Small Spiking Systems

For fun and learning.

↓ Slide 2 / 6

Simulations

Professional:

Multi-agent systems.

Hobby:

Emergent swarming of boids.
Boltzmann Machines.
An RSNN that trains in the browser.
RSNN with SNNTorch.
SNN with Numpy only.
Pathfinding with the Free Energy Principle.
Ising Model Visualisations.
Classifying spectrum from quantum dots.

↓ Slide 3 / 6

PING networks with a task.

✅ Small PING XOR network, not self-organising
✅ Larger self-organising PING network.
❌ Larger self-organising PING network with task.

↓ Slide 4 / 6

Tiny PING and XOR Model

Inputs A and B: Receive encoded external currents.
PING_I: Firing at ~40 Hz via E-I loop.
XOR_VETO: Silences XOR_OUT during 1,1 input via strong inhibition.
XOR_OUT: Spikes only under 1,0 or 0,1 conditions.

Hand tuned, not self-organising.

Raster Plot of Neuron Spikes

Input_A

Input_B

Ping ISI Histogram

Ping Frequency

38.79 Hz

Ping ISI

25.78 ms

ISI CV

0.06

↓ Slide 5 / 6

Larger PING-only network

80 / 20 E / I ratio.
LIF neuron threshold homeostasis.
Conductance-based synapses.
No plasticity.

15.1Hz average firing rate of E neurons.
86.3Hz average firing rate of I neurons.
27.0Hz oscillation frequency of E neurons.

Next steps: integrate a task:

XOR or another gate.
Coincidence detection.
Gate and reset.

Ping Spike Raster Plot

Spectrogram of E Neurons

↓ Slide 6 / 6

Notes on Workflow with AI

Trade of coding undergoing a revolution.
Have an idea, chat with the AI to make a coding plan.
Modularise the plan.
Closely guide the AI to write the modules.
Add multiple quantitative metrics.
Put it all together.
Iterate.
Save data and make plots.

RISK of shallow understanding.

Terminal output of the simulation

Slide 1 / 6