*Real-time* Kilosort + Neuropixels

[AtMostafa]

Is anyone trying this?

Other than this issue on real-time access to neuropixels data, I couldn't find any other resource.
How much work would it be to run (PY)kilosort in real-time, based on the templates of a few-minute-long recording just before the task starts? Has anyone tried?

Thanks for the great work,
Mostafa

[marius10p]

You would need an all new implementation of trackAndSort which does not rely on the CUDA kernels, since those need a large batch of data. There's a lot of code in there, so you'd need to work out a "lite" version of it that can run in real-time.

[pbotros]

FYI @AtMostafa we have a semi-realtime approach running, where: (1) we do a few-minute-long recording is done via OpenEphys & a Neuropixel, (2) we run Kilosort on that recording, (3) we "approximate" the fancy Kilosort sorting with a threshold-crossing & PCA sorting step, i.e. for each unit from a subset of units that we want realtime sorted, we find their "best channel", find an appropriate threshold-crossing threshold & PCA transform & PC-space template, and (4) use OpenEphys' SpikeSorter + a range of additions we had to make in our fork (https://github.com/carmenalab/plugin-GUI) to real-time spike sort. It turns out for our application, this approximation is "good enough": we end up computing firing rates with smoothing. Finally, everything is streamed out to a host of other processes & computers via a streaming library River https://github.com/pbotros/river.

Ideally we will merge a lot of the changes to OpenEphys to master since they're written to be generally applicable, but unfortunately all of it is in our fork at the moment: https://github.com/carmenalab/plugin-GUI. Some more information on our changes we made were open-ephys/plugin-GUI#372.

Feel free to reach out via email ([email protected]) if you have further questions.

[AtMostafa]

I'll have to dive deeper into (hopefully python implementation of) Kilosort's trackAndSort to see how big a problem that is. But your solution @pbotros sounds like a very good compromise. Thanks for the input :)