Lower the programming threshold of big model training architecture 

2022-05-09 13: 30 Source: Science and Technology Daily

In addition to hardware such as chips, artificial intelligence operating system is also the core key technology in the field of artificial intelligence that needs to be controlled independently. A few days ago, the first "National Disruptive Technology Innovation Competition" (hereinafter referred to as the Competition) sponsored by the Ministry of Science and Technology and hosted by Torch High-tech Industry Development Center of the Ministry of Science and Technology came to an end. After layers of selection, 36 high-quality projects won the highest award-the winning prize, including the "Artificial Intelligence Operating System" independently developed by the Chinese team.

It is understood that this competition is held in Suzhou, Chengdu and Qingdao, aiming at strengthening the supply of disruptive technologies, cultivating disruptive innovation culture, exploring a new mechanism of "discovery-selection-cultivation" of disruptive technologies, and focusing on discovering and excavating a number of disruptive technology directions. 2,724 technical projects from key universities, well-known research institutes, leading industry enterprises and emerging innovative enterprises all over the country participated in the competition.

In the field of artificial intelligence, the deep learning framework is the basic software, which is called "the operating system in AI field". Its positioning is similar to that of Windows in PC era, IOS and Android in mobile Internet era, and it is the key link to connect artificial intelligence elements such as data, algorithms and computing power. It connects chips below and applications above, and all algorithms and applications of artificial intelligence should be trained and deployed through the deep learning framework. Its importance is self-evident.

The United States, which started the fastest in the field of deep learning framework, has two products with the highest market share—TensorFlow developed by Google and PyTorch developed by Meta (formerly Facebook), which almost form a monopoly. At the same time, Amazon, Microsoft and other international technology giants are also developing similar products by themselves. In China's "14th Five-Year Plan", the "Deep Learning Framework" has been included in the "New Generation Artificial Intelligence" section, which has become a frontier innovation technology supported by the state. Baidu, Huawei, Defiance and other enterprises in China have also launched deep learning frameworks.

It is understood that the deep learning framework has three important dimensions: completeness, ease of use and high efficiency. The existing deep learning framework is very mature in solving completeness and ease of use, so high efficiency has become the breakthrough point of domestic framework. The "Artificial Intelligence Operating System" that won this award is the OneFlow distributed deep learning framework from Beijing's first-class technology, which is characterized by its high efficiency. It pioneered the static scheduling and streaming execution architecture, which solved the "network wall" problem faced by distributed deep learning. At the same time, the distributed programming interface based on the overall perspective enabled ordinary engineers to quickly train and deploy large models, which enabled more enterprises to have the ability to deeply mine their own data assets, lowered the programming threshold of the large model training architecture, and made it possible for large models to land quickly.

It is understood that at present, the deep learning framework, as the basic software independently developed by the Chinese team, has formed a purely domestic AI model training and deployment solution. In the future, standardized modules specially adapted to emerging hardware will be introduced.

(The above article is reproduced, and does not represent the views of our company. If there are any issues related to copyright, please contact us for handling.)