data in a private manner, leveraging synthetic data.
Employing a technique known as “differential privacy,” Apple disclosed its plan to create synthetic data first and then engage users’ devices (with consent to share device analytics) by providing them with snippets of the generated synthetic data. This process allows Apple to assess the accuracy of its models and make necessary improvements.
According to a blog post by Apple, synthetic data is designed to replicate the format and essential characteristics of user data without containing any actual user-generated content. By generating a diverse set of synthetic messages on various topics and deriving embeddings that capture key dimensions like language, topic, and length, Apple can create a representative set of synthetic emails.
These embeddings are then distributed to a select group of user devices participating in Device Analytics, which compare them with a sample of emails to determine the most accurate representations. Apple is utilizing this approach to enhance its Genmoji models and intends to extend its use to Image Playground, Image Wand, Memories Creation, Writing Tools, and Visual Intelligence in the future. Additionally, Apple plans to gather feedback from users who opt in to share device analytics using synthetic data to enhance email summaries.
By incorporating synthetic data and differential privacy techniques into its AI development process, Apple aims to address the shortcomings of its AI products and deliver more effective and personalized user experiences. This innovative approach underscores Apple’s commitment to leveraging cutting-edge technologies while safeguarding user privacy.
