There are a few neurons in the brain that have intrinsic bursting firing rates under certain conditions. For example, see this paper on “Thalamic model of awake alpha oscillations and implications for stimulus processing”. Have bursting effects of neurons been used effectively in Nengo yet?
P.S. This is my first attempt of using this forum instead of just sending emails to specific people. Let me know if I’m doing this right.
I remember Terry was doing stuff with Gruber spiny neurons a while back. I seem to remember them having bursting behaviour, but I could be wrong… Either way we don’t have a good understanding yet of how to exploit these cellular dynamics, other than taking an RC-like approach and explicitly simulating the network to solve for decoders. It would be nice to characterize classes of dynamics in the vector space that would be supported by these neurons. This is something I’ve wanted to look into more analytically, but the more the merrier.
We do have the Izhikevich neuron model, which exhibits bursting in some parameter regimes. But how you want to use these is up to you. In general, the NEF doesn’t have good support for more complex neuron dynamics like bursting, though Aaron’s been working on this.
Yup you’re doing it right :). Also, the purkinje cells in the cerebellum are highly active without input (as in, if you take them out of the brain, they fire a lot). Which isn’t exactly bursting, but it is a very different kind of neuron that what you usually find in cortex (pyramidal cells).
There were a few things that came up in trying to control DMP / iDMP networks that made me want something like bursting neurons. The basic idea was like this:
- I have some circuit that will cause some value to go above some threshold.
- When it goes above threshold, an ensemble becomes active.
- That ensemble inhibits certain things, including the circuit that causes the value to go above threshold.
The issue with that is that once the ensemble becomes active, it causes its own inactivation. It might not be active long enough for the ensemble to inhibit everything it needs to inhibit. I got around this by playing around with times and whatnot until it worked, but it seemed like just the situation where you would want bursting neurons.
It turns out that bursting Izhekivich neurons can be used to improved the computational power of Liquid State Machines (LSM). For those who don’t remember, LSM are the spiking version of Reservoir Computing, which are randomly recurrently connected neuron ensembles.