Looking back on the sixty-year history of artificial intelligence, the iteration of machine learning models has dominated the development of artificial intelligence technologies. Today, technologies like deep learning have taken a prominent position among artificial intelligence technologies, and deep learning frameworks have even been the core of the entire system of artificial intelligence technologies. The Dubhe Artificial Intelligence Open Source Platform developed by Zhejiang Lab focuses on the innovation of core technologies and takes advantage of its differentiating performance to stand out above other platforms. Bao Hujun, the deputy director of Zhejiang Lab and the chief architect of the platform said that “Dubhe will build up four core advantages: an independently developed high-performance core computing framework; a full-featured AI development kit; AI model integration and free deployment across edge, cloud, and terminal; and an intelligent collaboration runtime system.”
The independently developed performance computing framework of Dubhe is the most essential advantage of the platform which provides features such as concurrency, automatic orchestration and implementation, high-performance efficiency and stability. "Computing frameworks are just like the foundation of a house and the high-performance core computing framework of Dubhe can lay a powerful and stable foundation for developers,” said Dr. Shan Haijun, a platform architect of Dubhe. To take face recognition for example, the platform supports data concurrency, model concurrency, and mixed concurrency, and has successfully run model training on datasets of over ten million face images in the security and defense area. The resource usage rate is better than the average performance of existing frameworks.
For artificial intelligence developers, the development platform provided by Dubhe has a full-featured AI development kit and AI model integration services, which will greatly boost the research and development efficiency of artificial intelligence technologies. “You need the right tool for your job, and the AI development kit and AI model integration services on Dubhe are like a premium set of architecture design tools and developers can build their models using these AI development kits or use the AI models provided by the platform to build their own,” said Shan Haijun. Currently, the AI development kit on Dubhe provides features such as automatic data labeling and processing, model building, automatic machine learning and tuning, and one-click deployment. With these features, even a beginner can swiftly learn everything he needs to begin developing algorithm models on the platform. The set of artificial intelligent analysis models not only covers multiple types of models and examples for tasks such as image classifying, object detection, and super-resolution, but provides some unique cutting-edge algorithms for features like semantic video slicing and video behavior recognition.
Besides, the model knowledge generating technology provided by Dubhe has made self-adaptive algorithms possible. To take image recognition for example, current video monitoring technologies require different models to achieve pedestrian detection and tracking in the morning and at night. The model knowledge generating technology can integrate the two models to work both in the morning and at night, effectively enabling the same algorithm model to be used in more scenarios.
After model design and training, we will need to get our AI models up and running. For model deployment, Dubhe enables developers to freely deploy their tasks on terminal, edge, and cloud devices, and the runtime platform supports intelligent allocation and collaboration to ensure the performance of AI models.
The artificial intelligence open-source platform is the core of the entire system of AI technologies and serves as a bridge between the application of algorithms and low-level hardware. "We will carry out research on artificial intelligence software and hardware collaboration to solve the technological bottleneck in platform compatibility and performance and fundamentally improve the computing efficiency of the platform,” said Bao Hujun.