Choose timezone
Your profile timezone:
CIDeR-ML General Meeting
→
America/Los_Angeles
Description
https://u-tokyo-ac-jp.zoom.us/j/83932834349
Minutes:
Quick recap
This meeting was a project update session where team members presented their current work progress. Tituan shared improvements to his siren model training, including the creation of a 3D lookup table for PMT acceptance correction and addressing overfitting issues through second derivative corrections. Omar reported merging Refactor V2 with production codes and adding neutrino telescope functionality to Lucid. Riya discussed working on systematically fixing gradient problems in the Lucid code. Ryotaro provided an update on focusing on the ADD dataset from the Rio de Janeiro experiment. Yifan presented progress on addressing BDX distribution biasing issues and setting up losses for multiple events. Zhenxiong updated the team on code repository updates, validation work, and running the calibration pipeline with COSMIC data. The conversation ended with Kazuhiro agreeing to chair next week's Super K meeting.
Next steps
- Kazuhiro: Chair and share the meeting next week in Patrick's absence.
- Omar: Run the notebooks again and clean up documentation and functions before finalizing the merge to main branch.
- Riya: Continue systematically addressing the Lucid code gradient problem as discussed with Omar.
- Ryotaro: Continue checking and analyzing the LED experimental dataset.
- Titouan: Investigate ways to reduce model runtime
- Yifan: Continue working on resolving memory-related issues in the current model implementation.
- Zhenxiong: Perform validation to ensure the new code in the branch does not impact previous training and calibration, then split changes into separate pull requests for each feature and submit for review.
- Zhenxiong: Check the status of reconstructed variables and performance on real data before proceeding with siren tuning.
Summary
pmtSIREN Model Training Progress Updates
Titouan presented updates on his model training progress, focusing on improving performance using a 3D lookup table for PMT acceptance correction. He addressed overfitting issues by implementing a correction based on the second derivative and shared the resulting model performance. He also discussed runtime analysis of his model, noting that while 80 milliseconds is acceptable, it is slightly higher than Nahuel's waterSIREN.
LUCiD Refactor-V2 Production Merge Update
Omar reported that he merged Refactor V2 with Cesar's production code branch, making it close to becoming the main branch. The remaining work includes running notebooks and cleaning up documentation and functions. Omar also mentioned adding a new mode for neutrino telescopes, including Ice Cube functionality in Lucid, which will be merged soon. Riya expressed interest in learning more about this new feature.
Mie Scattering Gradient Fix Progress Update and LEDs
Riya reported working on systematically changing code to fix a gradient problem in the Lucid codebase following a discussion with Omar. Ryotaro indicated no progress on LED implementation but is focusing on experimental data for the Rio de Janeiro dataset.
dE/dx Distribution Biasing Updates
Yifan discussed updates on addressing dE/dx distribution biasing issues and implementing changes to propagate and collapse distribution after the recombination stage. They are also working on setting up loss functions for multiple events and dealing with memory-related challenges.
OpticSIREN
Zhenxiong updated the code repository with a new branch and plans to validate recent changes while running the calibration pipeline with COSMIC data. Patrick and Zhenxiong discussed the status of reconstructed variables and performance testing on real data, which had not yet been completed. Zhenxiong planned to double-check the impact on existing code before making a pull request, with Patrick suggesting to split the changes into separate pull requests for different features.
Kazuhiro agreed to chair next week's Super-K meeting in Patrick's absence.
