Marc de Kamps is a Lecturer in the School of Computing at the University of Leeds. After a PhD in particle physics, he moved to the Section Cognitive Psychology of University of Leiden where he was involved in research on visual attention and neural representation of language. He became a research associate at the Chair for Robotics and Embedded Systems at the Technical University of Munich where he helped coordinated EU funded Thematic Network nEUro-IT.net. Subsequently, he moved to the University of Leeds to become a lecturer in the School of Computing. From 2015-2020 he was PI in the EU funded Human Brain Project. He is a member of the Leeds Institute for Data Analytics (LIDA) and collaborates with LIDA colleagues and Turing Fellows Mark Gilthorpe and Peter Tennant on the interface between machine learning and causal inference. Through LIDA, he also (co-)supervises several PhD students in UKRI Centre for Doctoral Training in Artificial Intelligence for Medical Diagnosis and Care.
In the broadest sense, Marc is interested in how humans represent the outside world. This is not just an epistemological question, but also a practical one with potential ramifications for computer science and artificial intelligence. Animals are still far better at making sense of what they see around them than any deep network. This is probably due to neural representations carrying intrinsic meaning: a visual neuron is a visual neuron because it is in a pathway between optical receptors and higher cognitive function. As trite as that observation may be, it means that neural representations carry meaning in a way that bit strings do not; it also means that one of the fundamental operations of computer science, the copying of symbols, is not available in neuronal architectures. That is somewhat puzzling as language processing and high level reasoning very much seem to require symbol processing and if one carefully looks at the historical development of the notion of computation, Turing and others developed it on what they knew humans could do. Our proposal is that the brain temporarily links different brain areas, as and when needed, and the neural blackboard architecture is a concrete proposal for how a neural substrate may support this way of representing knowledge.
Specifically, the requirement of modelling neural activity in a biological way has led me to develop mean-field techniques for modelling the activity of groups of spiking neurons. These techniques allow the inclusion of biologically realistic point model neurons in mean-field population models that can be used to construct large-scale models of neural circuits. This in turn has led to a collaboration with Samit Chakrabarty (Biological Sciences, Leeds) to use neural models of spinal cord to help analyse EMG muscle activity generated data. I also have a strong interest in generative models, particularly normal flows and variational autoencoders and I am interested in the question of whether these techniques can be used in conjunction with more conventional techniques used for causal inference.