ASC19 Student Supercomputer Challenge (ASC19) entered its breathtaking final week. From April 21-25, teams from 20 renowned universities around the world are taking part in the final round of the competition at Dalian University of Technology. They need to design and build a system combining high-performance computing (HPC) and artificial intelligence (AI) to flexibly respond to the challenges brought by traditional scientific computing and emerging AI computing.
ASC competition promotes HPC and AI integration
Jack Dongarra, ASC Advisory Committee Chair, Member of National Academy of Engineering, and Distinguished Professor at the Oak Ridge National Laboratory of the United States and the University of Tennessee, believes that high-performance computing is the driving force behind the continuous development of science, while artificial intelligence computing, represented by deep learning, ushers in a new field for high-performance computing, especially in the fields related to mathematics and computational mathematics. “I am very happy to see that ASC sets different tasks in HPC and AI fields,” he said. “This will allow students to think about how to integrate the two and learn to build a supercomputer system to complete different types of computational tasks. It will be very important and inspiring for their future careers.”
The trend of convergence between AI and HPC is accelerating. As we all know, the larger the amount of AI training data, the more accurate the labeling and the trained models. At the same time, the amount of calculation required also increases by tens of times. This is also an important reason why AI computing has quickly evolved from single-machine single-card to single-machine multi-card, multi-machine multi-card, and heterogeneous clusters. Technologies such as multi-machine parallel, high-speed low-latency networks, and scheduling algorithms in the HPC field can greatly reduce the burden of managing and using AI clusters. From the perspective of either economy or convenience, the integration of HPC and AI will be an inevitable trend. After all, when solving problems in different fields, a single system is better than two independent systems.
“At present, the connection between AI and HPC is still in a primary state,” said Professor Qian Depei, head of expert panel of HPC Key Project, National High-tech R&D Program. “As for AI, a lot of data, inference, and training tasks requires supercomputing. As for HPC, AI can be applied to new areas, for example, using AI for more precise weather forecasts. But in the long run, there may be a deeper correlation between AI and HPC. With the original intention as computing inspired by human intelligence, AI may have a crucial impact on future computing models.”
AI challenges promote cultivation of HPC+AI talents
Since 2016, ASC has been working with leading AI enterprises such as iFLYTEK, Baidu and Microsoft to bring the latest AI technology to the competition, such as intelligent voice DNN applications, traffic prediction applications and machine reading comprehension, allowing contestants to use supercomputers to optimize AI applications in parallel environments. This has helped establish students’ overall understanding of artificial intelligence from the perspective of infrastructure and computing power, and cultivate talents with HPC+AI capabilities.
A challenge from the ASC19 Finals was Face Super Resolution (FSR), a domain-specific super-resolution (SR) problem. One of the ultimate goals in FSR is to explore image intensity correspondences between low resolution (LR) and high resolution (HR) faces from large scale datasets and generate HR facial images closed to the ground truth HR face images. Super-Resolution (SR) technology is a visual computing technology that has received great attention in recent years, aiming to recover or reconstruct low-resolution images into high-resolution ones. As deep learning techniques, especially Generative Adversarial Networks (GAN), are introduced into SR research, this technology can be widely used in satellite and aerospace image analysis, medical image processing, compressed image/video enhancement and other applications.
In the final, the 20 college teams are asked to design/tune their SR algorithm design in the preliminary competition to achieve 4x upscaling for face images used in the challenge. In order to meet the requirements and achieve better results, on one hand, the students need to try to integrate HPC technology into the AI computing process to achieve higher computing performance; on the other hand, they also need to fully consider the identity similarity of face images when designing the model. A study published in ECCV in 2016 showed that the traditional measurements like peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) of SR images could not valuate the reconstructed face images well. In face recognition algorithms such as Facenet and SphereFace, the 512-D feature vector is used as the only feature. In ASC19 final, the feature similarity is the cosine similarity of the 512-D identity vector extracted from the HR face and the reconstructed SR image.
Liu Jun, a member of the ASC Organizing Committee and General Manager of AI & HPC, Inspur, believes that AI opens a new field for HPC applications and they are converging and reinforcing each other. “We hope that, the scientific computing and artificial intelligence challenges in the ASC competition will help college students from all over the world think, learn and try to integrate HPC with AI, thereby promoting the cultivation of HPC and AI talents, and easing the competition for talents between industry and academia. This is a very important mission of the ASC competition.”
About ASC
The ASC Student Supercomputer Challenge is the world’s largest student supercomputer competition, sponsored and organized by China and supported by Asian, European, and American experts and institutions. The main objectives of ASC are to encourage the exchange and training of young supercomputer talents from different countries, to improve supercomputer applications and R&D efforts, to boost the development of supercomputing technologies, and to promote technical and industrial innovations. The annual ASC Student Supercomputer Challenge was first held in 2012 and has since attracted over 7,000 undergraduates from all over the world. ASC19 is jointly organized by Asia Supercomputer Community, Inspur Group and Dalian University of Technology.