We have a set of Neuropixels recordings with histologically validated locations of each recording site, by _post-hoc_ track reconstruction. <br> [on DENMANLAB NAS: /s1/localization/M310016] Using this data as a training set, we have tested several classification methods (linar SVM, deep networks) with the goal of identifying the brain region each electrode is in. From the data, a multi-dimensional vector of properties at each recording site is extracted: ![schematic param][1] | Property | Band | | ----------- | ----------- | | Gamma power | LFP | | Alpha power | LFP | Delta power | LFP | | Beta power | LFP | | Event rate | AP| | Event amplitude | AP| We have a pipeline for going from path to raw data to a `pandas` DataFrame with the properties in the above table. (phil lee's notebook repo.) To date, all approaches have done this on a per channel basis, either using linear classification: @[osf](arzvu) or a [deep network][2]: @[osf](mv5ef) We would like to explore classification that takes multiple channels, and the distance between them, into account. [i.e., spatial information] <br> <br> <br> **TODO: a set of recording from more brain areas with DiI tracks, aligned into HERBS.** This project has been developed by Jeffery Ni (while at UW, Allen Institute for Brain Science), Aidan Armstrong (while at CU Anschutz), and Phil Lee (while at Dartmouth). [1]: https://files.osf.io/v1/resources/bvfj5/providers/osfstorage/6363d00f3b5b3900b6b3dc22?mode=render [2]: https://docs.google.com/presentation/d/1n0w6pWE-nA8bbMyNI6H83gfrY8j527PucjNFSAdHFEE/edit#slide=id.p