A representative example of a cosmology image (left). Setting all pixels with higher-than-average brightness to white reveals a complex structure that is undetectable by means of visual inspection (right).
Semester project for the Computational Intelligence Lab 2020 at ETHZ
The first part of this project is to build a generative model that is able to produce realistic cosmology images. The provided dataset contains 10800 gray-scales images with 1000x1000 pixels resolution. The second objective is to repurpose the aforementioned generative model and use it to predict so-called "score-values": A function that measures the similarity of a cosmologic image to a "prototypically ideal" image is defined on a large subset of the data. Finally, a set of unseen query images needs to be scored and submitted to the Kaggle competition.
GitHub repository: Repo
Report PDF: Report
