A simple Neural network for classifying hand written images from MNIST in java.
The program is written completely from scratch with no help from library's.
The best result I have achieved is around 98.6% accuracy on the test data with a design of {784,800,10} neurons.
I have implemented learning techniquessuch as L2 regularisation, cross entropy loss function and expansion of the training data using a multitude of methods (rotations, squishing/stretching and thinning/thickening).
In the WB_folder there are sets of weights and biases which can be read to run the program or it can be trained given some training data.
Any help on how the design can be improved is appreciated!