Dubhe AI Open Source Platform is a shining star in ZJ Lab's history. Dubhe Platform went public in 2019, which initiated domestic new R&D institutes leading the R&D of high-performance basic R&D platform and promoting the independent and controllable innovation of AI core technology. In August 2020, Dubhe Platform Version 1.0 was officially released, featuring four core strengths namely developer-friendliness, efficient training, visual analysis and model refining. Since then, the platform maintains a three-month pace of version iteration. In less than a year, Dubhe Platform Version 2.0 was released in August 2021 as an open source, which greatly improved the ease of use, functional completeness and the richness of applications, increased the number of eco-partners to exceed 1,000, and further empowered developers from government agencies, universities and enterprises.

At the release conference of ZJ Lab’s Dubhe 3.0 and Dubhe 2022 Annual Developer Conference held today, ZJ Lab officially released Version 3.0 of ZJ Lab’s Dubhe AI Open Source Platform. This is the largest update iteration since Dubhe Platform was released as an open source. On the basis of greatly improving the overall performance of the platform, it focuses on the expansion of modules such as computing framework, algorithm model base, development suite and toolkit. The "new" is reflected in all aspects.

"Dubhe AI Open Source Platform is committed to providing developers with a more intelligent and more efficient AI basic platform,” said JIA Weiqiang, Technical Chief of Dubhe Platform Project and PI of ZJ Lab, "the newly added  'AI + computing' scientific computing framework is an independently developed algorithm model based on the integration of intelligent computing that is faster than conventional numerical solutions. So far, it enables highly efficient solutions in many scientific scenarios including hydromechanics, solid mechanics, molecular dynamics, thermodynamics and electromagnetics."

The updated version also includes some featured algorithms such as the 3D visual algorithm LoFTR, led by Prof. BAO Hujun, Vice President of ZJ Lab and Chief Architect of Dubhe Platform, a frontier algorithm targeting local image feature matching, and the neural rendering algorithms NeRF and NeuS that may rebuild 3D scenarios based on few images. According to JIA Weiqiang, Dubhe Platform’s algorithm model base has more than 150 algorithm models of eight categories, i.e. 2D vision, 3D vision, scientific computing, intelligent recommendation, natural language processing, intelligent audio processing, model compression and reinforcement learning, providing developers with a wider range of options.

Introduction of neural rendering algorithm NeRF

Demonstration of 3D vision algorithm LoFTR

In addition, a series of new toolkits, including intelligent data labeling, deep learning visualization, deep model refining, and data visualization analysis, provide whole-process efficient supports and services for developers to carry out data cleaning, analysis and decision-making, algorithm development, and model training.

"Model refining technology can obtain better performance models by recombining models with a small amount of labeled data, which helps developers reduce labeling dependence, reduce model waste, and realize on-demand, flexible and configurable model reuse." Prof. SONG Mingli of Zhejiang University shared the principles and methods of Dubhe Platform’s visual depth model refining technology at the Developer Conference. Prof. QIU Xipeng of Fudan University introduced the language model as a service of Dubhe Platform. Prof. XU Gang of Hangzhou Dianzi University demonstrated the highlights of Dubhe Platform’s deep learning visualization toolkit. Associated researcher XU Jiarong of Fudan University presented case studies on how Dubhe Platform’s graph neural network algorithm drives social relation computing. "Focusing on big data intelligent analysis and decision-making, Dubhe visual analysis toolkit integrates distributed analysis engine, intelligent analysis algorithm, all-stack analysis process and ‘turnkey’ system to help users complete data processing and analysis efficiently and assist decision making." Research Expert FU Siwei of ZJ Lab introduced the features of Dubhe visual analysis toolkit.

Dubhe Platform’s continuous evolution is most powered by the demands of scientific research and social and industrial development. At the meeting, representatives from the Committee of Political and Legislative Affairs of Yuhang District, China Telecom Zhejiang Branch, Hangzhou Sunyard Technology Co., Ltd. and Zhejiang Haishu Technology Co., Ltd. shared how Dubhe Platform supports intelligent governance at the basic level, digital reform, and intelligent financial system and postal system.

According to CHEN Zhefeng, Deputy Director of the Law Society of the Committee of Political and Legislative Affairs of Yuhang District, relying on the powerful computing framework of Dubhe Platform, the Committee of Political and Legislative Affairs of Yuhang District has built a grassroots intelligent governance brain and built a platform system with all-time, multi-dimensional, real-time dynamics sensing and long-term trend prediction. "Based on the core businesses of grassroots governance, we utilize Dubhe Platform’s intelligent risk scenario warning, uniform address matching, event intelligent flow, event handling evaluation, and other functional modules to implement the whole chain intelligent upgrade for complete event flow and handling. With the help of natural language processing and knowledge graph technologies, the knowledge and experience cases in the field of grassroots governance are solidified to achieve the accumulation and sharing of governance experience," said CHEN Zhefeng.

"After more than two years’ development and evolution, Dubhe Platform has rapidly accumulated massive data resources, further nursed and optimized AI model algorithms, promoted the evolution of deep learning, voice recognition, computer vision and other technologies at a faster pace, and promoted the innovative development of AI technology. In terms of the construction of Dubhe ecology, the Platform has provided industries with public fundamental R&D resources and makers with fast and convenient technical and data supports," said BAO Hujun, Vice President of ZJ Lab and Chief Architect of Dubhe Platform, "With the deepening of the R&D of platform and improvement of technical ecology, we hope in the future we can empower and cooperate in more fields to drive up the information industry and digital economy, build the eco-chain of the AI industry, promote and accelerate AI technical iterative innovation and application, and contribute more to realizing autonomous and controllable AI core technologies."

Dubhe AI Open Source Platform V3.0 is now available for download.

Developers may learn and download the latest version of Dubhe Platform through:

Dubhe Official Website:

http://tianshu.org.cn/ 

Project File:

http://docs.dubhe.ai/docs/ 

gitee：

https://gitee.com/zhijiangtianshu 

CodeLab:

http://codelab.org.cn/ 

FAQ:

http://docs.dubhe.ai/docs/troubleshooting