| name | exponential-family-predictive-coding |
| description | Extended predictive coding framework using exponential family distributions beyond Gaussian assumptions. Reveals biological neural network properties: nonlinearity, heterogeneity, biological plausibility. Maintains FEP-PC correspondence up to second cumulant. Derives biologically plausible local plasticity rules from EFD variational free energy. Use when: predictive coding, free energy principle, exponential family, variational inference, biological plausibility, local plasticity rules, neural heterogeneity, non-negative firing rates. arXiv: 2605.30882 |
Extended Predictive Coding with Exponential Family Distributions
Paper: Extended predictive coding framework as variational free-energy minimisation under exponential-family assumption
arXiv: 2605.30882
Authors: Asaki Kataoka, Kenji Doya
Category: q-bio.NC
Published: 2026-05-29
Core Concept
The Free-Energy Principle (FEP) → Predictive Coding (PC) correspondence has traditionally been limited to Gaussian assumptions with Laplace approximation. This paper extends the framework to the Exponential Family of Distributions (EFD), revealing biological properties previously missing:
Missing Properties Captured by EFD Extension
- Nonlinearity of input-output properties within neural networks
- Heterogeneity — different neurons have different response characteristics
- Biological plausibility — no negative firing rates (Gaussian PC allows negative rates)
Key Result
The FEP-PC correspondence is maintained up to the second cumulant of the posterior distribution when EFD is assumed for both variational posterior and prior.
Mathematical Framework
Traditional Gaussian PC Limitation
- Assumes Gaussian posterior and prior
- Uses Laplace approximation (matches only first two moments)
- Results in linear, homogeneous networks with potentially negative firing rates
EFD Extension
- Uses exponential family: p(x|θ) = h(x) exp(η(θ)·T(x) - A(θ))
- Natural parameters η(θ) and sufficient statistics T(x)
- Captures skewness, kurtosis, and other higher-order moments
- Maintains FEP-PC correspondence through second cumulant
Reusable Patterns
Pattern 1: Biologically Plausible Local Plasticity Rules
- The EFD-based PC model can be trained using local plasticity rules
- Derived from EFD variational free energy gradient:
Δw_ij ∝ E_q[∂log q/∂w_ij · (log p(x|z) + log p(z) - log q(z))]
- Each synapse updates based on local info (pre-synaptic activity, post-synaptic prediction error)
- Hebbian-like: Δw ∝ pre × error — no global error signal required
Pattern 2: Heterogeneous Network Design
- Different neurons can have different distributional assumptions
- This creates heterogeneous input-output properties within the same network
- More biologically realistic than homogeneous Gaussian networks
Pattern 3: Nonlinear Predictive Coding Layers
- Replace standard linear PC layers with EFD-based nonlinear layers
- Use the natural parameter space for prediction error computation
- Sufficient statistics become the nonlinear activation functions
Pattern 4: Cumulant-Based Approximation
- Track prediction errors through cumulants (mean, variance, skewness, kurtosis)
- Higher cumulants capture non-Gaussian structure in neural representations
- Truncate at second cumulant for computational efficiency while preserving key properties
EFD Family Selection Guide
See references/efd-family-selection-guide.md for complete comparison table, selection heuristics, and local plasticity derivation.
Quick reference:
| EFD Family | Best For |
|---|
| Poisson | Spike counts |
| Gamma | Positive continuous (firing rates) |
| Beta | Proportions/bounded variables |
| Von Mises | Circular (orientation tuning) |
Implementation Guidance
- For neural network design: Use EFD activations instead of ReLU/sigmoid
- For learning rules: Implement local prediction error minimization
- For variational inference: Use EFD families (Gamma, Beta, Poisson) instead of Gaussian
- For biological modeling: Map natural parameters to neural membrane potentials
Pitfalls
- The EFD extension maintains FEP-PC correspondence only up to second cumulant — higher cumulants require additional terms
- Not all exponential families are equally suitable — choose based on the data type (count data → Poisson, proportions → Beta, positive continuous → Gamma)
- Local plasticity rules derived from EFD may require careful initialization to avoid divergence
- Local plasticity converges to different solutions than backpropagation — don't expect identical benchmark performance
- The Gaussian PC is NOT wrong — it is a special case of the EFD framework
Connections to Existing Skills
predictive-coding-light: Base PC framework — this extends beyond Gaussian assumption
feedback-hebbian-continual-learning: Local learning rules — complementary to EFD local plasticity
free-energy-moe-routing: Free energy principle applications — shared FEP foundation
predictive-coding-exponential-family-plasticity: Sub-skill focused on implementation details of local plasticity rules
Related Skills (Consolidation Note)
Multiple skills cover arXiv:2605.30882 — this is the umbrella skill:
exponential-family-predictive-coding ← this skill, umbrella
extended-predictive-coding-exponential-family — duplicate, should be consolidated into this
extended-predictive-coding-free-energy-exponential-family — duplicate in neuroscience/
predictive-coding-exponential-family-plasticity — focused sub-skill on local plasticity implementation