Windows runtime error and Izhikevich neuron model

[ahmetozleraktas]

Hello, I hope you're all doing well. I am trying to solve this problem which I mentioned in the title. There is no ready-to-use model for Izhikevich neuron model. I saw the LIF but i couldn't manage to modify it for changing to Izhikevich model. Are there any documentation for it? Can you help me? Thanks for all answers! Have a nice day.

[weidel-p]

Hi,

you are right, there is no ready to use Izhikevich neuron model available in the process library yet. Have you seen the jupyter notebook tutorials? Especially the tutorial about creating custom ProcessModels from scratch should be helpful for you.

Please don't hesitate to ask if you need further assistance. If you succeed in implementing the Izhikevic neuron model, consider making it available to other users by creating a pull request.

[ahmetozleraktas]

I checked the MNIST classifier, but didn't notice that tutorial you mentioned. It seems a bit difficult to begin but maybe I can figure it out. You can be sure, I will add the model if I solve it. Thanks for your answer by the way. Best regards!

[felwang-sandia]

Did this ever get resolved/added?

By the way, a simple izhikevich model I've used previously (taken from [1]) would have the population-wide (constant) parameters: a, b, c, d, and the individual (mutable) states: v, u, I.
The forward equations would be (time-step split into two):
v = v + dt/2 * ((0.04*v + 5)*v + 140 - u + I
v = v + dt/2 * ((0.04*v + 5)*v + 140 - u + I
u = u + dt*a * (b*v - u)
On spike:
v = c
u = u + d

[1] E. Izhikevich, Which Model to Use for Cortical Spiking Neurons? IEEE Trans. on Neural Networks, v. 15, no. 5, 2004

[ahmetozleraktas]

Thanks. For completeness, could you also provide what you measure in terms of runtime on your end and what you would expect in terms of runtime? Your numpy code should not be the issue. It's likely something in the channel communication.

If you don't mind, please open an Issue for this. This way we can communicate on this particular speed problem and close it when the problem is resolved.

You are very right, at first I thought it was about me, so I thought I could handle it from here. If I can run @felwang-sandia 's code, I will create an issue with the outputs.

[awintel]

So @felwang-sandia can run it in jupyter (or .py?) but you, @ahmetozleraktas can't?
Maybe an operating system issue?
If you figure it out, please include that info in the issue.

[felwang-sandia]

Yes, I'm running the code on a jupyter notebook. My development environment is on an ubuntu 22.04 machine, I'm using anaconda for the python virtual environment with python 3.8, and I'm using lava 0.5 (installed with "conda install lava -c conda-forge").

[ahmetozleraktas]

@awintel I opened an issue: #401.

[awintel]

Thanks!

[awintel]

Ok. Thanks for the suggestions.
FYI: We have just released an E/I network tutorial, although not with Izhikevich neurons but normal rate and LIF neurons:
https://github.com/lava-nc/lava/blob/main/tutorials/end_to_end/tutorial02_excitatory_inhibitory_network.ipynb

[felwang-sandia]

Thanks, I've since taken a look at the E/I network tutorial. It's very nice, and I especially like the additional analysis function looking at criticality/correlation length. I know sparse connectivity and varied synaptic delays (tag_2, I believe?) are not quite implemented yet, but a network with those additional properties would be a great addition to the E/I network tutorial if/when they come about.

[awintel]

The compiler should support synaptic delays by now. But we do not have a tutorial for it yet.

[felwang-sandia]

That's great!

[ahmetozleraktas]

UPDATE: I tried the codes separately on Ubuntu and Windows. There weren't any of the problems I mentioned in Ubuntu, but Windows runtime and Jupyter Notebook problems continue. I opened an Issue about it (#401). I guess it's an operating system problem for now.