neuropixels
neuropixels copied to clipboard
cortex-lab
Number of clusters in Nfilt and merge_posthoc2
Hi,
Do people use the built-in post hoc merging function merge_posthoc2? Do you find that it works well?
I am getting mixed results and sometimes it is more work to split all the incorrectly merged clusters and re-merge them than just starting with fewer clusters in Nfilt and not using merge_posthoc2 at all.
Would you recommend to start with more clusters than you need e.g. 3-4 times the number of channels and then applying merge_posthoc2 OR to just use fewer clusters in Nfilt e.g. 1-2 times the number of channels and not use merge_posthoc2 at all?
In general, are there any recent developments, either in the Cortex lab or somewhere else to reduce the amount of manual curation necessary? If so, can you point me to the appropriate repository or paper? This takes up a lot of time and any improvement will be super helpful!
Many thanks,
Alex
Personally I'm not using this function - I think it has the right idea but is confounded by drifting units. I don't know which would work better between the options of using lots of clusters with merge_posthoc vs using few clusters without - instead we always use lots of clusters without.
It's definitely in everyone's interest to reduce the manual time... we're hopeful that drift correction will work at some point and will make a big difference...
Hi Nick,
Many thanks for your quick reply!
When you set the number of clusters in Nfilt, what do you find to be a good compromise between too much merging and too much splitting?
I use 960 for phase3A probes.
On Sat, Feb 3, 2018 at 1:10 PM, PhantomSpike [email protected] wrote:
Hi Nick,
Many thanks for your quick reply!
When you set the number of clusters in Nfilt, what do you find to be a good compromise between too much merging and too much splitting?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/cortex-lab/neuropixels/issues/15#issuecomment-362804916, or mute the thread https://github.com/notifications/unsubscribe-auth/AHPUP0lrXxGVKBUPnCpD_kpy2Hah0470ks5tRFqqgaJpZM4R3q6c .