Preeti
Apple introduced Apple Intelligence at the 2024 Worldwide Developers Conference, integrating it deeply into iOS 18, iPadOS 18, and macOS Sequoia. Apple Intelligence consists of generative models tailored for everyday tasks, such as writing, notifications, and creating images. The focus is on responsible AI development, guided by principles like empowering users, representing diversity, designing with care, and protecting privacy.
The foundation models are trained using Apple’s AXLearn framework and a combination of data parallelism and tensor parallelism. Data quality is crucial, and post-training algorithms are used to fine-tune the models. Various optimization techniques are applied for on-device and server models to ensure speed and efficiency.
Model adaptation is achieved through adapters, allowing the models to specialize for specific tasks without altering the base parameters. Performance and evaluation involve human grading to assess the models’ effectiveness in tasks like summarization and writing. The models outperform comparable models in terms of safety, helpfulness, and instruction-following capabilities.
Apple’s models are robust against adversarial prompts and are preferred by human graders for their safety and helpfulness. Ongoing evaluations and red-teaming activities ensure the models’ safety. The models excel in writing ability, instruction-following, and general capabilities compared to other models of similar size.
Overall, Apple’s foundation models and adapters power Apple Intelligence, providing users with advanced capabilities while maintaining a focus on responsible development and user privacy. Further advancements and details on a broader family of generative models are expected to be shared in the future.
Source link
Source link: https://machinelearning.apple.com/research/introducing-apple-foundation-models
GIPHY App Key not set. Please check settings