Meeting Minutes

Meeting Minutes === ###### tags: `Salting` `Meeting` `CDMS` `AI` `ML` :::info - **Location:** https://ucdenver.zoom.us/j/97994358165 - **Date:** Dec 11, 2023, 9 AM Mountain - **Participants:** - Amy Roberts (AR) - Zack Kromer - **Meeting Goal(s)** ::: ## Summary  ## Notes  - Previously, Sukee had the network look at each channel individually and then took an "AND" to determine if an event was a signal event - But this is different from what we did in analysis. For the analysis, we looked at the sum of the four channels. So Sukee looked at training on pT, the sum of the phonon channels, and the results weren't great - Now Sukee is training the network on all the channels (A, B, C, D) at the same time. So she's handing (4096, 4) to the CNN rather than (4096, 1). - This does much better, when spiking the background with 40 ADC units the signal identification just starts to decrease from 100%. Whereas with pT, signal identification for 40 ADC units has already decreased to 40% - Using this new model identifies 29 events that were classified as background by the Soudan cut, but the network identifies as signal. Do we trust that these are really signal? - Farnoush: wonder if looking in the frequency space might improve classification. Sukee: I did try this, but I don't totally understand these results - Sukee is trying to get an understanding of what the model does/looks for. Farnoush warns that the model's job is to distinguish, so while it's good to develop intuition it's tough to beat the model