During her internship, Aline van Driessche examined the possibility to generate fake but realistic looking invoices out of thin air. If this works, the fake data could be used to improve our machine learning model for invoice recognition, without any privacy concerns attached. Although the field is barely touched in research, this would be the holy grail in terms of invoice recognition.
Fearlessly, Aline jumped at this quest as a modern day Don Quixote, armed with a decent knowledge of generative adversarial neural networks (GANs). Unfortunately she has not found the holy grail (yet), but she was able to generate some interesting results non the less, including an impressive 16/20 as a final score for her thesis from the master in AI at the KULeuven.
A more detailed look on her approach and findings can be found on her blog on the matter.