Files
Jacob_32821900_2024.pdf
Open access - Adobe PDF
- 1.17 MB
Details
- Supervisors
- Degree label
- Abstract
- In the context of classification, neural networks are generally considered as state-of-the-art in terms of performance. Yet, these black-box models suffer from a lack of interpretability. This can however be alleviated through a careful design of the network's architecture. Such an approach has been followed in recent years to create the Decision Rules Network (DRNet) model. This neural network handles binary data and can be interpreted as a simple set of decision rules. In this thesis, we develop extensions of the DRNet model, with the aim of seamlessly handling datasets of various types and properties. This is achieved by adding a new layer to the network, which integrates the data binarisation closely with the rest of the learning process. At the same time, we wish to maintain a high degree of accuracy in our predictions, and favour rule sets that are composed of few, short rules. These are conflicting objectives, as high accuracy levels generally require more intricate decision rules, and conversely. Through our experiments, we find that the accuracy of our extended model remains high, and identify ways in which to further improve its simplicity.
Citations
Jacob, M. (2024). Neurosymbolic AI: combining deep learning and rule learning. https://dial-mem.test.bib.ucl.ac.be/handle/2078.2/38305