We discussed regarding the impossibility of creating non-gaussian distributions using independent samples in GAN. Paras suggested we could try some sort of recurrent network to try and correlate the samples with one another.
We also discussed the possibility of using a purely regression model which helps us find the inverse of the cumulative function, We could use such an approximate function along with binning to be able to generate our own distribution using uniform distribution.
One of the possible directions to take is to perform a statistical analysis where we could try to prove the impossibility of non-gaussian distributions using independent samples in GAN.
It was noted that, while trying to simplify the original problem we have simplified it to the point of trying to generate single parameter non-gaussian distributions. This seems to be a very different problem and is different from the initial problem of trying to simulate tracks in CORSIKA.
We discussed regarding CONEX and 2 additional approaches were noted.
- To consider this to be a supersampling problem, and try to perform simulations with lower resolution and then use GAN to improve the resolution. Similar to the work done here: https://research.fb.com/publications/neural-supersampling-for-real-time-rendering/
- Another approach was to consider the data as some form of time series (where height is the parameter for us) and perform recurrent GAN. An approach used in this paper (https://arxiv.org/pdf/1706.02633.pdf), uses GAN for generating time series data. Such an approach could be used on CORSIKA data.
