On November 30, the Wave Summit+ 2022 Deep Learning Developer Summit hosted by the National Engineering Research Center for Deep Learning Technology and Application was held as scheduled.

Director of Pengcheng Laboratory, Academician Gao Wen of the Chinese Academy of Engineering, Dean of the School of Electronics and Information Engineering of Shenzhen University, Deputy Director of the Technical Committee of the National Engineering Research Center for Deep Learning Technology and Application, and Academician Ding Wenhua of the Chinese Academy of Engineering were invited to deliver speeches. Wang Haifeng, director of the National Engineering Research Center for Learning Technology and Application, delivered a keynote speech entitled "Deep Learning Platform Enlarged Model, Industrial Intelligence Base".

　　Academician Gao Wen pointed out that developers are the core force for the development of the open source ecosystem and the backbone of technological innovation.

At this stage, it is very important to build a basic software and hardware platform for independent innovation in my country.

Flying Paddle is fully open source and open, with many developers and solid core technology. It has done a lot of leading work for the industry, and actively explores the combination with basic research such as scientific computing.

Many large models of Wenxin will be launched on Pengcheng Cloud Brain, and jointly release the Flying Paddle-Pengcheng Cloud Brain release.

　　Academician Ding Wenhua said in his speech that the National Engineering Research Center for Deep Learning Technology and Application is an important part of the national science and technology innovation system.

As the core research achievement of the engineering research center, the flying paddle platform has played an important role in ensuring the security of national information technology and promoting the large-scale implementation of artificial intelligence applications.

For the underlying core technology in the field of AI, the development initiative must be in our own hands.

The core technology with independent intellectual property rights is the source of core competitiveness.

　　Wang Haifeng announced the latest progress of the Flying Paddle ecology: As of now, Flying Paddle has gathered 5.35 million developers, served 200,000 enterprises and institutions, and created 670,000 models based on Flying Paddle.

Developers, scientific research institutes, enterprises and institutions, technical partners, hardware manufacturers, etc., are not only the builders of the flying paddle ecology, but also the beneficiaries.

Flying Paddle has built an all-round ecological system, where industry-university-research collaboration, co-creation, symbiosis, and win-win results are achieved.

Baidu CTO Wang Hai announced the latest developments in the flying paddle ecology

　　At present, the deep learning ecology continues to prosper, AI technology has made further breakthroughs, and industrial applications urgently need to reach a new level.

In this regard, Wang Haifeng pointed out that the deep learning platform plus the large model will penetrate the entire AI industry chain from hardware adaptation, model training, reasoning deployment to scene application, and consolidate the foundation of industrial intelligence, which will further accelerate the intelligent upgrade.

Large-scale model industrialization consolidates the foundation of digital-real fusion

　　The deep learning platform is a basic common platform. It is connected to the chip and the application, which plays a role of linking the past and the future. It is equivalent to the operating system in the intelligent era and strongly supports the intelligent upgrading of the industry.

　　The large model is an important direction for the development of artificial intelligence in recent years. It has the characteristics of good effect, strong generalization, and standardized R&D process, which brings new opportunities for the further development of artificial intelligence.

In addition, large models put forward higher requirements for the development, training, and inference deployment of deep learning models, leading to the development direction of deep learning platforms.

　　But at the same time, large-scale model development relies on comprehensive support of algorithms, computing power, and data, and it also faces a series of technical challenges at the application level: first, the data scale is large and the data quality is uneven; second, the model is large and the algorithm is difficult ; The third is the large scale of computing power and high performance requirements.

　　How to realize the industrialization of large models?

Wang Haifeng believes that enterprises with comprehensive advantages in algorithms, computing power, and data can encapsulate the complex process of model production, and provide large-scale model services for thousands of industries through a low-threshold, high-efficiency production platform, thereby forming a large-scale model industry. transformation path.

Baidu created a Wenxin industry-level knowledge enhancement model

　　This industrialization path has been verified in the industrial practice of the Wenxin model.

Based on the Flying Paddle platform, Baidu has built a Wenxin industrial-level knowledge enhancement large model, including a general large model represented by Pengcheng-Baidu Wenxin, a cross-modal large model, and a biological computing large model. Models, as well as toolkits that adapt to scene applications, large model APIs, products based on large models, and creative communities that explore ecological co-construction, etc.

"Let the landing of large models be as efficient as an assembly line"

　　At present, the Wenxin large model has been widely used in Internet products such as search, information flow, and smart speakers. All walks of life.

With the further expansion of application scenarios, Wenxin Big Model has jointly created more than 10 large-scale industry models, and continuously integrates and learns from the unique data and knowledge of industries and enterprises. The model capabilities are further enhanced, helping enterprises reduce costs and increase efficiency, and accelerate industry Upgrade.

　　With the rapid development of large models, the flying paddle deep learning platform that supports large model development, training, and inference deployment is also continuously evolving, with more significant advantages: unified development paradigm of dynamic and static, adaptive distributed architecture, load balancing of heterogeneous devices, etc. , to achieve flexible development and efficient training of large models; high-concurrency elastic service deployment, software-hardware collaborative sparse quantization acceleration, adaptive distillation and cutting, etc., to achieve efficient deployment.

　　In order to make the implementation of the large-scale model industry more efficient and convenient, Paddle provides a full-process industrialization tool and platform, including large-scale model development kits, scene model production lines, etc., which greatly reduces the application threshold.

Wang Haifeng pointed out that through efficient construction and rapid iteration of diverse scene models based on large models, the implementation of large models can be as efficient as an assembly line.

　　As Wang Haifeng said, the Flying Paddle Platform and the Wenxin Large Model, "Insist on technological breakthroughs and innovations, ecological cultivation and hard work, and consolidate the foundation of intelligence, so that every developer and all sectors of society committed to technological innovation and industrial development , are able to make great strides on the foundation of independence, strive to achieve high-level technological self-reliance and self-improvement, and promote high-quality economic and social development."