A fundamental question in neuroscience is why the hippocampus, critical for episodic memory in humans, primarily exhibits spatial representations like place cells in many species. This talk proposes that spatial coding emerges naturally when episodic memory systems are engaged in navigation tasks. We demonstrate this using a novel computational model based on memory-augmented neural networks. The model autonomously learns to store and retrieve information from an external memory buffer and shows remarkable flexibility: When solving visual tasks in supervised learning, it develops representations of categories and visual features. When learning to navigate a simulated maze based only on visual inputs using reinforcement learning, it spontaneously develops population-level representations that capture the maze’s 2D spatial structure. These findings suggest that spatial representations in the hippocampus are not primary, but rather emerge as a consequence of using episodic memory systems to solve spatial problems, reconciling two major lines of research on the hippocampus that have largely remained separate until now. 

About the Speaker:

Sen Cheng’s group investigates neural mechanisms underlying learning and memory. Focusing on the hippocampus – the brain region involved in episodic memory, as well as in the learning and memory of sequences, Professor Cheng’s research focuses on as yet unexplored dynamics of these processes. (Read further at https://www.ini.rub.de)

Certificate of attendance: Please contact team assistant serenella.brinati.1(at)hu-berlin.de

This invited talk is hosted by SFB1315 Deputy Speaker Richard Kempter (subprojects A01, B01).

Lecture Poster

Series Overview