From imaging algorithms to quantum methods Seminar

Monday 25 Nov 2024, 10:00 → 11:30 Europe/Warsaw

https://cern.zoom.us/j/66151941204?pwd=n7upvvZYibexBhbtyn5kvTpy36L0Wo.1 (Zoom)

Konrad Klimaszewski (NCBJ), Wojciech Krzemien (NCBJ)

Zoom link: https://cern.zoom.us/j/66151941204?pwd=n7upvvZYibexBhbtyn5kvTpy36L0Wo.1

Title: Review of generative deep learning for fast simulations
Presenter: Aleksander Ogonowski
Date: 25.11 2024

Participants:
Rafał Mozdzonek(RM)
Michal Mazurek(MM)
Dima Melnychuk (DM)
Aurelien Coussat(AC)
Mateusz Kmiec (MK)
Wojciech Krzemień (WK)
Konrad Klimaszewski (KK)
Michał Obara (MO)
Aleksander Ogonowski (AO)
Arkadiusz Ćwiek (ACw)
Roman Shopa (RS)
Lech Raczynski (LR)
Tomek Fruboes (TF)
Mateusz Bała (MB)

Questions/Remarks:

WK: Is it one particle (e.g. one photon) per event?
AO/MM: It is one particle per event that generates e.g. showers
KK: It makes sense to have one particle cause the goal is to have generative models. One can later combine the outputs as he wants.

MM: It is important to note that CaloShowerGAN is used in ATLAS production MC

WK: In normalizing flow is the z latent space coordinate? What is the difference between x and z?
AO: No it is just a confusion of notation with the previous page, the
equation corresponds to the next step transformation.    

KK: What is caloDREAM?
AO: Conditional Flow Matching which is very similar to Normalized Flows but has to be understood better

KK: Calo-VQ what is the representation of those vectors in embedding space?
AO: Embedding vectors have the same length as the dimensions of the latent space representation. The number of embedding vectors is predetermined - they are quantized. Each latent vector is assigned to the closest embedding vector.

KK: What is the conditional embedding and how is it introduced into the ResnetBlock?
AO: The kinematics of the particle. I don't know precisely how it is incorporated into the ResNet block. To be checked.

KK: Diffusion model - at which stage the noise is applied?
AO: We must train step by step e.g. between two images which are different
by the level of noise injected.

LR: What is more important AUC or processing time? What is your goal?
AO: There are many metrics. It must be faster than GEANT4, but the accuracy
should be "the best possible"
MM: It is difficult to say. The idea is to provide a variety of models to give a choice to the analyst.
KK: For some use cases you don't need a lot of quality (e.g. background studies). For others, one needs to go with the best available within  the time budget.  

MM: What are the next steps for you?
AO: I don't know yet :-).