This Machine Learning project implements the Expectation Maximization algorithm using Bernoulli mixes (1-32 mixes were tested) in order to classify digits.
Firstly, a train + test dataset is given on the mnist_all.mat file.
By running bernoullimix3.m file you train the dataset for all bernoulli mixes.
If you run draw_digits.m after that for different values of Kplot (number of mixes), you will see what we have learned from the test dataset for each mix (larger K means a larger pool of different forms for every digit).
Lastly, if you run the test_digits.m file you will get the final results (avg error percentage for every mix and for every digit/mix) on a results.txt file.
The code was executed in Matlab (version: R2011a).
Here you can find the presentation I did of this project for the Machine Learning Course, during my MSc (AUEB) in 2014.