Several models of the inflation predict a roughly perfect Gaussian primordial fluctuations which are generated by a single scalar quantum field in ground state. This is while there are alternative approaches to describe the inflationary period which lead to the non-Gaussianity of the CMB.
From this point of view, detecting non-Gaussian patterns in the CMB radiation plays a crucial role in confining cosmological models.
My primary aim in this research is to design a network which is able to indentify footprint of primordial non-Gaussianity in the simulated data fluctuations.
I used simulated data that are available here: http://dc.zah.uni-heidelberg.de/elsnersim/q/s/fixed
The mathematical background of data generation is available in this paper: https://ui.adsabs.harvard.edu/abs/2009ApJS..184..264E/abstract
According to the mathematical background, I considered three different amounts of f_NL and, therefore, my data set has three classes.
The utilized CNN is called Resnet18 that has 18 layers. Although the predefined input channel of mentioned network is 3, I altered it to 1 channel (it is possible to merge images using np.repeat to create a 3 channel input, but due to lack of free space on my drive, I preferred 1 channel case. The output of network will not change dramatically)
To prevent overfitting, I employed the cross-validation technique. As a result, the network reached a maximum of about 99% on validation accuracy.