The ASC 2019 Student Supercomputer Challenge (ASC19) is underway, as more than 300 student teams from over 200 universities around the world tackle challenges in Single Image Super-Resolution (SISR), an artificial intelligence application during the two-month preliminary. Teams are required to design their own algorithms and train AI models using PyTorch, so as to use supercomputers to restore 80 blurred images back to high-resolution ones in the shortest possible time and meet set standards for similarity. Simple, effective and easy to use, PyTorch has quickly gained popularity in the open source community since its release and become the second most frequently used deep learning framework.

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. GAN can bring more and finer texture details and make pictures look more delicate, real and natural to the naked eye. Therefore, it has become a hot area of research in the field of image SR.

The difficulty of the SR task lies in that participating students who mostly major in computer science and math have to study extensive papers on SR technology and deep learning within two months to design their algorithms, complete AI model training on the supercomputer system, and optimize their algorithms continuously.

Furthermore, to meet the requirements and obtain better results, the students must take into consideration the trade-off between the distortion parameter and perception parameter. A paper published in CVPR 2018 elaborated an interesting “paradox” in SR technology, that is, the higher the similarity between the restored or reconstructed high-resolution image and the original image, the worse the definition observed by the naked eye, and vice versa. The reason behind it is the difference in the focuses of the distortion parameter and the perception parameter.

Ding Wenhua, Academician of the Chinese Academy of Engineering expressed a hope that the SR challenge can help students lay a solid foundation for deep learning, model training and optimization and promote the application of SR technology in more scenarios. Cheng Jian, representative of supporting organizations of the SR challenge, researcher of the Institute of Automation, Chinese Academy of Sciences, said that the development of AI has led to a surge in demand for computing. Training an image classification model requires exascale floating-point operations while fast, large scale image SR requires ever more computing. He also noted that hopefully the SR task can drive college students to better combine supercomputing with artificial intelligence and provide new ideas for SR technology application.

The ASC Student Supercomputer Challenge is the largest student supercomputer hackathon in the world, aiming to promote exchange and development among young supercomputing talents across countries and regions, improving their application level and research capabilities, taking advantage of the driving force of supercomputing to promote technological and industrial innovation. The preliminary round of ASC19 has begun, with more than 300 university teams tackling challenges for CESM, the Community Earth System Model for studying climate change; SR, single image super-resolution; and HPL and HPCG, internationally accepted HPC benchmarks. The top 20 teams in the prelims will move on to the finals at Dalian University of Technology from April 21st to 25th.Learn more ASC at https://www.asc-events.org/