Powered by Indico v3.0.3
Zoom : 536 870 8448
Multi- disciplinary Building： 228
Abstract:
The tRoplcal DEep-sea Neutrino Telescope (TRIDENT) is a next-generation neutrino detector located in the South China Sea. High computational efficiency is required for event reconstruction methods in order to calculate the incident particle's direction and energy. In a typical neutrino event, less than 1% of photosensors are hit, making Graph Neural Networks particularly well-suited for their reconstruction. In this study, a Graph Neural Network has been constructed to achieve high resolution in direction and energy reconstruction. This talk will present results from TridentNet and make comparisons with traditional reconstruction method.
Biography:
莫岑，上海交通大学粒子与核物理研究所2021级直博生，师从李亮教授。专注于高能物理实验，包括中微子信号的重建与模拟、大气缪子研究和中微子望远镜数据获取系统的开发，以及深度神经网络在实验上的应用。获“上海市和教育部重点实验室优秀学生（2023）”称号。
Training of supervised neural network models requires a large
amount of high-quality datasets with true values. In computer vision tasks such as object detection and image segmentation, the process of annotating a large number of original two-dimension data segments is extremely costly, which greatly affects the application rate of AI for HEP (High Energy Physics). The SAM(Segment Anything Model) based on
transformer provides a promising solution to this problem. This paper proposes an intelligent image segmentation annotation method based on the SAM, by which the annotation efficiency can be increased by 50 times. Examples of annotations, the API (Application Programming Interfaces), and GUI (Graphical User Interfaces) are also provided. The use of this tool will greatly accelerate the process of transforming highenergy physics image-style data from raw data to AI-Ready data.
近年来，深度学习方法的引入使得Jet tagging分类任务的准确率大幅提高，其中以ParticleNet为代表的图神经网络在该任务中表现出色，关于模型的部署，常用的硬件有CPU、GPU、FPGA、ASIC，目前，由于FPGA具有低功耗、低延迟、硬件可编程等特性成为AI部署加速的前沿研究热点，相对于CPU，FPGA可以实现更好的并行操作和更低的延迟，相对于GPU平台，使用FPGA可以降低功耗。因此，将ParticleNet在FPGA上移植并优化，以实现快速、低功耗的执行粒子物学理中的分类任务，从而减少经济成本并加快粒子物理中数据处理进程。