Inferring networks in the presence of stimulation / intervention

[Adam W]

Hello,

I'm attempting to use IDTxl to infer connectivity in small networks of (simulated, leaky-integrate-and-fire) neurons in the presence of closed-loop and open-loop stimulation.

Are there any examples of applying IDTxl to systems with external stimulation in the literature or documented examples?

Naively, I expect one way to study this problem is to include the stimulus timeseries as a measurement, then at the analysis stage specify to only look for relationships FROM the stimulus and not TO the stimulus (i.e. it's a valid source but not a valid target).

Is this a reasonable approach? Is there any precedence for this? Does this violate any key assumptions of multivariate transfer entropy based approaches? Are there additional settings I should pay extra attention to for this sort of problem.

Thanks, and I'm looking forward to interacting more with the community and this toolbox!
Adam W.

[Joseph Lizier]

Hi Adam,

I can't think of examples where people have done this off-hand, but it seems fine enough to do. I think you've got the right approach to only look for links from the stimulus but not to it, that seems perfectly reasonable, and indeed so reasonable that I wouldn't worry about having any precedent for it to cite. Looks fine w.r.t. assumptions and I don't see any additional settings you would need.

One side comment is that if you were to leave the stimulus in as a target, then as you imply you may indeed identify links into it even though they are not causal. As always with these statistical/predictive methods, we don't expect a complete overlap with causality, so that would be not unexpected, and indeed there are many circumstances one can imagine where such links are meaningful from a predictive perspective. E.g. if the stimulus were regular but with some stochasticity, then in finite sized data sets it can be easy enough for one of the neurons to be providing additional information to decode the stimulus pattern over and above (or really, in the context of) what we can already see from the past of the stimulus. (and this would be exacerbated if the embedding dimension used for the stimulus is too low).

Best of luck!

--joe