Sorting in the presence of tons of multi-unit activity

[JoseGuzman]

I have very small neurons (similar to #123 i.e., spikes rates as low as < 0.1 Hz) which appear every 10 to 90 seconds together with bursts of high-frequency multi-unit activity. These last are probably unsortable, as originally suggested by @marius10p in #123 (comment). However, due to these bursts of activity, or maybe because most of the time the tissue is silent, KiloSort2 fails to detect any isolated units from spikes (60 to 100 uV) with the standard configuration (see image below, green is the multi-unit background and pink putative units).

Changing ops.minfe_goodchannels = 0; and ops.minFR = 1/300; allowed me to isolate the large units if they're in a separate channel. However, is the spikes are small amplitude (approx. 30 uV) or two- or more spikes are seen on the same channel, isolation is not possible (contamination is close to 70%, see figure). I think KiloSort2 is including some of these multi-unit spikes.

Despite the discussion, we had in the neuropixels 2009 (see also #59). I tried to change ops.spkTh = -7; or even to -8, but things did not change much. I also changed ops.Th = [8 4] following the suggestion in (#59), but the amount of detected spikes from the multi-unit is very large.

Is there any other alternative to discard all these noise spikes? I didn't play with ops.ThPre because I didn't find any documentation.

I would appreciate some comments here.

[JoseGuzman]

I realized half of the binary is corrupted. I will make a reference #123 to be consulted there.

In my case ops.criterionNoiseChannels = 0.2 alleviate the contaminations in the refractory period.