Apple has recently published two research papers through arXiv, which are studies on AI processing techniques that could potentially be applied to Apple’s products in the future.
The first research paper, titled “LLM in a flash,” presents a method to enable devices with limited memory, such as smartphones, to run complex AI models with a large number of parameters. This is achieved through two techniques: windowing, which reuses previously processed AI results instead of using new data, and row-column bundling, which reads data in chunks rather than individual words. These techniques are particularly suitable for LLM tasks and can increase the memory capacity of the AI LLM by up to 2 times.
The second research paper, titled “HUGS: Human Gaussian Splats,” introduces a method for generating 3D avatars of people using short video inputs. While this concept is not new, Apple claims that their method, called HUGS, is able to generate avatars faster, up to 100 times, and with improved details, such as clothing and hair.
Although Apple only presents these research papers, it gives us a glimpse of what AI features in future Apple products could be like.
TLDR: Apple has released two research papers on AI processing techniques. The first paper presents methods to enhance memory capacity for running complex AI models, while the second paper introduces a faster and more detailed method for generating 3D avatars. These research papers hint at the potential AI features in future Apple products.
Leave a Comment