On the evening of December 25, the 5th @ World "Touching Academic Frontiers" international exchange event of the SCS was held as scheduled in the conference room at the corner of the 2nd floor. At this event, two outstanding students from the SCS, Ye Yiwen and Wang Haoyang, were invited to share their current developments and latest research findings in the hot fields of 3D medical image segmentation and graph neural network accelerators.
The theme of Ye Yiwen's report is a universal 3D segmentation model for medical images based on prompt learning. 3D medical image segmentation has long been perplexed with the challenge of multiple tasks and difficult annotation, which has led to current methods mostly training a proprietary model on different tasks. These models, however, often face the problem of insufficient training due to limited data. In addition, different models require separate training and optimization, which incurs huge costs. In this report, Ye Yiwen focuses on how to build a universal 3D segmentation model for medical images, sharing her insight into how to use prompt learning to achieve a high-precision, low optimization cost, multi domain and multimodal universal segmentation model, and exploring the opportunities and challenges of future universal segmentation models with the attending classmates.
Wang Haoyang shared his experience of attending the ISCAS2023 conference and briefed on the paper on graph neural network accelerator. Graph neural network (GNN) operations involve a large amount of irregular data operations and sparse matrix multiplication, leading to inefficient utilization of computing resources. With regard to large-scale graph training, it can become more complex and challenging. GNN training based on mini-batch is an effective solution. However, the current GNN operation accelerator does not support small batch structures. This paper proposes some methods to analyze the operational characteristics of GNN from multiple aspects, while considering the acceleration requirements during GNN training and inference processes, and suggesting the SaGNN system structure. SaGNN offers multiple working modes to provide acceleration solutions for different GNN frameworks, while ensuring system configurability and scalability. Compared to related work, SaGNN has brought a 5.0-fold improvement in system performance.
The @ World "Touching the Academic Frontiers" international exchange activity has attracted the active participation of many students, and aroused the strong interest and thirst for knowledge among students in the SCS.
Through this activity, students not only got the chance to communicate face to face with outstanding senior students, but also gained a deeper understanding of the latest research achievements and cutting-edge technologies in the field of computer science. At the same time, the activity also provided a platform to showcase their talents and abilities, and to inspire their innovative thinking and academic enthusiasm. In the future, the SCS will continue to hold a series of activities to provide students with more academic exchange opportunities and help them grow into excellent computer talents with an international perspective.
