In this small project (bonus exercise sheet for the course of Machine Learning and Physics) the objective was to build an autoencoder using various two different sub-architectures, the first one being MLPs and the second one being CNNs. As a dataset we got a set of jet pictures from professot Tilman Plehn and we had to come up with a series of observations on how the two architectures that we came up with were able to reconstruct the images that were given as inputs.
I still want to put some effort in this project, even though the due date is already over, because I want to understand a couple of things that didn't really sit right with me, here is an approximative roadmap of the process:
- Clean the code and make it easier to understand
- Understand why the CNN has lower accuracy than the MLP substructure I found a way to make sure that the CNN has higher accuracy than the MLP substructure but it takes 5 times to run and the ROC curve is kind of a weird result.
- Make the Convolutional Neural Network more efficient.
Even though our solution was not totally complete it was deemed close to perfect, thus it's possible for anyone to see it as a tutorial that is very close to beginner level (bragging, but still down to earth).