| name | predictive-coding-equilibrium-propagation-imagenet |
| description | Training Predictive Coding Networks on ImageNet using Equilibrium Propagation. Biologically plausible training framework for PCNs achieving near-backpropagation performance. Use when: (1) Training predictive coding networks at scale, (2) Implementing biologically plausible learning rules, (3) Scaling equilibrium propagation beyond small tasks, (4) Energy-based model training without backpropagation. Keywords: predictive coding network, PCN, equilibrium propagation, EP, energy-based model, ImageNet, biological learning, centered EP, equilibration scheme. |
Predictive Coding Network Training with Equilibrium Propagation at ImageNet Scale
Overview
Predictive Coding Networks (PCNs) are energy-based models from computational neuroscience traditionally trained with specialized algorithms. Equilibrium Propagation (EP) is a physics-based training framework. This skill documents the first successful combination of both at ImageNet scale, achieving 13.23% top-5 error rate (close to backpropagation's 12.2%).
Core Innovation
Centered Equilibrium Propagation for PCNs: Novel training method combining centered EP variant with specialized equilibration scheme for predictive coding networks.
Key Results:
- 10-layer convolutional PCN (VGG10) trained on full ImageNet
- 13.23% top-5 test error rate vs backpropagation baseline 12.2%
- First demonstration of both PCNs and EP at ImageNet scale
Methodology
Predictive Coding Network Architecture
PCNs are hierarchical energy-based models where:
- Each layer predicts the activity of the next layer
- Errors propagate backward through predictions
- Network minimizes a global energy function
VGG10 PCN Structure:
Input → Conv layers (10 total) → Output
↑ Predictions ↓ Errors
Centered Equilibrium Propagation
Standard EP has bias issues. Centered EP removes systematic bias:
Training phases:
- Free phase: Network relaxes to equilibrium given input
- Clamped phase: Output nudged toward target, network re-equilibrates
- Gradient computation: Difference between states
Centering modification:
- Subtract systematic bias from gradient estimate
- Use running averages to center the updates
- Enables unbiased gradient estimation
Novel Equilibration Scheme
PCNs require different equilibration dynamics than Hopfield networks:
Iterative relaxation:
for iteration in equilibration_steps:
for layer in reversed(layers):
update_predictions()
compute_errors()
minimize_local_energy()
Key parameters:
- Equilibration steps: ~100-500 iterations
- Step size: Adaptive based on energy gradient
- Convergence criterion: Energy change threshold
Training Procedure
Step 1: Initialize network with input image
x_input = image
Initialize all layers to predicted values
Step 2: Free phase equilibration
Relax network to equilibrium without output constraint
Monitor energy convergence
Step 3: Weakly clamped phase
nudge_output(epsilon) # epsilon ~ 0.01-0.1
Re-equilibrate network
Step 4: Gradient computation (centered)
gradient = (clamped_state - free_state - centering_term) / epsilon
centering_term = running_average_state_difference
Step 5: Weight update
weights += learning_rate * gradient
Update running averages for centering
Implementation Details
Energy Function
Total energy for PCN:
E(x) = Σ_l ||e_l||² + Σ_l ||r_l - f(r_{l+1})||²
Where:
e_l: Error at layer l
r_l: Representation at layer l
f: Prediction function (typically linear + activation)
Computational Considerations
Memory efficiency:
- Store free and clamped states separately
- Use gradient checkpointing for deep networks
- Batch processing reduces memory overhead
Speed optimization:
- Parallelize layer updates where possible
- Use momentum in equilibration dynamics
- Early stopping when energy converges
Hyperparameters
| Parameter | Typical Value | Notes |
|---|
| Equilibration steps | 100-500 | More for harder tasks |
| Nudge amplitude (ε) | 0.01-0.1 | Larger for faster convergence |
| Learning rate | 0.001-0.01 | Similar to backprop |
| Centering decay | 0.99 | Running average smoothing |
| Batch size | 32-128 | Standard ImageNet settings |
Comparison to Backpropagation
Advantages:
- Biologically plausible (no weight transport)
- Local learning rules possible
- Natural uncertainty quantification
Disadvantages:
- Slower training (requires equilibration)
- More memory (store two states)
- Slightly lower final accuracy
Performance gap: ~1% absolute (13.23% vs 12.2%)
Extensions
Three-Factor Learning Rules
EP naturally implements three-factor Hebbian learning:
Δw_ij = η * (x_i^clamped - x_i^free) * (x_j^clamped - x_j^free) * neuromodulator
Where neuromodulator = output error signal.
Hardware Implementation
Neuromorphic systems:
- Inference: Direct equilibrium dynamics
- Training: On-chip equilibration possible
- Energy: Potentially lower than backprop
FPGA acceleration:
- Parallel equilibration iterations
- Custom energy minimization circuits
Pitfalls
- Insufficient equilibration: Network not reaching true equilibrium leads to biased gradients
- Wrong nudge amplitude: Too large distorts dynamics, too small requires more iterations
- Centering not applied: Systematic bias accumulates without centering term
- Memory overflow: Storing two full states doubles memory requirements
Verification
Test equilibration convergence:
def check_equilibrium(network, input_data, tolerance=1e-5):
energy_prev = network.compute_energy()
for _ in range(10):
network.relax_one_step()
energy_curr = network.compute_energy()
if abs(energy_curr - energy_prev) < tolerance:
return True
energy_prev = energy_curr
return False
Activation
Trigger keywords: predictive coding, equilibrium propagation, EP training, energy-based learning, biological backprop, centered EP, PCN training, ImageNet PCN
References
See references/architecture_details.md for VGG10 PCN specification.
See references/mathematical_derivation.md for centered EP theory.
Source
arXiv:2606.03584 - "Training a Predictive Coding Network on ImageNet using Equilibrium Propagation" (June 2026)
Authors: Tugdual Kerjan, Rasmus Høier, Benjamin Scellier