Use Phy to compute template similarity in SpikeInterface

[alejoe91]

Hi @rossant

In SpikeInterface we have our own version of the compute_similarity function. @samuelgarcia suggested that it's it's probably better if we have a soft dependency on phy/phylib.

Could you point me to the correct function that I should use?

Cheers,
Alessio

[rossant]

Hi @alejoe91 , there is no good such function in phy at the moment, phy uses the similar_templates.npy matrix computed by kilosort2

[samuelgarcia]

Hi Cyril.
How do you refresh this value when you process a merge or a split inside phy ?

[rossant]

It considers the similarity between the templates from which the clusters originate from (the ones with the most spikes from each cluster). Certainly not an ideal solution and this should be improved eventually. Note that there's a plugin example showing how to implement another similarity function.

[alejoe91]

Hi @rossant

Thanks for the clarification. Do you know how KS implements such similarity? At the moment we use the correlation coefficient, but I think that's not really good as it doesn' consider differences in amplitude, for example.

As the similarity between templates is cheap to compute, maybe it would be a better option to have it centralized in phy? So you could recompute it after merges/splits... What do you think?

[rossant]

Yes that would make sense. I think the KS implementation of the similarity matrix is currently done on the GPU using a custom CUDA kernel. For the integration in phy I would prefer if we could have a pure NumPy version, don't know if that's feasible. Looping in @marius10p

[alejoe91]

For reference, here is the SI implementation (line numbers have changed from the link shared before)