World Models in Brains and Machines: From Normative Hippocampus Models to Robust Planning in Artificial Intelligence and Back

Ziel

A hallmark of intelligence is the ability to mentally simulate future scenarios and plan actions accordingly. This cognitive ability requires an internal model of the world. How the brain constructs, and uses such a world model is a fundamental, yet unresolved question. Normative models in neuroscience can explain the emergence of abstract representations from sequential observations, but their rigid formulation restricts them to static toy models. In artificial intelligence, world model based agents have matured significantly. However, despite impressive successes, leading state-of-the-art agents fail at long-horizon navigation and, as recent findings show, their internal dynamics do not match the structured representations observed in the brain, suggesting a fundamental architectural gap.
This project addresses world model learning in both brains and machines upfront with a cross-disciplinary approach that explains failure modes of state-of-the-art agents, overcomes their limitation with a novel brain-inspired architecture, and finally validates artificial world models against a unique dataset of brain activity recorded from mice navigating 11-different environments.
Key deliverables are a) a precise characterization of inductive biases required for well-structured internal representations, b) a new model class that pushes the frontiers for world model based reinforcement learning and that serves as powerful normative model for brain dynamics across contexts, and c) an in-depth analysis of hippocampal activity to reveal how the brain creates and reuses internal representations.
This project establishes a virtuous NeuroAI cycle: Insights from neuroscience will help to build more robust and interpretable AI, which in turn will serve as a powerful computational framework to deepen our understanding of how the brain builds its model of the world.

Wissenschaftliches Gebiet (EuroSciVoc)

Dieses Projekt wurde noch nicht bei EuroSciVoc klassifiziert.
Schlagen Sie die Wissenschaftsbereiche vor, die Ihrer Einschätzung nach besonders relevant sind, und helfen Sie uns, unseren Klassifizierungsdienst zu verbessern.