Jenia Jitsev

Panel Member

Jenia Jitsev is heading the Cross-Sectional Team "Deep Learning" (CST-DL) at the Jülich Supercomputing Center (JSC). His research focuses on mechanisms of plasticity and learning in biological and artificial neural networks. Working in the overlap of computational neuroscience and machine learning, he seeks to shed light on learning as a generic process of building up a valid model of unknown phenomena from its incremental observations and how it is implemented in neural circuitry of the brain.

He studied computer science and psychology at the University Bonn, with focus on neuroscience and machine learning. He then obtained his PhD in computer science at the Frankfurt Institute for Advanced Studies (FIAS) and the University Frankfurt under supervision of Prof. Christoph von der Malsburg, working on plasticity and unsupervised learning in recurrent hierarchical neural networks of the visual cortex, with applications to object and face recognition. Further stations of his research were Max-Planck Institute for Neurological Research in Cologne and Institute for Neuroscience and Medicine (INM-6) in Jülich, where he studied models of plasticity, unsupervised and reinforcement learning in the brain's basal ganglia network and in cortico-basal ganglia loops. For his work on reinforcement learning in the basal ganglia, he received Best Paper Award from International Society of Neural Networks.

The long-term aim of his current work is a draft for generic end-to-end deep learning neural architecture that is capable of both unsupervised and reinforcement learning. This research line will require understanding how both forward generative and inverse inference model can be learned from self-generated and real observation data, which in turn has parallels to theories of predictive coding in brain circuits and active Bayesian inference. Ultimately, this line of research can give rise to autonomous systems that have the ability to improve how they learn ("learning to learn") and also can perform active learning - search for and query, or generate, data from those sources that are most promising for successful learning of given tasks.