Xiaomi executives talk about the deep learning framework MACE: integrated into APPs by many software companies

Cui Baoqiu, vice president of Xiaomi Group and chairman of Xiaomi Technology Committee, talked about the deep learning framework MACE last night.

Cui Baoqiu said that in order to implement the many technical achievements of Xiaomi's algorithm team on mobile phones as soon as possible through the AI ​​framework, in 2017, after examining the popular deep learning frameworks at the time and repeatedly weighing the pros and cons, Xiaomi decided to develop its own mobile deep learning reasoning framework, and in 2018 it launched and open-sourced MACE, a deep learning reasoning framework specifically optimized for mobile devices.

MACE established four goals when the project was launched: universality, high performance, ease of use, and portability. In subsequent technical practice, the MACE team designed and developed technical solutions around these four goals, and made special optimizations in multi-framework support, performance, system response, power consumption, memory usage, model protection, etc.

In 2020, MACE launched and open-sourced the microcontroller AI inference engine MACE Micro to fully support ultra-low power inference scenarios for mobile phones and IoT devices. In addition to inheriting the general, high-performance, easy-to-use and portable features of the MACE engine, MACE Micro also focused on portability in consideration of the diversity of microcontroller chips and software systems. MACE Micro does not rely on any OS, file system, heap memory allocation, C++ library, or other third-party libraries other than math libraries. It only requires the microcontroller's compilation tool to support C++. In addition, the MACE microcontroller engine is also optimized for special chips in order to achieve industry-leading performance.

Cui Baoqiu revealed that after the unremitting efforts of the MACE team, in addition to the visual field, the MACE framework has also been widely used in multiple fields such as speech, NLP, and human behavior recognition. At present, the MACE framework has become the preferred Al framework within the company and is deployed on all models released by Xiaomi in recent years. The number of calls to the MACE framework on MIUI exceeds 50,000,000 times per day. Outside the company, the MACE framework has also been integrated and deployed by APPs of many domestic and foreign software companies. In the future, MACE will continue to work on ease of use, high performance, and compatibility to provide solid AI support for applications such as Xiaomi Camera.

Finally, Cui Baoqiu said that in order to adapt to the new demands of the community, MACE is still being improved, and he hopes to attract more developers to join in the future to jointly create a more complete community ecosystem. Xiaomi will continue to increase its investment in open source, contribute to the open source community, unite more Chinese Internet companies in the fields of AI, big data, cloud computing, loT, etc., and promote China's open source power to the world.