Biases/Leak are not encouraged#

We suggest to stay away from the use of biases, especially on layers with a large number of neurons. They introduce a lot of overhead on the chip as it is not a sparse operation and therefore increase the amount of computation and consequently power.

However, if you still like to try the leak/bias, please refer to “how to leak neuron”

Event driven vs time-step based evaluation#

These are not the same thing, even though ALL of our pytorch model simulations are time-step based! So you might run into cases where your hardware produces a different result to that from your simulation. In the “Troubleshooting and Tips” Section you can find some advice on how to reduce such discrepancies.

Spike generation function#

Sinabs supports multiple spike generation functions such as SingleSpike, MultiSpike, etc. We have observed empirically, that the best match between Sinabs simulation and on-chip inference is achieved with the MultiSpike method for common network architectures. Therefore, we suggest to use this method if the end goal is to run the trained model on one of our chips. The intuition is that the hardware is event driven. Therefore, with sufficient density of input spikes, a neuron can emit multiple spikes per simulation time step.