| name | local-synaptic-rules-sigreg-gradient |
| description | Local synaptic learning rules (STDP+ and homeostatic plasticity) can implement exact SIGReg-like self-supervised learning gradients without backpropagation, global error signals, or weight transport. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["neuroscience","spiking-neural-networks","synaptic-plasticity","gradient-learning","biologically-plausible"],"related_skills":["spiking-neural-network-analysis","stdp-bernoulli-message-passing","feedback-hebbian-continual-learning"]}} |
Local Synaptic Rules Implement SIGReg Gradient
This skill implements the methodology from arXiv:2607.21622 "Local Synaptic Rules Can Implement a SIGReg Gradient Without Backpropagation" by Martin Andrews.
Core Insight
Two canonical local synaptic learning rules together can implement the exact gradient of a SIGReg-like self-supervised learning objective:
- Potentiation arm of spike-timing-dependent plasticity (STDP⁺)
- Homeostatic plasticity (instantiated via flashlight granule-cell-like neurons)
This equivalence requires:
- No gradient calculations
- No global error signals
- No weight transport
- No label information
- Only inputs: pre- and post-synaptic firing rates, local firing statistics, and temporal contiguity of natural sensory streams
Implementation Steps
1. Network Architecture Setup
2. STDP⁺ Rule Implementation
def stdp_potentiation(pre_spike_time, post_spike_time, A_plus, tau_plus):
if post_spike_time > pre_spike_time:
delta_t = post_spike_time - pre_spike_time
return A_plus * exp(-delta_t / tau_plus)
return 0
3. Homeostatic Plasticity Implementation
def homeostatic_plasticity(firing_rate, target_rate, eta_homeo):
return eta_homeo * (target_rate - firing_rate)
4. Combined Weight Update Rule
def combined_update_rule(weight, pre_rate, post_rate, local_stats, temporal_context):
stdp_component = compute_stdp_potentiation(pre_rate, post_rate, temporal_context)
homeo_component = homeostatic_plasticity(post_rate, target_rate, eta_homeo)
delta_weight = stdp_component + homeo_component
return weight + learning_rate * delta_weight
5. Temporal Ordering Requirement
- Input presentation must preserve temporal contiguity of natural sensory streams
- Random ordering fails to recover class structure
- Ordered presentation enables cluster separation through temporal statistics alone
Validation Tasks
Synthetic Clustering Task
- Purpose: Probe whether class structure can be recovered from temporal ordering alone
- Metric: Cluster Separation Ratio (CSR)
- Expected Results:
- Ordered presentation: CSR ≈ 2.49 (≈3.5σ separation)
- Random ordering: CSR ≈ 0.83 (near baseline)
Temporally Ordered MNIST
- Architecture: Two-layer network trained entirely with local rules
- Evaluation: Linear-probe accuracy on learned representations
- Expected Result: ~87.3% accuracy
Biological Plausibility Advantages
- No weight transport problem: Updates use only locally available information
- No global error signals: Learning driven by local firing statistics and temporal contiguity
- Biologically realistic mechanisms: Uses established synaptic plasticity rules
- Self-supervised: No labels required, learns from natural temporal structure
Applications
- Biologically plausible deep learning: Bridge between neuroscience and AI
- Neuromorphic hardware: Energy-efficient learning without backpropagation
- Unsupervised representation learning: Extract structure from temporal data streams
- Continual learning: Natural integration with online learning scenarios
Activation Keywords
- local synaptic rules
- STDP gradient learning
- biologically plausible backpropagation
- SIGReg without backprop
- temporal ordering clustering
- homeostatic plasticity gradient
References
- Andrews, M. (2026). Local Synaptic Rules Can Implement a SIGReg Gradient Without Backpropagation. arXiv:2607.21622
- Original SIGReg framework: Self-supervised learning through temporal prediction
- Biological STDP: Caporale & Dan (2008), Markram et al. (1997)
- Homeostatic plasticity: Turrigiano (2012), Zenke et al. (2013)