I recently have been learning about something called a Generative adversarial network (GAN) which basically two neural networks competing them each other, with the first creating an image and the 2nd determining whether that image is fake or not. It’s a bit more complicated than that, but that’s the general idea.
My main goal is to create a GAN for galaxies, specifically those in the Green Valley, which is the type of galaxy that I like the most. I have tried making my own images of galaxies but there is still lots of room for improvement, shown below. Though I have to admit this progress is acceptable. It LOOKS like it could be a high redshift galaxy, so that part is realistic. Though It would be much more amazing to recreate Hubble level images, but that may be too big of a goal. I haven’t learned all the techniques involved in create high resolution images, but I’ve started reading a book on GANS so hopefully I can learn more from it.
As you can see above, the neural networks learn over time how to create convincing images of galaxies over time. This was over 12 hours of training, and it looked like I may have been plateauing. I need some more practice with something not as difficult.
Cosmalo's Blog 