issue with whitening when electrode locations are duplicated

[magland]

I realize it should never happen that two electrodes have identical location. However, unless I am making an error in reading the data file, I have come across an example of this in an IBL dataset on the DANDI archive. I still need to double check and confirm that this is the case.

But in any event, here's what happens when you have a duplicated location - and I think this is something that should be fixed.

Here's the relevant code

Kilosort/preProcess/whiteningLocal.m

Lines 1 to 14 in d55179f

	function Wrot = whiteningLocal(CC, yc, xc, nRange)
	% function to perform local whitening of channels
	% CC is a matrix of Nchan by Nchan correlations
	% yc and xc are vector of Y and X positions of each channel
	% nRange is the number of nearest channels to consider
	Wrot = zeros(size(CC,1), size(CC,1));
	for j = 1:size(CC,1)
	ds = (xc - xc(j)).^2 + (yc - yc(j)).^2;
	[~, ilocal] = sort(ds, 'ascend');
	ilocal = ilocal(1:nRange); % take the closest channels to the primary channel. First channel in this list will always be the primary channel.
	wrot0 = whiteningFromCovariance(CC(ilocal, ilocal));
	Wrot(ilocal, j) = wrot0(:,1); % the first column of wrot0 is the whitening filter for the primary channel
	end

Note the comment "First channel in this list will always be the primary channel." This is not the case when you have duplicates. And so what ends up happening is that two rows (or maybe columns) of the whitening matrix end up being identical. This in turn causes two channels to be identical signal in the filtered data. And this causes a problem with detection where all spikes are considered non-isolated. Finally this in turn causes the crash of #566

[magland]

As a follow-up. It seems the IBL dataset is actually okay... you just need to read the rel_x, rel_y fields from the NWB file and not the x, y, z. Still investigating.