Skip to content

Online Generalised Predictive Coding

Home / Papers / Online Generalised Predictive Coding

Online Generalised Predictive Coding

Authors: Mehran H. Z. Bazargani, Szymon Urbas, Adeel Razi, Thomas Brendan Murphy, Karl Friston Date: 2026-05-04 Paper ID: openalex:2605.02675

Summary

This paper presents Online Dynamic Expectation Maximisation (ODEM), a framework for online data assimilation that jointly estimates latent states, model parameters, and uncertainty. By employing a separation of temporal scales, the method allows for fast Bayesian belief updating of hidden states while performing slower, iterative updates for parameters and precision. Numerical experiments with chaotic, non-linear systems confirm the method’s ability to track latent dynamics effectively, even when the generative model is misspecified. This approach provides a computationally efficient, neuro-mimetic solution for inference and learning in volatile, real-time environments.

Key Contributions

  • Introduces Online Dynamic Expectation Maximisation (ODEM), which extends the standard DEM framework for real-time online data assimilation.
  • Implements a temporal scale separation strategy that decouples fast Bayesian state inference from the slower updating of model parameters and precisions.
  • Demonstrates the robustness of ODEM in tracking latent states for non-linear and chaotic generative processes, even under model mismatch scenarios.

Open Questions & Future Work

Key Concepts

Archivist Review

I have approved the Online Dynamic Expectation Maximisation (ODEM) framework as a novel, reusable approach to real-time joint inference and the corresponding open question on non-stationary kernels. The candidate regarding learnable smoothness matrices was rejected as it represents a specific technical parameterization choice rather than a fundamental research bottleneck.

Approved Concepts

  • Online Dynamic Expectation Maximisation (ODEM): Introduces a novel online adaptation of the Dynamic Expectation Maximisation (DEM) framework for triple estimation (states, parameters, uncertainty) in dynamic environments.

Approved Open Questions

  • Non-stationary Kernels in GPC: Non-stationary kernels would significantly increase the flexibility of predictive coding models for real-world scenarios where system dynamics are non-homogeneous over time, a critical requirement for robust online inference.

Rejected Candidates

  • [open_question] Learnable Smoothness Matrices (learnable-smoothness-matrices) - subcomponent_of_broader_mechanism: This is a specific implementation detail regarding model parameterization rather than a foundational, high-level open research problem.

Metadata & Links

url
https://arxiv.org/abs/2605.02675
paper_id
2605.02675
paper_source
openalex
domain
time-series
tags
time-series-forecastingbayesian-inferenceonline-learning
architectures
datasets
concept_slugs
online-dynamic-expectation-maximisation-odem
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-07T06:05:01Z