To successfully control the body our brain needs to generate stable states (i.e. keep a hand somewhere), but also full dynamic behaviours (i.e. a full arm swing). The brain does this highly efficiently through networks of spiking neurons - the efficiency and success of which is yet to be reproduced in AI. Crucially, neural activity during dynamic tasks is often observed to lie on lower-dimensional manifolds. Why this is useful, and how networks generate and adapt such activity manifolds, is the subject of many theoretical models but still very much an open question. In this NWO-funded 4-year PhD project, using tools from both AI and neuroscience, you will develop neural networks which produce such low-dimensional manifolds and then use these to learn efficient and effective control. This project is expected to yield more capable AI systems and further our understanding of brain computations.

Dependent on your background and interests, the project can either go deep into applying the resulting models to AI control problems, or reproduce and predict neuroscience experiments. You will be supervised by Dr. Sander Keemink and Prof. Marcel van Gerven, and advised by Dr. Nasir Ahmad. The project is hosted at the AI department of the Donders institute, where there are additional opportunities to collaborate with experts in neuroscience, robotics, AI, and other related disciplines.