Alexander van Meegen

Theory of Neural Networks

I’m a scientist who is fascinated by neural networks—both the biological ones inside our head as well as their artificial counterparts powering AI. These networks work amazingly well (for the most part). However, we still have a rather shallow understanding of the underlying principles. The goal of my research is to address this challenge and to develop a solid theoretical understanding of the principles underlying biological and artificial neural networks. The main tools to achieve this goal are methods from theoretical physics.

PostDoc

Currently, I’m a PostDoc at EPFL Lausanne in Wulfram Gerstner’s Lab. In extensive collaborations with various of the amazing members of the team, I work on a range of topics from loss landscapes to recurrent networks and self-supervised learning. Prior, I was a Swartz Fellow at the Center for Brain Science at Harvard University working with Haim Sompolinsky. With Haim, I developed a theory of supervised learning deep in the feature learning regime, which enabled a thorough understanding of the learned representations—from the network level down to the level of single neurons.

PhD

During my PhD in computational neuroscience in the Theoretical Neuroanatomy Group at INM-6, Research Centre Jülich, I worked with Moritz Helias and Sacha van Albada. My PhD research focused on the link between the single-neuron level and the collective dynamics of neural networks. I approached this problem along different lines of attack: I developed and employed methods from statistical physics to gain analytical insights into the dynamics and I constructed and simulated biologically constrained, large-scale models of cortical networks. Conceptually, I tried to keep a tight balance between the overwhelming neurobiological complexity and the simplifications that are inherent in modeling. In particular, I tried to neither create a “map of the empire that […] coincides with it point for point” (Borges, On Exactitude in Science; see also Abbott, 2008) nor examine oversimplified systems that neglect all of the relevant complexity.