| name | extended-predictive-coding-exponential-family |
| description | Extended Predictive Coding framework using exponential-family distributions for variational free-energy minimization. Captures biological network properties: nonlinearity, heterogeneity, positive firing rates. Biologically plausible local plasticity rules. Activation: predictive coding, exponential family, free-energy principle, variational inference, local plasticity, 预测编码, 自由能原理. |
Extended Predictive Coding under Exponential-Family Assumption
Source: arXiv:2605.30882 | Submitted: 2026-05-29
Authors: Asaki Kataoka, Kenji Doya
Category: q-bio.NC (Neurons and Cognition)
Core Contribution
Extends Free-Energy Principle (FEP) and Predictive Coding (PC) from Gaussian assumption to Exponential Family Distributions (EFD), capturing biological neural network properties:
- Nonlinearity of neural responses
- Heterogeneity of input-output properties
- Positive firing rates (biological constraint)
Problem with Gaussian Assumption
Traditional FEP-PC correspondence under Gaussian + Laplace approximation:
- Linear input-output relationships
- Homogeneous neuron populations
- Negative firing rates (biologically implausible)
- Limited explanatory power for sensory cortex dynamics
Solution: Exponential Family Distributions
EFD includes:
- Bernoulli (binary neurons)
- Poisson (spiking neurons)
- Exponential (positive firing rates)
- Gamma (firing rate distributions)
- Beta (bounded activity)
Key result: FEP-PC correspondence maintained up to second cumulant of posterior.
Biologically Plausible Properties
1. Nonlinearity
- Nonlinear transfer functions emerge naturally
- Matches cortical neuron response curves
- No artificial linearization required
2. Heterogeneity
- Different neuron types → different EFD members
- Specialized populations (excitatory/inhibitory)
- Population-specific priors
3. Positive Firing Rates
- EFD naturally constrains to positive domain
- No negative firing rate artifacts
- Matches physiological observations
Local Plasticity Rules
Critical contribution: Model trained via biologically plausible local rules:
- Prediction errors computed locally
- Synaptic updates depend on local signals
- No global error propagation (vs. backprop)
- Compatible with cortical microcircuits
Plasticity Mechanism
Δw_ij = η * (prediction_error_i * activity_j)
where:
prediction_error_i = mismatch between predicted and observed
activity_j = presynaptic firing rate
- Updates are local (no global optimizer)
Variational Free-Energy Minimization
Under exponential-family assumption:
F = E_q[log q(z) - log p(z, x)]
≈ KL[q(z) || p(z|x)] - log p(x)
Minimized via:
- Variational posterior
q(z) from EFD family
- Prior
p(z) also from EFD
- Matching moments up to second order
Neural Implementation
Cortical Microcircuit Model
Layers:
- L4: Sensory input → prediction error computation
- L2/3: Prediction generation via EFD
- L5/6: Feedback to lower areas
Dynamics:
- Input arrives at L4
- L2/3 generates prediction (EDF parameters)
- L4 computes error (observed - predicted)
- Local plasticity updates predictions
- L5/6 sends feedback
Heterogeneous Populations
- Excitatory: Exponential/Gamma distribution (positive firing)
- Inhibitory: Beta distribution (bounded suppression)
- Binary: Bernoulli (decision neurons)
Advantages over Gaussian PC
| Property | Gaussian PC | Exponential Family PC |
|---|
| Firing rates | Can be negative | Always positive |
| Nonlinearity | Linear transfer | Natural nonlinearity |
| Heterogeneity | Homogeneous | Population-specific |
| Plasticity | Global gradient | Local biologically plausible |
| Explanatory power | Limited | Rich biological properties |
Applications
- Sensory cortex: V1/V2 perceptual inference
- Motor cortex: Action prediction
- Hippocampus: Memory prediction
- Spiking networks: Neural coding models
Methodological Checklist
- [ ] Select appropriate EFD member for neuron type
- [ ] Set prior parameters (natural parameters)
- [ ] Initialize variational posterior
- [ ] Compute prediction errors locally
- [ ] Apply local plasticity rules
- [ ] Validate positive firing rates
- [ ] Compare with Gaussian baseline
Implementation Outline
import numpy as np
class ExponentialFamilyPC:
def __init__(self, distribution_type='poisson'):
self.dist_type = distribution_type
self.prior_params = np.array([0.5, 1.0])
def compute_prediction_error(self, observed, predicted):
error = observed - predicted
return error
def local_plasticity(self, error, presynaptic_activity, eta=0.01):
delta_w = eta * error * presynaptic_activity
return delta_w
def enforce_positive_firing(self, firing_rate):
return np.maximum(firing_rate, 0)
Research Questions
- How do different EFD members affect prediction accuracy?
- What's the optimal prior for each neuron type?
- How does heterogeneity improve inference?
- Can local rules achieve convergence?
- What's the relationship to STDP?
Key References
- Free-Energy Principle: Friston (2010)
- Predictive Coding: Rao & Ballard (1999)
- Exponential Family: McCullagh & Nelder (1989)
- Local Learning: Doya (2000)
- Biological Plausibility: Lillicrap et al. (2020)
Activation Triggers
Use this skill when:
- Implementing biologically plausible predictive coding
- Modeling positive firing rate constraints
- Building heterogeneous neural populations
- Studying local plasticity rules
- Comparing Gaussian vs. exponential family inference
Citation
@article{kataoka2026extended,
title={Extended predictive coding framework as variational free-energy minimisation under exponential-family assumption},
author={Kataoka, Asaki and Doya, Kenji},
journal={arXiv preprint arXiv:2605.30882},
year={2026}
}
Theoretical Bridge
This work bridges computational theory (FEP) and biological implementation:
- Free-energy principle → variational inference
- Exponential family → biological neuron properties
- Local plasticity → cortical microcircuits
Insight: Theoretical frameworks gain explanatory power when accounting for biological constraints.