Justify way of data generation in experiments

[svenvanderburg]

I think we should maybe justify the rather complicated way we do training data generation. There is 2 things we should show I think: 1) that balancing improves the test score 2) that seeing a different set of training examples every epoch improves the test score.