| name | predictive-coding-light |
| description | Predictive Coding Light+ (PCL+) methodology for spiking neural network sequence prediction. A spiking neural network architecture for unsupervised sequence processing that learns recurrent excitatory connections with delays to enable short-term retention of information. Combines spike timing-dependent plasticity (STDP) with synaptic delays to learn predictive representations. Successfully reproduces classic visual cortex sequence learning findings and learns to fill in missing inputs in gesture recognition tasks. Activation: Predictive Coding Light+, PCL+, STDP sequence learning, spiking neural network prediction, synaptic delay learning, unsupervised sequence processing, visual cortex sequence learning, short-term memory SNN, predictive coding spiking, spike timing-dependent plasticity sequence, gesture recognition SNN. |
Predictive Coding Light+ Methodology
Overview
Predictive Coding Light+ (PCL+) (arXiv:2605.12732v1) addresses a fundamental question: how can biological or artificial spiking neural networks learn to maintain past sensory information to predict the future? It proposes a novel SNN architecture combining spike timing-dependent plasticity (STDP) with synaptic delays to learn recurrent excitatory connections for short-term information retention.
Core Problem
Successfully predicting future sensory input requires maintaining a memory of the recent past. In spiking neural networks, it remains unclear how to learn this short-term retention without complex recurrent architectures or external memory modules.
Key Contributions
1. PCL+ Architecture
PCL+ learns recurrent excitatory connections with delays to maintain a record of the recent past:
- STDP-driven learning: Synaptic weights adapt based on spike timing correlations
- Synaptic delays: Different connection delays create a distributed temporal buffer
- Recurrent excitatory connections: Enable short-term retention without external memory
- Unsupervised sequence processing: No labeled data required
2. Sequence Learning in Visual Cortex
PCL+ reproduces classic findings on sequence learning in visual cortex, demonstrating:
- Temporal prediction of visual sequences
- Learning of stimulus order statistics
- Emergent selectivity for temporal patterns
3. Missing Input "Fill-in"
PCL+ learns to reconstruct missing input in a challenging gesture recognition task:
- Given partial temporal sequences, predict missing components
- Demonstrates robust temporal pattern completion
- Shows how short-term retention supports inference
Architecture Details
Core Components
Input Layer → [Recurrent Layer with Delays] → Prediction Output
↑ STDP learning rule
← Synaptic delay distribution
STDP with Delays
The key insight is that synaptic delays create a temporal buffer:
- Short delays → recent input retention
- Long delays → extended history access
- STDP learns which delays are useful for prediction
- Emergent delay distribution matches task temporal structure
Learning Rule
def stdp_update(pre_spike_time, post_spike_time, delay, weight):
"""STDP rule incorporating synaptic delay"""
effective_time = post_spike_time - (pre_spike_time + delay)
if effective_time > 0:
delta_w = A_plus * exp(-effective_time / tau_plus)
else:
delta_w = -A_minus * exp(effective_time / tau_minus)
return weight + delta_w
Application to Skill Development
When to Apply PCL+
- Building energy-efficient sequence prediction models
- Implementing biological short-term memory in SNNs
- Unsupervised learning of temporal patterns
- Gesture recognition with temporal context
- Visual sequence prediction tasks
- Any task requiring "filling in" missing temporal data
Comparison with Alternative Approaches
| Approach | Memory Mechanism | Supervision | Biological Plausibility |
|---|
| PCL+ | Synaptic delays + STDP | Unsupervised | High |
| RNN/LSTM | Hidden state | Supervised | Low |
| Transformer | Self-attention | Supervised | Low |
| Liquid State | Reservoir dynamics | Supervised (readout) | Medium |
Implementation Considerations
- Delay Distribution: Use a range of delays (e.g., 1-20ms) to capture different timescales
- STDP Parameters: Tune A+, A-, τ+, τ- for the specific temporal structure
- Network Size: Larger networks capture more complex temporal dependencies
- Input Encoding: Use latency coding or rate coding depending on the task
Research Implications
- Biological plausibility: Synaptic delays are a known biological mechanism for temporal processing
- Energy efficiency: SNNs with STDP are significantly more energy-efficient than traditional RNNs
- Unsupervised learning: No labeled data required — learns from raw temporal structure
- Short-term memory: Emerges naturally from recurrent connections with delays, no external memory needed
Related Skills
stdp-synaptic-delay-learning - STDP learning rule extension
spiking-neural-network-analysis - SNN paper analysis methodology
spiking-computational-neuroscience-survey - SNN applications survey
spikingjelly-framework - SNN deep learning framework
References
- Predictive Coding Light+: arXiv:2605.12732
- Classic visual cortex sequence learning studies
- STDP learning rule literature