Meetings - philippeller/freeDOM GitHub Wiki
20. May 2020
- Aaron implemented a GA+simulated annealing optimizer, which requests 250 llh evaluations in batch mode. using that he can reco an event in ~500-700 ms. Resolutions look already quite reasonable (except for ~10% of events where the minimization failed, those could be outside of where our NN is valid, needs to be investigated).
- Time has some outliers, and especially cascade energy has a larger tail on the positive side (fitted energy too high).
- it's plausible these limitations are from the likelihood itself. Charge net being the suspect
- ways to improve
chargenet
:- train a per-DOM
chargenet
to assess how good we can get - train a per-string or per-z-layer
chargnet
- include more training data
- include additional info like number of hits
- get rid of charge net and include charge in
hitnet
and add some complicated/undefined things to make it work
- train a per-DOM
- Jan did studies with events that have the same
MCTree
(i.e. same particles from the interaction), but separate photon propagation = different hit patterns in the detector. 1-d scans show reasonable behaviour. Could also use those to assess validity of estimated reco uncertainties. - ...
07. May 2020
Initial Meeting, participants: Doug, Aaron, Justin (PSU), Jan (JGU), Martin, Philipp (TUM)
Ideas and suggestions that were discussed:
- use raw waveforms instead of pulses. Was decided to first go with pulses as it is a known entity, but keep waveforms in mind. One comment that was raised: ADC digitizer bins are no longer independent/uncorrelated
- Resuse parts from the retro framework where it makes sense, or think about merging the two
- Directly sampling the posterior (MCMC, HMC, Multinest, ...)
- First step: use SRTTW pulses and CRS minimizer and DeepCore events, compare to known recos. We may not expect too much of an improvement in resolutions, but in speed
- Should be rel. straight forward to then use the same technique applied to ICU
- Validity range of parameters should be assessed and corresponding constraints used in the minimization process. Could restrict ourselves to a smaller parameter space for the beginning to keep things simple
What I meant here was to intentionally shrink the parameter space so that we could study the impact of exceeding the validity range, using a more easily understood system. Looking instead at (say) events at very low energy, or with a vertex outside the detector's physical edge, will make it harder to disentangle the impact of the validity range from other factors. –Doug
- PSU has and will likely get even better Deep Learning optimized compute resources: accounts could be requested if needed by Doug
Tasks:
We decided it would be a good idea to separate the likelihood part (NN) from the exploration part (minimizer/sampler)
- PSU wants to take care of the latter, i.e. provide a general reco framework that takes any likelihood function as input
- Jan will use his experience of validating tables-based lieklihoods to validate the NN likelihood. For example compare to repeated simulation of a given event
- Philipp to keep working on the NN training, i.e. providing the likelihood function
Next meeting: Week of May 18, need to find suitable time