Reproduce “The generalized Hierarchical Gaussian Filter” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Lilian Aline Weber, Peter Thestrup Waade, Nicolas Legrand, Anna Hedvig Møller, Klaas Enno Stephan, Christoph Mathys

Researchers at Osnabrück University, University of Zurich & ETH Zurich, and Aarhus University generalized the Hierarchical Gaussian Filter (HGF) by incorporating linear and nonlinear value coupling and dynamic observation noise, providing approximate Bayes-optimal update equations and a modular network architecture. This extension unifies HGF with principles from Predictive Coding, enhancing its ability to model diverse hierarchical dependencies and uncertainties for applications in computational psychiatry.

Key insights

  • The generalized HGF is conceptualized as a modular network of belief nodes, each performing generic prediction, update, and prediction error steps, simplifying the construction of complex hierarchical models.
  • The unified message-passing scheme in the generalized HGF requires the bottom-up transmission of lower-level precision estimates (alongside prediction errors) for accurate updating of higher-level belief precisions.
  • The framework successfully unifies key computational principles of the HGF (modeling volatility and uncertainty) with those of Predictive Coding (value-based hierarchical predictions), offering a more comprehensive Bayesian brain model.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.