Choose timezone
Your profile timezone:
CIDeR-ML General Meeting
→
America/Los_Angeles
Description
https://u-tokyo-ac-jp.zoom.us/j/83932834349
Link to Recording
Minutes:
Quick recap
This meeting was a team check-in following a workshop and holiday period, where team members provided updates on their recent work. Titouan presented improvements to a SIREN network model for PMT acceptance correction, sharing results of model performance optimization including adjustments to omega values and addressing normalization issues, with the model now showing 5% maximum error. To Titouan reported success in reducing computation time to approximately one millisecond per iteration while working on lowering training losses, though validation loss remained around 0.01 with some overfitting concerns. Other team members including Zhenxiong, Junjie, and Ryotaro reported no updates, while Yifan mentioned ongoing work on position-related tasks and basic checks.
Next steps
- Nahuel: send a message to Omar to confirm if the current one millisecond per iteration step time is acceptable for the Adam optimizer step
- Zhenxiong: commit documented work and relevant notebooks/scripts to the repository
Summary
Team Progress
Patrick opened the meeting by welcoming participants after the holiday and workshop, asking for reports from team members. Zhenxiong reported having no updates this week. Patrick suggested documenting and committing the work done since the last workshop to a repository, to which Zhenxiong confirmed she will do. Junjie had no additional report to share.
SIREN Neural Network Performance Issues
Titouan discussed his work on training a SIREN neural network to learn corrections for PMT acceptance based on incident angle and hit position. He mentioned having performance issues with his models and was investigating what might be causing them during the Golden Week. Adjusting omega helped with overfitting. He resolved normalization issues in the visualization and achieved a maximum relative error of around 5%, training in full 3D. Will continue to investigate detailed structures in 2D error plots.
Nahuel reported on improving computation time to about one millisecond per iteration while working on the SIREN, though the model showed some overfitting with validation loss not decreasing below 0.01. He will investigate some potential reasons for the overfitting, such as still too complex central region, issue from random sampling to line sample (try random sampling with new true function), etc. The team agreed to check with Omar regarding the one-millisecond requirement for iterations and discussed the validation set approach used by Nahuel.
