SORN: Self-Organizing Recurrent Neural Network

A pure-Julia implementation of a biologically plausible spiking neural network combining five plasticity mechanisms with excitatory-inhibitory balance.

Unique Contribution

First pure-Julia SORN combining:

1. Pair STDP — Hebbian learning on E→E connections
2. Inhibitory STDP — Anti-Hebbian learning on I→E connections
3. Synaptic Scaling — Homeostatic weight normalization
4. Intrinsic Plasticity — Adaptive spike thresholds (LIF)
5. Structural Plasticity — Dynamic synapse formation/elimination

with biologically realistic 80:20 E/I balance.

Architecture

• Excitatory neurons: Leaky Integrate-and-Fire with adaptive threshold
• Inhibitory neurons: Izhikevich fast-spiking model
• Connectivity: Sparse random with 80:20 E/I ratio
• Weight matrices: E→E (plastic), I→E (plastic), E→I (fixed), Input→E (fixed)

Performance

• Sparse weight matrices (SparseMatrixCSC) — only stores 15% non-zero connections
• Struct-of-arrays layout for SIMD vectorization
• Zero-allocation simulation loop
• All inner loops @simd-vectorized (neuron updates, current computation, weight scaling, recording)
• Analytical LIF solution (exact ODE integration, not Euler approximation)
• Exponential decay for threshold adaptation and synaptic scaling
• Structural plasticity rebuilds from IJV triplets (avoids O(nnz) per-insertion cost)

Scaling (Celeron N4020)

Neurons	500ms sim	1000ms sim
100	~50ms	~100ms
200	~60ms	~130ms
400	~480ms	~960ms
800	~2s	~4s
1600	~10s	~20s

Quick Start

julia --project=. examples/demo.jl

Project Structure

SORN/
├── src/
│   ├── SNN.jl           # Main module
│   ├── utils.jl         # Poisson sampling, trace decay
│   ├── neurons.jl       # LIF + Izhikevich pools
│   ├── plasticity.jl    # 5 plasticity mechanisms
│   ├── network.jl       # SORN architecture
│   ├── simulation.jl    # Simulation loop
│   └── encoding.jl      # Input encoding
├── examples/
│   ├── demo.jl          # Full demonstration
│   └── benchmark.jl     # Performance benchmarks
├── Project.toml
└── README.md

Usage

using .SNN
using .SNN.Network: create_sorn
using .SNN.Simulation: simulate!
using .SNN.Encoding: poisson_encode

net = create_sorn(n_exc=80, n_inh=20, n_input=100, seed=42)
input = poisson_encode(100, 1000; rate=0.1)
result = simulate!(net, input; verbose=true)

Use Cases

• Computational neuroscience research (cortical microcircuit modeling)
• Neuromorphic computing prototypes
• Studying emergent memory/sequence learning in recurrent networks
• Educational tool for spiking neural network dynamics
• Baseline for comparing plasticity rule combinations

References

• Lazar et al. (2009) "SORN: a self-organizing recurrent neural network"
• Turrigiano & Nelson (2000) "Homeostatic plasticity in developing networks"