Kilosort: realtime spike-sorting for extracellular electrophysiology with hundreds of channels

Kilosort: realtime spike-sorting for extracellular electrophysiology with hundreds of channels Pachitariu Steinmetz Kadir Carandini D

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Kilosort: realtime spike-sorting for extracellular electrophysiology with hundreds of channels 2016 10.1101/061481 Pachitariu2016-kilosort-realtime-spike-sorting-for-extracel Pachitariu2016-kilosort-realtime-spike-sorting-for-extracel.txt, the project's staged copy 50de3bb9bf3d1b55008adf109c79a18f4b762bb6 not-redistributable neuroscience

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them method Quotation withheld because this record is not cleared for redistribution. spike sorting of high-channel-count extracellular recordings Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms Replacing the full matrix convolution by a rank-3 low-rank template decomposition cuts the number of filtering operations by the ratio of channel count to template rank, which is what makes near-realtime sorting of hundreds of channels possible Quotation withheld because this record is not cleared for redistribution. mechanisms Over-split clusters arise because the fitting algorithm splits a high-variance cluster across several templates in order to reduce total reconstruction error, which is why the resulting cluster pairs form one continuous density and can be merged automatically Quotation withheld because this record is not cleared for redistribution. Manual operator curation is still needed despite the automatic merging heuristic, mainly because of electrode drift and other non-stationarities. what this limits: any claim that the pipeline is automatic: OPERATOR CURATION IS STILL REQUIRED, PRIMARILY DUE TO NON-STATIONARITIES IN THE RECORDINGS SUCH AS ELECTRODE DRIFT Quotation withheld because this record is not cleared for redistribution. The optimisation can settle in non-optimal local minima, which the authors only ameliorate by annealing several parameters. what this limits: the reproducibility of a given sort: LIKE MOST CLUSTERING ALGORITHMS, THIS MODEL IS PRONE TO NON-OPTIMAL LOCAL MINIMA Quotation withheld because this record is not cleared for redistribution. The speed comparison against the established framework was not run at the largest channel count, so no 384-channel baseline exists. what this limits: the speed comparison as a guide at the largest channel counts: the established framework NEEDED 480 MINUTES TO SORT THE 32 CHANNEL RECORDING, AND 15,000 MINUTES (10 DAYS) TO SORT THE 128 CHANNEL RECORDING, and no 384-channel comparison was run Quotation withheld because this record is not cleared for redistribution. The ground-truth spikes are the curated output of one of the two algorithms being compared. what this limits: the accuracy comparison reported here: the hybrid ground truth data set was built from all the clusters of a recording previously analysed with KlustaKwik and curated by a human expert, and KlustaKwik is the algorithm Kilosort is compared against Quotation withheld because this record is not cleared for redistribution. The accuracy benchmark was made by cropping a larger one. what this limits: the 32-channel accuracy results: the authors detail only the 32 channel benchmarks, which they created by cropping the full 128 channel hybrid benchmark Quotation withheld because this record is not cleared for redistribution. Scores are also reported after a best-possible merging step that no user could perform. what this limits: any score quoted from this figure: panels g-i repeat the results after greedy best possible merges, and a further panel gives the number of merges required to reach the best score Quotation withheld because this record is not cleared for redistribution. the denoised waveform SUBTRACTED from its original location after being added elsewhere eliminates: an artificial increase in spike density at any point on the probe - a hybrid-data test that adds spikes without removing them changes the very statistic the sorter is being judged on · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. hybrid ground truth hybrid ground truth Quotation withheld because this record is not cleared for redistribution. KlustaKwik KlustaKwik Quotation withheld because this record is not cleared for redistribution. human expert human expert Quotation withheld because this record is not cleared for redistribution. greedy best possible merges greedy best possible merges Quotation withheld because this record is not cleared for redistribution. ground truth neurons ground truth neurons Quotation withheld because this record is not cleared for redistribution. ground truth, here - MANUFACTURED, NOT OBSERVED: the paper's own term is 'hybrid ground truth', and the data set is made by taking human-curated clusters from a previous analysis, denoising their waveforms and re-inserting them elsewhere in a real recording, so the target the sorters must recover is a cleaned copy of an earlier judgement Quotation withheld because this record is not cleared for redistribution. the reference standard's provenance - THE COMPETITOR'S OUTPUT: the clusters come from a recording previously analysed with KlustaKwik, which is the algorithm Kilosort is being compared against; re-insertion at new locations makes detection a fresh problem, but the standard is still derived from one of the things measured against it Quotation withheld because this record is not cleared for redistribution. the reported score - GIVEN TWICE, ONCE AFTER AN ORACLE MERGE: the same panels are repeated after greedy best possible merges, with the number of merges needed to reach that best score shown separately, so one of the two numbers is an upper bound rather than an outcome Quotation withheld because this record is not cleared for redistribution. extends → a previous template-matching method (numbered reference 6) Quotation withheld because this record is not cleared for redistribution. extends → masked KlustaKwik (numbered reference 18) Quotation withheld because this record is not cleared for redistribution. builds-on → the hybrid ground truth benchmarking technique (numbered reference 18) Quotation withheld because this record is not cleared for redistribution. singular value decomposition singular value decomposition Quotation withheld because this record is not cleared for redistribution. k-means clustering k-means clustering Quotation withheld because this record is not cleared for redistribution. covariance matrix estimation covariance matrix estimation Quotation withheld because this record is not cleared for redistribution. Gaussian distribution Gaussian distribution Quotation withheld because this record is not cleared for redistribution. Kilosort represents the raw voltage as a sum of template waveforms triggered at spike times, which allows spatiotemporally overlapping spikes to be identified and resolved rather than missed. Quotation withheld because this record is not cleared for redistribution. The established KlustaKwik framework needed 480 minutes to sort a 32-channel recording and 15,000 minutes a 128-channel one on a CPU cluster, and Kilosort was a significant improvement on that. Quotation withheld because this record is not cleared for redistribution. On hybrid ground-truth data Kilosort yielded well-isolated units (score above 0.9) for 53% of ground-truth neurons versus 35% for KlustaKwik. Quotation withheld because this record is not cleared for redistribution. After greedy best-possible operator merges Kilosort still led, 69% of units above score 0.9 against 60% for KlustaKwik. Quotation withheld because this record is not cleared for redistribution. Kilosort's accuracy advantage came from missing fewer spikes rather than from fewer false positives, which the authors attribute to its resolution of overlapping spikes. Quotation withheld because this record is not cleared for redistribution. Fitting each spike with its own three-component spatiotemporal SVD ("private PCs") left about five times less residual waveform variance than one common per-channel PCA basis. Quotation withheld because this record is not cleared for redistribution. Kilosort's run time scales linearly with the number of recorded neurons rather than the number of channels. Quotation withheld because this record is not cleared for redistribution. Kilosort Kilosort Quotation withheld because this record is not cleared for redistribution. spike sorting spike sorting Quotation withheld because this record is not cleared for redistribution. common average referencing common average referencing Quotation withheld because this record is not cleared for redistribution. zero-phase component analysis (ZCA) zero-phase component analysis (ZCA) Quotation withheld because this record is not cleared for redistribution. hybrid ground truth hybrid ground truth Quotation withheld because this record is not cleared for redistribution. private PCs private PCs Quotation withheld because this record is not cleared for redistribution. run time run time Quotation withheld because this record is not cleared for redistribution. spike density spike density Quotation withheld because this record is not cleared for redistribution. miss rate miss rate Quotation withheld because this record is not cleared for redistribution. false positive rate false positive rate Quotation withheld because this record is not cleared for redistribution. 10, 29 and 140 minutes — Kilosort run times for 32-, 128- and 384-channel recordings on a single GTX 980Ti workstation Quotation withheld because this record is not cleared for redistribution. 480 minutes / 15,000 minutes (10 days) — KlustaKwik run times for the 32-channel and 128-channel recordings, the comparison baseline Quotation withheld because this record is not cleared for redistribution. 53% vs 35% — share of ground-truth neurons sorted with a score above 0.9 by Kilosort versus KlustaKwik before manual merges Quotation withheld because this record is not cleared for redistribution. 69% vs 60% — share of units scoring above 0.9 for Kilosort versus KlustaKwik after best possible merges Quotation withheld because this record is not cleared for redistribution. ∼5 times — reduction in residual waveform variance from per-spike private PCs relative to a single per-channel PCA Quotation withheld because this record is not cleared for redistribution. 16 — number of hybrid ground-truth dataset instantiations on which both algorithms were run Quotation withheld because this record is not cleared for redistribution. Nchan/Nrank — factor by which the low-rank template decomposition reduces filtering operations, with channel count typically above 100 and template rank typically 3 Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1101/061481.