Kilosort

UPDATE 3/8/24: this version has the "spike holes" bug from issue 594 and is only left here as a reference. Please use the new release above "Kilosort 3.0 patch0" where the main bug has been fixed. We will continue to investigate if other bugs are left that produce this behavior.

This is the last version of Kilosort 3.0 and the last Kilosort version in Matlab. No further support or code changes for Kilosort 3 will be provided beyond this release.

This release fixes both bugs for the spike holes problem (issue https://github.com/MouseLand/Kilosort/issues/594). The default value of the NT parameter has changed to 65,472 in configFile384. You need to use these defaults to fix the spike holes problem (or the rules specified in configFile384 for NT and ntbuff). Please recompile the CUDA functions with mexGPUall.

The release will be updated if further modifications are made (patch2, 3 etc).

Note: the previous release 2.5.1 had incorrect spike times and only fixed one of the two bugs.

UPDATE 3/8/24: this version has the "spike holes" bug from issue https://github.com/MouseLand/Kilosort/issues/594 and is only left here as a reference. Please use the new release above "Kilosort 2.5 patch0" where the main bug has been fixed. We will continue to investigate if other bugs are left that produce this behavior.

This release fixes a bug for the spike holes problem (issue https://github.com/MouseLand/Kilosort/issues/594). The default value of the NT parameter has changed to 65,472 in configFile384. You need to use these defaults to fix the spike holes problem (or the rules specified in configFile384 for NT and ntbuff). Please recompile the CUDA functions with mexGPUall.

The release will be updated if further modifications are made (patch2, 3 etc).

Note: this aligns the changes in Kilosort 2.0.2 to those in 2.5.2 and 3.0.2.

This release is the final Kilosort 2.0 build, before moving on to v2.5. A new drift correction algorithm is introduced in v2.5, which makes obsolete the drift tracking approach from v2.0. In addition, the two algorithms are incompatible and we are dropping support for drift tracking from v2.5 onwards. Users of Neuropixels 1.0 and 2.0 should transition directly to Kilosort 2.5.

Nonetheless, we expect some users will get better performance from v2.0, at least initially. Recording configurations that may benefit more from v2.0 than v2.5 are: tetrodes, single-channel recordings such as from a Utah array, probes with >30um vertical spacing.

To learn more about Kilosort 2.0, read the supplementary material and figures in [Stringer, Pachitariu et al, Science 2019](https://science.sciencemag.org/content/364/6437/eaav7893.abstract), and the wiki page of the v2.0 release. The following talks may also be useful: [talk1](https://youtu.be/clq50N7V_wA), [talk2](https://youtu.be/XWnc1xpIwK4).