Issues training a clusterless classifier

[chris-angeloni]

I am using Neuropixels 2.0 recordings in hippocampus to decode position using your really great decoding tool. I was able to get this to work very well with single probe recordings. However, I have recently started trying to decode using data from recordings from two probes simultaneously and I'm having issues with the clusterless decoder not converging.

I have narrowed the issue down to the multiunit_likelihood step, in recordings with successful decoding, there is a variable called "summed_ground_process_intensity" that has some real values (plotting it looks like this:
)

However, if the decoder fails to converge, this variable is all nan.

During fitting, I printed the sum of the marginal density that is used to calculate the ground_process_intensity and found that in recordings that failed, the sum was nan for several of the multi-unit groups which had very low firing rates.

At this point I am not sure how to proceed. I think it is possible that there are several mua groups ('pseudo tetrodes' in my case) that just have very few spikes and are throwing off the likelihood estimation. Does this intuition make sense? Do you have any recommendations?

[chris-angeloni]

I figured out the issue. The "tetrodes" that were contributing nan to the likelihood estimate had only a single spike, removing them from the mua data passed to the classifier fixed the problem.

It looks like fit_multiunit_likelihood_gpu tries to avoid this by only including tetrodes with > 0 spikes (line 301 in multiunit_likelihood_gpu.py), but maybe this should be changed to > 1 spikes to avoid this failure mode (or maybe a more general argument to exclude mua groups below a certain spike count).