Recently, at the Huawei Developer Conference 2020 (Cloud) – HDC.Cloud, Huawei fully shared the achievements it has made in computer vision. Huawei opens up papers and algorithm code for global developers to break new ground in AI research, development, and deployment. In addition, Huawei released the Computer Vision Plan (referred to the Vision Plan) and invited global AI experts to participate in the research. The Atlas AI computing platform powered by Huawei Ascend AI processors will provide powerful computing to support this plan.
The research results will be implemented in Huawei’s MindSpore, the all-scenario AI computing framework. MindSpore is open to the industry, enabling global AI developers to continuously innovate, break through boundaries, and build pervasive intelligence.
Opening Up Huawei’s Basic Research Achievements in Computer Vision
Investing in basic research is an important part of Huawei’s AI strategy. Huawei has been tirelessly building basic capabilities in fields such as computer vision, natural language processing, and decision-making and inference. These capabilities are critical to developing efficient, green, secure, trustworthy, and autonomous machine learning.
Tian Qi, Chief Scientist of Computer Vision, Noah’s Ark Laboratory at Huawei, and IEEE Fellow, shared the latest research progress in computer vision. “Huawei has invested heavily in the basic research of computer vision with focus on data, knowledge, and models. In the past two years, Huawei has published more than 80 papers at leading AI conferences and journals such as CVPR, ICCV, NeurIPS, and ICLR. Huawei has made many groundbreaking achievements. These research results are opened to the industry in the forms of academic papers and source code. We invite global AI developers to research, develop, and deploy AI based on Huawei’s existing research results.”
The Vision Plan for Every AI Developer
Professor Tian Qi released the Huawei Computer Vision Plan. “Huawei will continue extensive investment in basic research, tackling the three fundamental obstacles to efficiently mine knowledge from massive data, design efficient models for visual recognition, and represent and store knowledge. Conquering the obstacles will help drive general intelligence.”
The Vision Plan incorporates six sub-plans. They are:
- Data Iceberg Plan: Use a small amount of annotated data to unleash the potential of massive unannotated data and support model training in small sample scenarios.
- Data Magic Cube Plan: Use multi-modal quantification, alignment, and fusion strategies to enhance the learning capability of models in real-world scenarios.
- Model High-touching Plan: Build large models on the cloud to explore the performance limits of various vision tasks.
- Mode Slimming Plan: Build efficient computing models on the device side to help various chips complete complex inference.
- Generic Vision Plan: Define vision pre-training tasks to build generalized vision models.
- V-R Integration Plan: Direct computer vision to real artificial intelligence via virtual-real integration.
At the same time, Professor Tian Qi said, “We welcome global AI researchers to join Huawei’s Vision Plan to innovate and explore the future together. The powerful computing capability of Huawei Atlas AI computing platform will accelerate the implementation of the Vision Plan. The research results will be fully implemented on Huawei’s all-scenario AI computing framework MindSpore and open to the industry, helping every AI developer.”
