Made By Nikita Kogan and David Abramov.

As part of the Intro to Machine Learning course we took during our 2nd year at Bar Ilan University, we were requested to implement a
Fully-Connected Neural Network in order to learn and predict labels of items in the "Fashion-MNIST" dataset (given to us in a pair of files to learn on, train_x (learning items) and train_y (learning labels) and a file (test_x) with items we had to output labels of (test_y, the file which is generated by our python code). The dataset consists of 28X28 single-value pixel images. There are 10 possible labels for each image, representing 10 diffetent fashion items.

After exploring different hyper-paramters (such as learning rate, number of epochs in the learning phase and the amount of hidden layers
and their size), we achieved a (k = 5, 80%(learning set)-20%(validation set)), cross validation verified, over 85& success rate.
The values we found giving this rate are:
learning rate = 0.01
one hidden layer sized 100
and 50 epcochs during the learning phase

A PDF report of our work is also included.