Study of a learning method of multi-step deep learning model in particle physics experiment
-
- SAITO Masahiko
- International Center for Elementary Particle Physics, The University of Tokyo Institute for AI and Beyond, The University of Tokyo
-
- MORINAGA Masahiro
- International Center for Elementary Particle Physics, The University of Tokyo Institute for AI and Beyond, The University of Tokyo
-
- GANGULY Sanmay
- International Center for Elementary Particle Physics, The University of Tokyo Institute for AI and Beyond, The University of Tokyo
-
- KISHIMOTO Tomoe
- International Center for Elementary Particle Physics, The University of Tokyo High Energy Accelerator Research Organization Institute for AI and Beyond, The University of Tokyo
-
- TANAKA Junichi
- International Center for Elementary Particle Physics, The University of Tokyo Institute for AI and Beyond, The University of Tokyo
Bibliographic Information
- Other Title
-
- 素粒子物理実験における多段深層学習モデルの学習
Abstract
<p>In particle physics experiments, for the data processing, several steps are required from raw experimental data to statistical analysis. In recent years, deep learning has been used in each step, contributing to the improvement of data analysis. If each deep learning model can be connected at once, its simultaneous training is expected to improve the performance of the last step, that is, results of the statistical analysis. In this talk, we will discuss how to connect models and learn them simultaneously. We will show that (1) propagating the loss function for each step through an MLP mitigates degradation of the last step's performance, and (2) deep learning models having multiple loss functions can be effectively trained by applying techniques of multi-task learning.</p>
Journal
-
- Proceedings of the Annual Conference of JSAI
-
Proceedings of the Annual Conference of JSAI JSAI2022 (0), 3N4GS1001-3N4GS1001, 2022
The Japanese Society for Artificial Intelligence
- Tweet
Details 詳細情報について
-
- CRID
- 1390855656055932928
-
- Text Lang
- ja
-
- Data Source
-
- JaLC
-
- Abstract License Flag
- Disallowed