Apple Plans to Boost AI Using Private User Data Analysis

Following criticism of its AI tools—particularly for weak notification summaries—Apple on Monday outlined efforts to enhance its AI models by privately analyzing user data, supplemented with synthetic data.

Apple to Boost AI Performance with Differential Privacy, Using Synthetic Data While Safeguarding User Privacy

Apple explained that it will use a technique called “differential privacy,” which involves generating synthetic data and then sending snippets of it to users’ devices—if they’ve opted in to share analytics. This allows Apple to assess model accuracy and make improvements, all while preserving user privacy.

In a blog post, Apple explained that “synthetic data” are designed to replicate the structure and key characteristics of user data without including any real user-generated content. To build a representative collection of synthetic emails, Apple begins by creating a wide range of artificial messages on different subjects. Each message is then transformed into an “embedding” — a type of data representation that reflects important features such as language, topic, and length.

Apple Sends Embeddings to Opted-In Devices to Refine AI Accuracy Through On-Device Email Comparisons

Apple stated that these embeddings are sent to a limited number of user devices that have opted in to Device Analytics. The devices then compare the embeddings with a sample of emails to help identify which ones are the most accurate.

Apple noted that it’s applying this method to enhance its Genmoji models and plans to expand the use of synthetic data to features like Image Playground, Image Wand, Memories Creation, Writing Tools, and Visual Intelligence. Additionally, it will use synthetic data to poll opted-in users’ devices to refine email summary capabilities.

Read the original article on: TechCrunch

Read more: Apple iPhone 16e Review: A18 Chip and Apple Intelligence for $599