| name | hybrid-ann-snn-local-plasticity |
| description | Hybrid ANN-SNN pipeline with local plasticity: couples pretrained ANN encoders with spiking classifiers using biologically-inspired local learning rules, bypassing end-to-backprop. Use when building energy-efficient spiking neural networks from pretrained models, implementing local Hebbian/plasticity rules, or converting ANN features to spike trains. |
| metadata | {"arxiv_id":"2606.20151","published":"2026-06-18","authors":"","tags":["spiking-neural-network","local-plasticity","ann-snn-conversion","neuromorphic","rate-coding"]} |
Hybrid ANN-SNN Pipeline with Local Plasticity
Core Idea
Combine a pretrained ANN encoder (e.g., EfficientNet) with a spiking classifier trained using local, biologically-plausible learning rules. This bypasses end-to-end backpropagation while achieving performance comparable to deep networks (99.09% on 64-class ImageNet).
Architecture
Input → [Pretrained ANN Encoder] → Rate Coding → [Spiking Classifier] → Output
(activations → spike trains) (local learning rules)
Key Components
1. ANN Encoder Selection
- Use any pretrained feature extractor (EfficientNet, ResNet, etc.)
- Extract intermediate layer activations as feature embeddings
- No fine-tuning needed — encoder stays frozen
2. Rate Coding Conversion
- Convert continuous ANN activations to spike trains
- Higher activation → higher firing rate
- Temporal window determines spike count resolution
- Trade-off: longer windows = more accurate but higher latency
3. Spiking Classifier (CoLaNET)
- Spiking neural network with local learning rules
- Receives spike trains from rate-coded ANN features
- Trained independently — no gradient flow back to encoder
4. Local Learning Rules
- Hebbian plasticity: strengthen co-active connections
- Spike-timing dependent plasticity (STDP)
- Local error signals without global backpropagation
- Biologically plausible — no weight transport problem
Workflow
- Select pretrained encoder: Choose based on task domain and feature requirements
- Extract features: Run input through encoder, capture activations
- Configure rate coder: Set temporal window and firing rate mapping
- Build spiking classifier: Design SNN architecture for the task
- Train with local rules: Apply Hebbian/STDP learning to SNN weights
- Evaluate: Measure accuracy vs energy efficiency trade-offs
Advantages
- Biological plausibility: No end-to-end backpropagation
- Energy efficiency: SNN inference is sparse and event-driven
- Leverages pretrained models: Benefits from large-scale pretraining
- Modular: Encoder and classifier can be developed independently
- Scalable: Local rules parallelize naturally
Pitfalls
- Information loss in rate coding: Continuous values discretized to spikes
- Temporal latency: Rate coding requires time windows for accuracy
- Local rule limitations: May not capture complex feature interactions
- Hyperparameter sensitivity: Learning rates, thresholds, time constants need careful tuning
- Task dependency: Works best when pretrained features are already discriminative
Activation Keywords
- hybrid ann snn
- local plasticity
- spiking neural network conversion
- rate coding
- biologically plausible learning
- CoLaNET
- hebbian spiking
- snn classifier
- neuromorphic pipeline