Are spikeTimes in ms? + general comments

[alejoe91]

Hi @noamroth @nsteinme

I'm trying out the sliding RP computation on a KS2.5 output (NP2, 405 units, duration 732s) output and I have a few questions:

• when I run the slidingRP function and provide times as ms (as the docstring suggests), I get 402 NaNs out of 405 (for minContWith90Confidence). I think this is not expected and times should in fact be in s instead, since the phylib.correlograms function expects them in seconds (see here)

• Should the rpBinSize be equivalent to the params['binSizeCorr']? In that case, the correct param should be propagated here (using params['binSizeCorr'])

• When using spike times in seconds, I still get 236 NaNs. How can I interpret these? If I understand correctly, this means that there are no pairs of contamination levels/refractory periods with > 90% confidence. HOw would you deal with these units? And do you have an idea of why that would be the case?

Thansk in advance!
Alessio

[alejoe91]

Another point, as I'm porting the metric to SpikeInterface I noticed that the way the firingRate is computed might not be very robust.

In the same above-mentioned dataset, I have a unit with 1605 spikes (and a firing rate of ~2.92 Hz). The slidingRP function estimates the firing rate of this unit (with the correlograms approach) to be 81.2!

[nsteinme]

I'm going to let Noam reply about NaNs and units, but regarding firing rate computation, this is intentional - I think if you plot a raster or histogram of that unit's spiking over the duration of the session, you will find that it is silent for much of the session and fires very rapidly (around 80 sp/s) for only a short part of the session. Try it and let me know? If so, it is correct to consider that the firing rate for purposes of estimating contamination is 80 rather than 3 sp/s. We have done simulations with neurons that are only present for part of the recording to confirm that this definition of firing rate gives the correct answer.

[alejoe91]

Thanks for the explanation Nick!

Indeed:

Makes sense to me. Then I'll also use a similar approach :)

[noamroth]

Hi @alejoe91,

1. That's right, times should be in s. This was wrong in the slidingRP_all docstring (sorry about that - fixed now!).

2. Yes, it was hardcoded on that line but should be params['binSizeCorr'], I've fixed that now as well, thanks for noticing!

3. That's right, if a neuron never goes above 90% in the entire matrix we are testing, it will return a NaN, because minContWith90Confidence finds the lowest contamination at 90% confidence. We do note the possibility of NaNs in the docstring ("note: minContWith90Confidence, timeOfLowestCont will return np.nan for neurons with too few spikes -- these neurons have "empty" confidence and should be rejected."). But this should only affect neurons with very few spikes, or possibly very high contamination at all time points, I have never seen this affect such a large fraction of a dataset. Maybe you can send me the dataset in question and I can have a look? Thanks for your comments!!