Lei Chen - Deep Learning and Practice with MindSpore

AI has spread much faster and wider than any other general-purpose technology in history. There is little doubt that this new general-purpose technology will have a profound impact on every aspect of the economy and society. However, different industries and organizations have adopted AI at different rates: Traditional industries are still in the initial stages of adoption, whereas the Internet industry embraced it from the beginning. Although most industries are already using AI to some extent, they have not witnessed the expected changes, meaning that there is still huge room for improvement. Some industries have still not found an effective way to realize the value of AI.

In general, there are three main reasons for this. First, the development of AI applications and solutions is challenging, requiring an expert skill set that millions of IT engineers have yet to acquire. Second, unlike the Internet, which has a relatively simple environment, AI applications have complex operating environments and must adapt to meet a diverse range of requirements. For example, the various cloud, edge computing, and terminal device scenarios have different requirements, and AI applications must not only collaborate with existing production systems in various industries and related processes, but also integrate with traditional IT applications. Third, the diversity and complexity of applications in various industries bring new requirements, such as robustness, security, and interpretability, for AI algorithms and theories. AI has therefore undergone rapid, continuous improvement and adaptation, creating a range of challenges for new entrants wanting to master AI technologies.

Central to Huaweis AI strategy is the continuous investment in full-stack, all-scenario AI solutions with the aim of accelerating AI development. To this end, Huawei released its full-stack, all-scenario AI solution in October 2018, and officially unveiled MindSporean all-scenario AI computing frameworkin August 2019. Many people have asked why Huawei launched MindSpore despite there being many other frameworks available. The main considerations for investing in MindSpore at that time were as follows. First, no existing framework was able to support all scenarios seamlessly. Because Huawei services cover device, edge, and cloud scenarios, and privacy protection is becoming increasingly important, we need a computing framework that supports all scenarios to empower all-round AI. Second, we have learned from Huaweis AI applications and research over the past few years that the development of AI computing frameworks is far from converging. Exploratory research into technologies such as AI acceleration for just-in-time compilation (which is used in automatic differentiation and tensor calculation) and automatic parallelization for ultra-large neural networks, as well as new research areas such as deep probabilistic programming and graph neural networks, require further development of AI computing frameworks.