Spiking Network: float vs 8-bit fixed-point

50 leaky integrate-and-fire neurons with random connections, driven by random input spikes. The same network runs twice on identical input: once with ordinary decimal numbers, once with small integers like a hardware chip would use. Change the number of bits and watch how the behaviour changes.

Float (decimal numbers)

Fixed-point (integers)

■ excitatory neuron ■ inhibitory neuron Each dot is a spike. Bars at the bottom: spikes per step.

Input spike rate Chance per step that a neuron gets an outside spike.

Excitatory weight How strongly an excitatory spike pushes its targets.

Inhibitory weight How strongly an inhibitory spike pulls its targets down.

Leak shift K Each step the neuron loses 1/2^K of its charge, so beta = 1 - 1/2^K. Both networks use this beta.

Weight bits Fewer bits means coarser weights in the fixed-point network.

Membrane bits Range of the stored membrane potential.

Random seed Same seed, same network and same input every time.

What to try

With defaults, both networks fire at similar rates, but the exact spike times drift apart after a few steps.
Lower the weight bits to 3 or 4. Weights get rounded to coarse values, and the fixed-point network can behave very differently.
Set membrane bits to 8 and raise the inhibitory weight. The potential hits its limits and the saturation counter climbs.
Rates are a better comparison than exact spike times, because small rounding differences grow in a recurrent network.

This is a teaching model with no learning in it. It is not an accurate copy of any specific chip.