FCC-hh Calo meeting 04.12.:
attended: Anna, Coralie, Jan, Michele
connected: Clement, Jana
News from DNN studies (Jan)
- short introduction in the updates/optimisations on DNN models, using ResNET like building blocks to avoid over-training
- Tests on single pion performance:
- PU0 and PU200 look good, without electronics noise
- performance including electronics noise has been tested, without re-training the network!!!, and fails
- New training not feasible for CDR, downtime of GPU machine
Update of single pion performance for topo-clusters on EM scale — with electronics noise (Jana)
- Including the linearity plots to the results shown for different threshold settings —> smallest non-linearity for default 4-2-0
- Comparison to cells on EM scale within a cone of R<0.4 around the generated particles —> including the cone selection on the topo-clusters :
- Before cut, best resolution with 5-2-0
- After cut, best performance for 4-2-0
- Next steps, PU scenarios
There are minutes attached to this event.
Show them.