| name | equilibrium-propagation-lif-snn |
| description | Equilibrium Propagation (EP) with Predictive Learning in Leaky Integrate-and-Fire Spiking Neural Networks. Biologically plausible alternative to backpropagation for training SNNs using energy-based two-phase learning. Use when training SNNs without backpropagation through time, implementing biologically realistic learning rules, or applying equilibrium-based optimization to spiking neuron networks. |
| activation_keywords | ["equilibrium propagation SNN","EP leaky integrate-and-fire","predictive learning SNN","biologically plausible SNN training","energy-based spiking learning","EP without BPTT","equilibrium propagation LIF","spiking neural network backprop-free"] |
| tags | ["spiking-neural-network","equilibrium-propagation","biologically-plausible-learning","leaky-integrate-and-fire","predictive-learning","energy-based-models","backprop-free-training"] |
Equilibrium Propagation in LIF Spiking Neural Networks
Description
Equilibrium Propagation (EP) is a biologically plausible alternative to backpropagation that trains neural networks by comparing free equilibrium states with nudged (target-perturbed) equilibrium states. When extended to Leaky Integrate-and-Fire (LIF) spiking neural networks, EP provides a backpropagation-free training method that respects biological constraints while achieving competitive performance.
Based on crossref:2026.05.19.726261 "Equilibrium Propagation with Predictive Learning in Leaky Integrate-and-Fire Spiking Neural Networks."
Core Concepts
Equilibrium Propagation Principle
EP computes gradients by comparing two equilibrium states:
- Free phase: Network relaxes to equilibrium with input only
- Nudged phase: Network relaxes with input + target perturbation
The gradient estimate:
∂E/∂θ ≈ (F_nudged - F_free) / β
Where β is the nudging strength and F is the equilibrium energy.
EP for LIF Neurons
For LIF spiking neurons, EP is adapted by:
- Spike-based equilibrium: Equilibrium defined as steady-state firing rates rather than voltage convergence
- Predictive learning: Network learns to predict future spikes based on current state
- Local plasticity rules: Weight updates computed from pre/post spike correlations in each phase
- Membrane potential dynamics: LIF differential equations govern neuron dynamics during both phases
Two-Phase Learning for SNNs
Free Phase (duration T_free):
1. Present input x as spike trains
2. Let network relax to steady-state firing rates
3. Record firing rates r_free and spike timings
Nudged Phase (duration T_nudged):
1. Present input x + target perturbation β·(y_target - y_output)
2. Let network relax to perturbed steady-state
3. Record firing rates r_nudged and spike timings
Weight Update:
Δw_ij ∝ (r_nudged_i - r_free_i) · r_pre_j
Mathematical Framework
LIF Neuron Dynamics
τ_m · dv/dt = -v + R·I_syn(t)
if v >= v_th: spike, v ← v_reset
EP Gradient Estimation
For weight w_ij connecting neuron j to neuron i:
∂E/∂w_ij ≈ (1/β) · [⟨r_i·r_j⟩_nudged - ⟨r_i·r_j⟩_free]
Where ⟨·⟩ denotes time-averaged firing rate correlation.
Predictive Learning Extension
The predictive learning variant adds:
- Temporal prediction error: Compare predicted next-state spikes with actual spikes
- Predictive nudging: Perturbation applied to prediction error rather than output error
- This enables learning temporal dependencies without BPTT
Usage Patterns
Pattern 1: Classification with EP-SNN
Train a spiking neural network for classification:
- Encode input as Poisson spike trains
- Build feedforward or recurrent LIF network
- Run free phase: present input, record steady-state rates
- Run nudged phase: add target-dependent perturbation, record rates
- Update weights using EP gradient estimate
- Repeat for multiple epochs
Pattern 2: Temporal Sequence Learning
Train on temporal sequences using predictive EP:
- Encode sequence as time-varying spike trains
- Use recurrent LIF connections for memory
- Free phase: let network predict next timestep
- Nudged phase: correct prediction errors
- Weight updates capture temporal dependencies
Pattern 3: Neuromorphic Hardware Deployment
Deploy EP-trained SNN on neuromorphic hardware:
- Train using EP simulation on conventional hardware
- Convert trained weights to neuromorphic format
- Deploy on Loihi/SpiNNaker with LIF neuron models
- No backpropagation needed during inference or online fine-tuning
Key Parameters
| Parameter | Description | Typical Value |
|---|
| β (nudging strength) | Perturbation magnitude | 0.01-0.1 |
| T_free | Free phase duration | 50-200 timesteps |
| T_nudged | Nudged phase duration | 50-200 timesteps |
| τ_membrane | LIF membrane time constant | 10-30ms |
| v_threshold | Spike threshold | 1.0 (normalized) |
| v_reset | Reset potential after spike | 0.0 |
Advantages Over BPTT
- Biological plausibility: No error backpropagation through time
- Local learning rules: Each synapse only needs local pre/post activity
- Energy efficiency: Can be implemented with event-driven neuromorphic hardware
- Temporal learning: Predictive EP handles sequences without storing full trajectories
Error Handling
Network Doesn't Converge to Equilibrium
- Increase free/nudged phase duration
- Reduce learning rate
- Check LIF parameters (τ_m, threshold) are stable
EP Gradient Too Noisy
- Average over multiple trials per phase
- Increase β slightly (but not too large to break linear approximation)
- Use exponential moving average of gradients
Poor Classification Performance
- Ensure input encoding preserves relevant features
- Try deeper network architecture
- Compare with surrogate gradient baseline to isolate EP-specific issues
Related Skills
- energy-based-neurocomputation: General energy-based learning frameworks
- snn-learning-survey: Comprehensive SNN learning rules survey
- decolle-snn-learning: Deep continuous local learning for SNNs
- selective-alignment-kd-snn: Knowledge distillation for SNNs
- spiking-free-energy-control: Spiking control via free energy principle