## Title: Introducing a Condtional 3D Dataset for Generative Models¶
## Presenter: Aleskander Ogonowski
## Date: 24.11 2025
## Participants:
Wojciech Krzemień (WK)
Konrad Klimaszewski (KK)
Roman Shopa (RS)
Lech Raczynski (LR)
Aurelien Coussat (AC)
Aleksander Ogonowski (AO)
Mikołaj Mrozowski (MM)
Arkadiusz Ćwiek (AĆ)
Krzysztof Nawrocki (KN)
Michał Obara (MO)
Varvara Batozskaya (VB)
AC: What is the meaning of the term energy when using the Wasserstein distance?
WK (I looked it up later): The Wasserstein distance is related to the optimal transport problem. As a very simplified (and not fully precise) example you can imagine that the two distibutions p(x) and q(x), we compare, correspond to some distribution of the "mass" or dirt on some space x, and you define a cost function (or the energy you use to move it), that allow you to move the dirt (mass) e.g. for p(x) around in such a way, that it becomes a q(x) distribution. There are many ways of going from p(x) to q(x), and for each, there is some cost (or energy used). Then, among all those "ways", one chooses the one with the smallest cost (energy used) and treats it as a distance between the p(x) and q(x) original distributions.
RS: Those symbols in the table are hiragana not kanji, why are they so badly written? So that the character "re" (れ) resembles "ki" (き) in the penultimate row?
KK: Those are real handwritten input data
KN: Maybe you could try to use some software like OCR) to recognise generated symbols and translate them. To be able ot quantify it.
KK: Somehow KID metric does it, but it is a good idea.
KN: Second question concerning your plans -> Mamba? Mamba is not necessarily better in the sense of classification, but rather because of the performance.
AO: I checked only Mamba at GitHub.
RS: Slide 13. How is the choice of refinements for GUN, e.g. auxiliary critic loss and gradient penalties, made?
AO: I added it step by step from the simplest GAN architecture.
RS: The question is, rather, what was the motivation for every step, such as using the Wasserstein distance? Is the architecture based on references or randomly built?
KK: It is explained in our paper. The ideas are taken from some papers: Wasserstein GAN.
The main originality: 3-D. The way the data are provided to the model (regression parameter etc).
WK: Can you use it with the Calo data?
AO: In principle yes. We will see how it will perform.
WK: What about the trade-off between precision and performance? My understaning is that the final goal is to have something precise enough to replace MC.
AO: We will concentrate on GAN improvement. Diffusion models are slow, some of them slower than MC event.
WK: Is there any standard limit in precision - in the context of physics that a given model must fullfil to be accepted?
KK: My understanding is that for the Calo challenege we will leave the choice for the future user.At some moment, one should add some comparision at the more physics level.
R.S: About those Hiragana (Kuzushiji-MNIST dataset). I asked Gemini and it confirmed that it was deliberately built from real handwritten samples of Kuzushiji Hiragana, not polished calligraphy. But deciphering is much easier in the context that character-by-character. Can GAN-based ML work together with text recognition tools that account for context, e.g. as in the recent KN presentation?
KN:
There are minutes attached to this event.
Show them.
