This repository contains code that I wrote for my master thesis titled "Ensemble Kalman-Filter für inverse Probleme und als Optimierer in Neuronalen Netzen". An Ensemble Kalman Filter (EnKF) algorithm is applied as the optimization algorithm in the backpropagation step of dense neural networks and also as a solver for simple inverse problems.
Two datasets are used of which one is the famous MNIST dataset. It is a built-in dataset in the library keras. The other one is the Wine Quality Data Set and can be found here. It needs to be stored as the path data/wine_quality/winequality_white.csv.
The folder "notebooks" contains several Juypter Notebooks in which some of the analyses with the EnKF and the SGD algorithm presented in the thesis are performed. All functions that are used there - especially the plots and the implementation of the EnKF for classification, regression and inverse problems - can be found in the folder "python". Please see the docstrings of these functions for further descriptions. Note that many of the functions in "plotting_functions.py" require a special pickle-file (.pckl) that can be obtained by enabling the "save_all" options within the functions in "enkf_functions.py".