PhD Neuroscience

Learning should leverage previous knowledge, rather than overwrite it. This process is natural in animals including humans, but remains a challenge in artificial systems, highlighting our knowledge frontier in lifelong learning. The main problem amounts to maintaining stable representations/knowledge of previous tasks (stability) while acquiring new information (plasticity). How animals solve this stability/plasticity tradeoff problem remains unknown and is rarely investigated over long periods of time.

Paradoxically, recent studies reveal that neural representations in the hippocampus change gradually over time while supporting stable behaviors, a phenomenon known as representational drift. The functional role of this drift remains debated: it may serve to separate similar experiences across time, promote generalization and continual learning, or simply reflect underlying biological dynamics such as synaptic turnover. Understanding whether drift is an adaptive computational mechanism or a byproduct of synaptic plasticity could transform our view of how the brain learns across the lifespan. While neural activity sheds a light on learned representations in biological systems, studying synaptic connections through learning in live animals remains a daunting task. As artificial neural networks share key properties with biological ones, they provide a novel and useful model to investigate some of these questions. Thus, a cross-disciplinary study of live animals and artificial models provides a powerful window into learning and memory.

In this project, you will combine computational modelling with in vivo neurophysiology in freely behaving mice to uncover how drift supports learning and memory. Specifically, you will train deep learning models on perceptual tasks to evaluate continual learning in novel classes of biologically inspired neural networks that implement representational drift. In parallel, you will record hippocampal population dynamics longitudinally using miniaturized calcium imaging in freely behaving mice during learning to identify neural signatures of drift and their behavioral relevance. By establishing a link between representational drift and continual learning, we aim to advance our understanding of hippocampal function, inform the design of artificial systems capable of lifelong learning while bridging neuroscience and AI.