| name | closed-form-predictive-coding-hgf |
| description | Closed-form predictive coding via hierarchical Gaussian filters (HGF) methodology from arXiv:2605.20293. Restores precision-weighted prediction errors to predictive coding networks by expressing them as deep hierarchical Gaussian filters. Enables biologically plausible, local learning without backpropagation, with dynamic uncertainty estimates and Hebbian-compatible update rules. Activation: predictive coding, hierarchical Gaussian filter, free energy principle, precision-weighted prediction error, biologically plausible learning, HGF, predictive coding network. |
Closed-Form Predictive Coding via Hierarchical Gaussian Filters
Methodology from arXiv:2605.20293 (May 2026). Authors: Aleksandrs Baskakovs, Sylvain Estebe, Kenneth Enevoldsen, Kristoffer Nielbo, Chris Mathys, Nicolas Legrand.
Overview
Predictive coding (PC) offers a local and biologically grounded alternative to backpropagation in training artificial neural networks. However, current PC networks suffer from two key problems: they are slower than backpropagation, and performance degrades sharply as network depth increases.
This paper traces both problems to a single simplification: current PC networks fix the precision matrix to the identity, discarding precision-weighted prediction errors that the variational derivation requires. The authors close this gap by expressing predictive coding networks as deep hierarchical Gaussian filters (HGFs) and restore precision-weighted message passing.
Key insight: The HGF framework provides closed-form variational updates for all parameters — activations, weights, and precisions — under a single free-energy objective, requiring no global error signal, no iterative relaxation, and no automatic differentiation.
Core Methodology
1. Precision-Weighted Predictive Coding
Traditional predictive coding networks use fixed identity precision matrices, which loses the precision-weighting that makes PC theoretically grounded. This paper restores precision-weighting by expressing PC networks as hierarchical Gaussian filters.
2. Hierarchical Gaussian Filter (HGF) Framework
The HGF provides:
- Dynamic uncertainty estimates at every layer
- Hebbian-compatible update rules derived from closed-form variational inference
- Simultaneous learning of activations, weights, and precisions under a single free-energy objective
- No global error signal — all updates are local
- No iterative relaxation — inference resolves in closed form
- No automatic differentiation — all updates are analytic
3. Free-Energy Objective
The entire network optimizes a single free-energy objective:
- Prediction errors at each layer are precision-weighted (precision = inverse variance)
- Precision parameters are learned online alongside weights and activations
- This provides natural uncertainty quantification
Key Results
- FashionMNIST: Approaches backpropagation in epoch-level wall-clock cost while converging in fewer epochs
- Online learning: Outperforms backpropagation on online (streaming) tasks
- Data efficiency: Better performance with fewer training samples
- Concept drift: Superior adaptation to changing data distributions
- Depth scaling: Maintains performance as network depth increases (unlike previous PC networks)
Practical Implications
For Computational Neuroscience
- Provides a biologically plausible learning algorithm that rivals backpropagation
- Precision-weighting connects predictive coding to attention and uncertainty estimation in the brain
- Local Hebbian-like updates align with observed synaptic plasticity mechanisms
- Offers a testable framework for how cortical microcircuits implement precision-weighted prediction errors
For AI/ML
- A practical alternative to backpropagation for biologically inspired AI
- Natural uncertainty estimation built into the learning process
- Superior performance in streaming/online learning scenarios
- Better handling of non-stationary data distributions
When to Use This Skill
- When implementing predictive coding networks for biologically plausible learning
- When studying free-energy principle applications in neural networks
- When building models that require online learning or uncertainty estimation
- When comparing biologically motivated learning rules to backpropagation
- When working with hierarchical Gaussian filters for neural computation
Key Concepts
| Concept | Description |
|---|
| Predictive Coding (PC) | Neural network training via local prediction errors instead of global backpropagation |
| Precision Matrix | Inverse covariance matrix; weights prediction errors by their uncertainty |
| Hierarchical Gaussian Filter (HGF) | Multi-level Bayesian filtering framework with precision-weighted message passing |
| Free-Energy Objective | Single variational objective optimizing all parameters simultaneously |
| Hebbian-Compatible | Learning rules that involve local, pre/post-synaptic activity correlations |
| Precision-Weighted Prediction Error | Prediction error scaled by its estimated precision (inverse variance) |
References