Currently I am an assistant professor at the Internet Technology and Data Science Lab (IDLab), Ghent University - imec, Belgium. I'm co-leading the Text-to-Knowledge research cluster with prof. Chris Develder, where we work on natural language processing in general, with focus on information extraction for applications in the media and medical domain. Some of our recent work includes multi-task information extraction on the document level, and augmenting clinical predition models with text-based features.
Besides sequence modeling with neural networks, I'm particularly interested in combining knowledge and neural network models. During a research visit in 2016 at UCL NLP lab at the University College London, I worked on injecting first-order logic into neural link prediction models. Together with the DTAI research group at KULeuven, I have been working on the combination of neural networks and probabilistic logic programs. These research topics fit within my general interest in joining deep learning models with systems better suited to reason.
Since I'm interested in AI in general, I'm having fun with a couple of side tracks. In combining my deep learning experience with my electrical engineering background, I worked on system identification from unequally spaced time series. Some new work is coming up on time series in the clinical domain, using techniques from complexity science. Also in the clinical domain, we came across some consistent misuse of data augmentation techniques, and tried to explain the issues to the medical AI community in this work.
In 2005, I received my M.Sc. degree in electrical engineering at Ghent University, after finishing my final year and master thesis at ETH Zurich, Switzerland, where I worked as a student assistent. In 2009, funded by a grant from the Research Foundation - Flanders (FWO), I obtained my Ph.D. at the Ghent University Department of Information Technology, under the supervision of prof. Daniel De Zutter, in the area of computational electromagnetics. Shortly afterwards, I got involved in research on Information Retrieval, in collaboration with the Database Group at the University of Twente in The Netherlands, and from there my interests moved to Natural Language Processing.
Provides a simple and natural extension of RNNs (in particular the GRU) towards unevenly sampled time series,
for modeling stationary MIMO systems.
It doesn't always need to be NLP ;)
Proposes models with predefined sparseness in embeddings and recurrent layers. Predefined sparse RNNs are designed such that they are equivalent to several small dense RNNs in parallel, sharing inputs. Word embeddings can made "sparse" up front by reducing effective embedding sizes for less frequent words.
Presents techniques for combining neural networks and probabilistic logic programs, by introducing neural predicates at the input of a ProbLog module and calculating loss gradients at the interface between both, to be used for backpropagation through the neural networks.