framework handles neuron threshold of <=0 differently than hardware

[BGull00]

The solution here may just be to restrict the threshold such that it cannot be <=0, but the behaviors still differ currently. The framework does not cause neurons to spike without input while the hardware does (e.g. neuron with threshold 0 spikes every timestep).

[keegandent]

Hmm, I get why this happens in the simulator. @jimplank and I were discussing earlier about its event-driven behavior. However, it seems that a neuron with threshold <=0 not spiking is the aberration.

Also, it prevents the creation of what I would call "generative spiking networks" which don't require constant external stimuli to produce fires. For example, say you wanted a network whose decoded output produced RF data that mimicked some drone. I get you could "kickstart" such a network with an external stimulus, but I don't see why it should have to be a strict requirement.

[keegandent]

Also to compare with the "XOR" network, you need to"spikes by default" behavior to create the "NOT gate" equivalent

[BGull00]

My opinion here is based on the idea that SNNs are event based. If a neuron does not receive an input event at a timestep, then it should not fire under the default behavior. At least that is what makes sense to me. @keegandent, I do see the usefulness for neurons that continuously fire without requiring any input to "kickstart" activity, so maybe that could be a processor parameter @jimplank?

[keegandent]

They are event based, but some neurons are "event generators" in this model and can still be disabled with inhibitory spikes (basically does it get enough negative charge in a timestep to push it below that threshold, otherwise I think it would reset back up to 0 every fire). The implementation of a "don't do anything if nobody provides input" is kind of hideous as you check every synapse port to find charge and don't do anything if you don't, oh but except leak which has to be done immediately... very gross. It may also affect inclusion of multiplicative leak when we try implementing multiplicative leak for a full Leaky-Integrate-and-Fire (LIF) model.

[keegandent]

@jimplank, I just had a discussion with somebody from a neuroscience background about this, and the simulation's behavior for neurons with negative threshold (despite that already being a shaky concept) is very clearly nonsensical. If you look at the table below, it's very clear that the software will decline to fire a neuron with negative threshold when there is net-zero charge activity, but will fire with an inhibitory net charge if it doesn't put that neuron below the threshold. This completely defies logic of how neurons work. I understand it may be a pain to simulate, but the way the simulator does it is frankly just incorrect.