Neural Network code for the MNIST dataset and different interations of it.
Several runs with one hidden layer of size 64 and training of 3 epochs.
Used sigmoid activation function and mean squared error for error.
Accuracy ~93%.
Same setup for layers.
Instead uses softmax and cross-entropy for error.
Accuracy ~96%.