Event-Driven Simulation of Networks of Spiking Neurons

Lloyd WattsNeurIPS 1993

Introduction

Traditional artificial neural networks (ANNs) often simplify biological complexity by relying on "mean firing rates," representing a neuron's activity as a continuous numerical value. However, in biological systems, the precise timing of individual spikes is frequently critical. For example, barn owls rely on microsecond-level axonal delays to localize sound, and blowflies can make split-second flight decisions based on just a few spikes.

To study these intricate systems, researchers typically use "compartmental models" that solve complex differential equations. While highly accurate, these models are computationally expensive and scale poorly. This paper introduces Spike, an event-driven simulator designed to bypass the heavy computational cost of traditional models. By treating currents as piecewise-constant pulses, Spike predicts future firing events and simulates large-scale, biologically realistic networks in a fraction of the time required by standard methods.

Core Methodology: The Piecewise-Constant Assumption

The central innovation of the Spike simulator is a strategic simplification known as the piecewise-constant assumption:

  1. Boxcar Pulses: The simulator assumes all currents injected into a cell—whether synaptic or tonic—are "boxcar" pulses, meaning they remain constant over a specific duration.
  2. Piecewise Linear Voltage: Because the input current is constant, the integration of that current (the membrane voltage) becomes piecewise linear.
  3. Event-Driven Queue: Instead of calculating the state of every neuron at every millisecond (time-stepping), the simulator evaluates the current state and predicts when a neuron will hit its firing threshold. It then places that "spike event" into a priority queue.
  4. Dynamic Rescheduling: If a new input arrives before a predicted spike occurs, the simulator recalculates the trajectory and reschedules the event accordingly.

This simulation engine is paired with NeuraLOG, a schematic capture tool that allows researchers to visually "draw" neural circuits. Users can define parameters like refractory periods, axonal delays, and synaptic saturation directly on the diagram.

Schematic of a simple bursting circuit with inhibitory feedback

Validation and Biological Fidelity

Because this research focuses on simulation infrastructure rather than a specific machine learning task, it validates the simulator using known biological circuit architectures rather than standard datasets. Key validation models include:

  • Simple Bursting Circuit: A two-neuron model used to demonstrate adaptation and inhibitory feedback.
  • Locust Central Pattern Generator (CPG): A complex circuit developed to model rhythmic locomotor patterns.
  • Cortical Oscillation Models: Large-scale simulations involving hundreds of neurons and thousands of synapses.
Simulation results of the bursting circuit showing piecewise linear voltage integration and spike events

The tool successfully replicated the adapting bursting behavior of neurons. In these simulations, the firing rate slows down over time due to self-inhibitory feedback—a behavior typically associated with complex calcium-dependent potassium channels in biology, achieved here through simple network topology.

Key Insights and Results

  • The Power of Topology: The paper demonstrates that "primitive" building blocks (like boxcar pulses) can produce highly sophisticated biological behaviors, such as gradual adaptation and bursting, provided the network topology is correct (e.g., using a summating synapse in a negative feedback loop).
  • Hardware-Software Symmetry: A major insight is the direct link to hardware. The simulator’s primitives (synapse, tonic input, and neuron) correspond exactly to analog CMOS VLSI circuit primitives. This allows engineers to simulate and optimize complex neural chips before they are fabricated in silicon.
  • Efficiency via Prediction: By shifting from "calculating what is happening now" to "predicting what will happen next," the simulator removes the overhead of simulating "quiet" periods in a network. It focuses computational power exclusively on active events.
  • Performance and Scalability: The simulator achieved a significant speedup, running large networks in a fraction of the time required for full compartmental simulations. It successfully handled networks of hundreds of neurons and thousands of synapses, a substantial achievement for its time.

Limitations and Future Directions

While highly efficient, the piecewise-constant current assumption is a double-edged sword. It may not capture the subtle non-linearities of real biological membranes or complex dendritic processing that does not fit the "boxcar" profile.

Future directions suggested by the author include applying these tools to more complex "system-level" questions, such as spiking associative memories and chaotic spiking circuits. By doing so, tools like Spike continue to bridge the gap between abstract artificial intelligence and biological reality.

Summary updated

BibTeX
@inproceedings{watts1993event,
  author = {Lloyd Watts},
  title = {{Event-Driven Simulation of Networks of Spiking Neurons}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1993},
}
Figure