Asc18 students excel in ai reading comprehension task provided by microsoft gas x strips instructions

In the final round of the 2018 ASC Student Supercomputer Challenge (ASC18), the ShanghaiTech University team made of five students with an average age of 20 years old gave a strong performance in the AI reading comprehension task, winning ASC18’s e Prize Award. The team made improvements to model algorithm and the performance of training and completed the parallel model training of a large-scale dataset within 8 hours. The team’s prediction precision reached 46.46, close to the world cutting-edge level.

The e Prize Award aims to be the Gordon Bell Prize for young talents, encouraging young people to tackle challenges in supercomputing applications, challenge the limits of computing performance and apply supercomputing to make breakthroughs in science and engineering applications. The “e” in the name symbolizes the most important natural constant in science and the next target in supercomputer – exascale.

Teams competing for the ASC18 e Prize Award were presented with an AI reading comprehension challenge provided by Microsoft. The challenge required teams to independently develop algorithm models of machine reading comprehension and ask and answer with independently built supercomputer systems and the CNTK deep learning framework. Teams were also required to train the models with the latest supercomputer technology in combination with the MS MARCO dataset, in order to facilitate more accurate answers.

Developed by Microsoft, MS MARCO is a large scale dataset for machine reading comprehension and question answering based on real data collected from Bing and Cortana and consisting of 100,000 questions, 1 million paragraphs and over 200,000 file links. The dataset used in the ASC18 finals includes 100,000 manual-tagged questions with answers. With the benchmark code and a single node NVIDIA Tesla P100 GPU, training takes nearly a day with benchmark precision being 30 (ROUGE-L=30). In the final round, however, only 12 hours were allowed for training. To reach higher ROUGE-L, teams had to find optimum parameters as fast as possible by optimizing code performance in training clusters and innovating and improving the algorithm models.

First-time finalist ShanghaiTech University demonstrated exceptional capability to design and optimize AI models that wowed the judging panel, making notable improvements on data, algorithms and training methods. The team expanded the dataset provided by the organizing committee to involve more training samples, and reorganized answers by the simultaneous training of multiple tasks to obtain several excellent models. As a result, the team designed an advanced accelerated heterogeneous supercomputer system with Inspur’s AI server NF5280M5 servers and Tesla V100, and implemented the parallel model training with large-scale dataset including 100,000 samples within 8 hours. The team’s prediction precision reached 46.46, approaching the performance of the world’s most advanced algorithm, winning them the e Prize Award with the full allotment of points.

A Microsoft representative noted that given the knowledge structure and the optimization capability, it was truly remarkable for undergraduates to train a model with such high precision within 8 hours. Their success demonstrated the team’s extraordinary creativity, sophisticated understanding of AI models and the equally impressive hands-on skills.

Teams from Tsinghua University, the ASC18 Champion, and NTHU, winner of the Highest Linpack, also stood out with a prediction precision exceeding 40. The performance of the Tsinghua University team in particular was 3-30 times better than the benchmark code in training and data processing.

Over the past several months, participating teams have pored over the latest papers and focused on optimizing algorithms and performances. Their efforts and achievements were seen in the task of AI reading comprehension in the preliminaries to the final round. For most participants, it was their first time handling deep learning tasks. Several teams referred to the latest academic achievements such as S-NET and QANet. Through ASC, students deepened their understanding of ways to better integrate HPC and AI, reflecting ASC’s mission to provide a strong foundation for students’ future career planning and development, mitigate the current shortage of AI talents, and lay a sound foundation for AI to be implemented in production, life and society.

The ASC Student Supercomputer Challenge was initiated by China and is supported by experts and institutions from Asia, Europe and America. Through promoting exchanges and furthering the development of talented young minds in the field of supercomputing around the world, the ASC aims to improve applications and R&D capabilities of supercomputing and accelerate technological and industrial innovation. The first ASC challenge was launched in 2012. Since then, the competition has continued to grow into the largest supercomputer challenge with more than 1,100 teams and 5,500 young talents from around the world having participated. Jointly organized by the Asia Supercomputer Community, Inspur Group, and Nanchang University, ASC18 featured more than 300 teams from around the world.