Deep neural networks exploit composition hierarchies of data, which makes it particularly suiting for finding trends in complex accounting records. This project will explore the potential of deep neural networks for automating big accounting data, particularly: (1) Generative adversarial networks to generate accounting data including invoices. (2) Image segmentation for invoices for quality assurance. (3) Pattern recognition for invoice classification. The candidate will specifically work with automated data generation to increase the accuracy of pattern recognition tasks. The candidate is expected to explore existing techniques, specifically in combinations of UBL and EHF, and develop and verify new variants of deep neural networks with real customer data. The candidate will develop a fully functional prototype. Goal 1: Automatic generation of accounting data for text using deep generative adversarial network. Goal 2: Automatic detection of crucial information from invoices including account number, recipient, and so on using deep convolutional neural networks. Goal 3: Automatic classification of invoices using a combination for real data and generated data (from goal 1).
Project leader: Robert Kristiansen