experimenting with my neural network library, puffpastry, to generate mnist numbers.
so far, the results are not very good. there is a lot of noise caused by the loss function performing badly with empty images.
with target number = 2 over 200 iterations:
