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
Quick recap
The meeting focused on updates and discussions regarding the angular response analysis using SIREN and cosmic data, as well as the performance evaluation of different SIREN models for PMT waveform prediction. Zhenxiong presented updates on generating angular responses using SIREN, comparing methods with cosmic analysis and diffused LED data, and noted that while SIREN results were less sensitive to delamination than LED data, they still showed differences from cosmic analysis. Junjie shared detailed findings on the performance of three different SIREN models (single MLP PCA, dual PCA, and quantile models) trained on various input/output formats, highlighting that while models performed well for peak visibility regions, they showed biases in high and low visibility tails, particularly near PMT surfaces. The team discussed potential solutions including regularization on the omega parameter and using Poisson loss instead of Huber loss to improve model performance.
Next steps
Zhenxiong
- Send the draft of the WCTE presentation to Patrick de by the end of tomorrow.
- Try the "shell method" with the cosmic analysis to compare with current results.
Junjie
- Test using Poisson loss instead of Huber loss for Siren training, especially for low-visibility voxels.
- Investigate applying regularization on the omega parameter in Siren, possibly in a voxel or distance-dependent way, to address overfitting issues.
Summary
Angular Response Analysis Updates
Zhenxiong presented updates on the Angular response analysis using cosmic data from WCTE, comparing pre-tuned and post-tuned SIREN models with cosmic analysis and diffused LED data. The analysis showed that while SIREN results were similar to cosmic analysis and not sensitive to delamination, they still differed from the diffused LED results, particularly in the low cosine theta range. Zhenxiong noted challenges with the shell method due to boundary effects and detector curvature, and planned to present these findings at WCTE for further discussion.
Siren Performance Analysis Review
Junjie presented a detailed analysis of sirenTV's performance across three different scientific models with varying input/output methods and architectures. The models showed generally good performance in predicting visibility and waveform quantiles, with the dual PCA model performing best overall. However, significant biases were identified in both high and low visibility regions near PMTs, particularly within 20 centimeters of the wall, and in early quantile bins. Junjie proposed potential solutions including applying regularization to the omega parameter and using Poisson loss instead of Huber loss to better handle low visibility voxels. The team discussed the possibility of exploring these improvements further and agreed to review the detailed bias analysis plots for potential applications to WC Siren studies.