pyhgf: A neural network library for predictive coding

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Bayesian models are increasingly used in neuroscience and AI for understanding cognition.
• Predictive coding theory explains learning and behavior through hierarchical probabilistic inference.
• Implementing these models in standard libraries faces challenges due to software limitations.

Purpose of the Study:

• Introduce pyhgf, a Python package for creating and manipulating dynamic networks for predictive coding.
• Overcome limitations of existing libraries by providing a transparent and modular framework.
• Enable biologically realistic and adaptable autonomous agents.

Main Methods:

• Developed pyhgf using JAX and Rust for efficient computation and differentiation.
• Enclosed network components as transparent, modular variables for flexible message-passing.
• Implemented arbitrary algorithms via belief propagation and dynamic network adaptation.

Main Results:

• pyhgf facilitates the creation of dynamic networks for predictive coding.
• The framework supports self-organization, structure learning, and meta-learning.
• Differentiable functions seamlessly integrate into sampling and optimization workflows.
• Generalized Bayesian filtering and hierarchical Gaussian filter are included examples.

Conclusions:

• pyhgf offers a powerful and flexible tool for developing advanced AI and computational neuroscience models.
• The package promotes biologically plausible and adaptable autonomous agents.
• It enhances the implementation of predictive coding by addressing software development challenges.