Apple, which is often accused of lagging on artificial intelligence, is actually actively working on breakthroughs in this area. At the same time, the company prefers not to advertise its developments in order to avoid high expectations. However, when it comes to specific achievements, this is of great importance for the entire industry.
Recently, Apple introduced two new scientific articles demonstrating significant progress in research. In the first article scientists offer the HUGS (Human Gaussian Splats) for generating animated 3D models taken on one camera. The algorithm automatically learns to separate the static and dynamic parts of the frame in order to create a completely animated human avatar in less than 30 minutes.
According to the authors of the study, compared with previous methods, Hugs is 100 times faster both in training and in rendering. This impressive achievement opens up new opportunities for virtual examples, television prisms and synthetic media in the near future.
the second scientific article is devoted to optimizing the work of large language models on limited memory devices, such as smartphones and tablets. The fact is that modern models like GPT-4 contain hundreds of billions of parameters, which draws their conclusion resource-intensive for consumer equipment.
The proposed Apple system minimizes the data transfer from flash memory to the operational one. First, a model is built, which evaluates computational costs based on the features of flash memory. This model allows you to optimize two key points:
1. Reduce the amount of data transmitted.
2. Read data from memory larger blocks.
Development will allow you to launch complex AI helpers and chat bots on the iPhone and other mobile devices. Although the company does not advertise its achievements in AI, it is obvious that it has already achieved impressive results. Perhaps the public will be able to get acquainted with them in more detail very soon.