Following criticism of its AI products, particularly in notification summaries, Apple announced on Monday its strategy to improve its AI models through private user data analysis, utilizing synthetic data.
The company explained that it will implement “differential privacy,” first generating synthetic data and then polling users’ devices—if they have opted in to share device analytics. This involves comparing snippets of synthetic data to assess and refine the accuracy of its models.
ICYMT: Miracles Aboagye to NPP: Keep Internal Matters Private
“Synthetic data mimics the format and key properties of user data without containing any actual user-generated content,” Apple noted in a blog post. “To create a representative set of synthetic emails, we generate a variety of synthetic messages on different topics. We then derive an embedding for each message that captures essential dimensions such as language, topic, and length.”
These embeddings will be sent to a select group of devices that have opted into Device Analytics, allowing Apple to determine which embeddings are most accurate by comparing them with sample emails.
Apple stated this method will enhance its Genmoji models and plans to extend the use of synthetic data to tools like Image Playground, Image Wand, Memories Creation, Writing Tools, and Visual Intelligence. Additionally, the company will consult users who opt in to improve email summaries using synthetic data.
SOURCE: TECH CRUNCH
