Welcome to Spruce’s documentation!¶
Spruce (2018-2020) was a project by LeafLabs, LLC and the Synthetic Neurobiology Group at MIT to gather high-quality neurophysiological data, with simultaneous extracellular and intracellular recordings, and use it to provide a ground-truth benchmark for new and existing spike sorting algorithms. Details about the data collection can be found in Brian Allen’s paper, and the raw data can be downloaded from {link TBD}. This website exists to describe the benchmarking results and some of their conclusions, as well as to provide an interactive viewer for analyzing them.
The top-line results are as follows:
MountainSort4’s performance is strongly dependent on electrode density; it requires a low density to be effective.
Kilosort and KiloSort2 tend to have a “plateau” of performance; adding more data does not improve performance, once they have enough channels to reach their peak.
KiloSort2 seems to be a “high-risk, high-reward” improvement of Kilosort; it performs worse on more difficult datasets, but better on higher-quality datasets.
The traditional spike-sorting pipeline (detection, feature extraction, clustering) is not as powerful as more sophisticated approaches. We explored using deep learning to improve traditional spike-sorting performance, and it still did not beat Kilosort.