spiking-neural-net

A spiking neural network simulation library.

Usage

You will need a Rust development environment, which you can install by visiting https://rustup.rs/ and following the instructions.

Once you have Rust and Cargo installed, you can run a simulation with:

make

You should see output like this:

This will produce plot images called neuron-trace.png and spikes.png like the ones below.

Note that you can also build the debug version by omitting the --release flag, but it will run slowly. This is good if you want to use a debugger, but if you want to simulate any reasonably large networks in real-time you will need to use release mode.

Sample output

It should create output images like this:

Filename	Image
neuron-trace.png	This shows the activity of a single neuron. Each of the vertical lines is an action potential, shown as a white dot on the plots below. There's a noise floor representing random thalamic input.
spikes.png	You'll need to download this image and zoom in to see the activity.
spikes.png (cropped)	This represents about 1s of time across 1100 neurons.

The spikes.png image shows some fascinating results:

• In the cropped photo you can see vertical lines of dots. This shows that all the neurons in the network have synchronized and are pulsing together at about 12Hz. This is a similar rate to Alpha waves in mammalian brains.
• Also in the cropped photo can also see horizontal lines. This is the motor layer of each column; this layer currently is weakly connected and doesn't synchronize with the rest of the neurons.
• In the full spikes.png image there are three distinct phases:
  • 0s-4s: The network quickly settles into a 5Hz rhythm, similar to Theta waves. I believe this is when the columns are organizing themselves.
  • 4s-10s: The network becomes disorganized. I believe this is when the columns are learning to wire together and are trying to sort out their connections.
  • 10s+: The network becomes synchronized at a much faster 12Hz rhythm.

About the simulation

This library simulates networks of biologically-inspired neurons. Spiking neural models are implemented as ordinary differential equations integrated using Euler integration at 1 millisecond resolution.

The simuation is written in the Rust programming language and uses Specs, an Entity-Component-System framework with excellent parallelization and performance. This allows it to simulate simulate about 1800 neurons and 100k synapses, including an online learning algorithm, in real-time on a typical laptop.

Features

Neural models

• Izhikevich neurons
  • Model: https://www.izhikevich.org/publications/spikes.htm
  • Code: https://github.com/michaelmelanson/spiking-neural-net/blob/master/src/simulation/models/izhikevich.rs
• Hindmarsh-Rose neurons
  • Model: https://en.wikipedia.org/wiki/Hindmarsh%E2%80%93Rose_model
  • Code: https://github.com/michaelmelanson/spiking-neural-net/blob/master/src/simulation/models/hindmarsh_rose.rs

Learning models

• Spike-timing dependent plasticity
  • Code: https://github.com/michaelmelanson/spiking-neural-net/blob/master/src/simulation/learning/stdp.rs

Network organization

• Columnar organization
  • Networks are organized into 100-neuron columns composed of five layers (motor, sensory, afferent, efferent, internal)
  • Code: https://github.com/michaelmelanson/spiking-neural-net/blob/master/src/simulation/mod.rs#L70-L160

TODO

• Multiple morphologies. Currently all neurons are Izhikevich 'regular spiking' neurons.
• Configuration. Currently you need to change the source code to change the network design or neural model parameters.