Poster Presentation

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Probabilistic World Models Provide a Shared Account for Learning from Surprise, Schemas, and Imagination

Shrey Dixit¹, Caswell Barry², Christian F. Doeller¹, Andrej Bicanski¹; ¹Max Planck Institute for Human Cognitive and Brain Sciences, ²University College London

Presenter: Shrey Dixit

According to predictive coding, the brain predicts sensory inputs and updates an internal world model based on prediction errors. In discrete settings, a world model can generate predictions for a finite set of possible observations. This becomes challenging in continuous environments, where observations cannot be enumerated. Stochasticity exacerbates this problem because the same observation may lead to multiple plausible futures, without a single ``best'' prediction. We introduce a probabilistic world model that generalizes predictive coding to continuous, stochastic settings and identify potential correlates in the brain. The model maximizes the likelihood of the observed data instead of minimizing pointwise prediction error. Crucially, we model the environment’s probability distribution using normalizing flows. Empirical data suggest that probability computation may be mapped onto the medial prefrontal cortex (mPFC), which has been implicated in schema formation and abstract task representations. In this algorithmic mapping, the probabilistic world model corresponds to the hippocampal system, here modeled as an RNN that encodes temporal context and interacts with the mPFC-like component. The fact that the model can quantify the likelihood of an episode can then be used to guide the generation of novel imagined episodes. The model learns sequence probabilities in a self-supervised manner without a biologically implausible reconstruction loss and provides a shared account for learning from surprise, schemas, and imagination.

Topic Area: Memory, Learning & Knowledge Structures