In the wake of criticism in areas such as the overwhelming performance of AI products, particularly notification overviews, Apple on Monday detailed how it can improve its AI model by using synthetic data to personally analyze user data.
Using an approach called “differential privacy,” the company said it would first generate synthetic data, then use a snippet of the generated synthetic data with the user’s device (if you opted in to share device analysis with Apple) to compare the accuracy of the model and then improve it.
“The synthetic data is created to mimic the format and important properties of user data, but does not contain any actual user-generated content,” the company wrote in a blog post. “To curate the representative set of synthetic emails, we start by creating large synthetic messages on a variety of topics. […] Next, we derive a representation called embedding of each composite message that captures some of the important dimensions of the message, such as language, topic, and length. ”
The company said these embeddings will be sent to a small number of user devices that have opted in to device analysis, and the devices will compare them to samples of emails to tell you which embeddings are the most accurate.
The company said it is using this approach to improve its Genmoji model, and will use synthetic data in the future for image playgrounds, image wands, memory creation, writing tools and visual intelligence. Apple said it would also vote for users who would choose to share device analytics with synthetic data to improve their email overview.
Source link
