Choose timezone
Your profile timezone:
CIDeR-ML General Meeting
→
America/Los_Angeles
Description
https://u-tokyo-ac-jp.zoom.us/j/83932834349
Recording
Minutes:
Quick recap
The team reviewed results from Monte Carlo simulations of cosmic muon tracks and discussed improvements in charge deposition predictions, though some issues with over-training and data limitations were identified. They explored various methods for data processing and randomization, including the use of Poisson loss and sequential randomization techniques, while also addressing the need for better training data and charge weighting implementation. The group examined LED run data and charge distributions, discussed trigger calibration requirements, and considered ways to improve predictions near detector edges through various optimization approaches.
Next steps
- Ka Ming: Implement inclusion of horizontal-going muon events into the fine-tuning dataset and show results next week.
- Ka Ming: Investigate and implement a global light yield (fetch factor) as an optimizable parameter in the fine-tuning process.
- Ka Ming: Implement randomization (or sequential randomization) of batches during training to avoid repeating patterns at epoch boundaries.
- Ryotaro: Categorize PMTs by 4 types and refine the fitting range/algorithm for charge distribution, including consulting with the MPMT group on proper multi-PE fitting functions.
- Ryotaro: Pick one PE gain for each MPMT type, check data quality, and resolve issues with Monte Carlo job submission (including consulting with Ka Ming if needed).
- Ryotaro: Ensure calibration data spans the full operation period and covers both software and hardware trigger runs to account for time variations and detector performance changes.
- Zhenxiong: Contact Ka Ming regarding collaboration on fine-tuning/validation tasks.
Summary
Cosmic Muon Track Simulation Improvements
Ka Ming presented results from fine-tuning simulations of cosmic muon tracks using Monte Carlo data. The simulations showed improvements in charge deposition predictions after tuning, but also revealed issues with over-training and lack of data for top PMTs. Kazuhiro suggested exploring the use of a global light yield factor to address the non-uniform efficiency problem. The team discussed the need to consider PMT saturation effects and improve the training data by including more horizontal muon tracks. Patrick inquired about the implementation of charge weighting in the loss function, which Ka Ming confirmed was not yet included. The group agreed to follow up on these issues and consider ways to improve the training process for better predictions near the detector edges.
Data Randomization and Trigger Calibration
The team discussed randomization methods for data processing, with Kazuhiro explaining that sequential randomization is faster than complete randomization while still providing adequate results. They explored the use of Poisson loss and its potential issues with large charge variations, agreeing that normalization might be necessary. Ryotaro presented findings on LED run data and charge distributions, receiving feedback from Kazuhiro and Ka Ming about fitting multiple PE levels for better accuracy. The group also addressed the differences between software and hardware triggers, with Patrick explaining the need for calibration data throughout the entire WCTE operation period to account for detector performance variations.
