Case Study: Cortical Column Simulation
Demonstrating the full power of Reaktor's stack—including distributed actors, C++ Highway SIMD, Reaktor Mesh, and the Blueprint editor—is the 100,000-neuron cortical column simulation.
Architecture & Components
| Component | Implementation |
|---|---|
| SpikeNeuronActor | Models the LIF/Izhikevich neuron. Uses C++ Highway SIMD for vectorized weight accumulation across synapses. Receives and fires spike messages via typed ports. |
| SynapseEdge | A weighted edge with STDP (Spike-Timing-Dependent Plasticity). Uses C++ SIMD for exponential decay calculation during Hebbian learning. |
| ColumnSupervisorActor | Supervises a single cortical layer. Detects pathological synchronization, modulates inhibition, and reports telemetry to the Blueprint editor. |
| SimulationCoordinator | The global clock for the simulation, broadcasting ticks every 0.1–1ms and managing checkpointing to ObjectStore. |
Distributed Simulation
- Horizontal Scaling: Each mobile phone in the mesh hosts 1,000–5,000 neurons, while server-side actors host the coordinator and heavily connected neurons.
- Intra-device Communication: Uses in-process Channels with sub-microsecond latency.
- Inter-device Communication: Uses the Reaktor Mesh with unreliable
DataChannels(1–10ms latency), matching biological noise tolerance. - Visualization: The Blueprint editor visualizes the simulation in real-time. Neurons are colored by their firing rate, edges by synaptic weight, and the system displays real-time spike rasters.
Why Reaktor for Neural Simulations?
This simulation proves a key architectural insight: neural network implementations do not require custom infrastructure. They are simply actors sending typed messages through the graph over the Mesh. The same runtime that powers a social chat application also powers a distributed neural simulation, with the Blueprint editor providing the same high-level visualization for both.