Fig. 3: Computing similarity scores and setting up the classifier.

a, The average weighted waveform for the example unit in Fig. 2 (black) and for a different unit (the nearest physical neighbor) from the same recording (blue). b, Centroid trajectories for the two units. The average position is shown by the shaded circle on the trajectory. c, The six similarity scores between the two units and their average, the total similarity score T. d–f, The same as a (d), b (e) and c (f), comparing the example unit (black) with the most similar unit across days (red), which was very likely to be the same neuron. g, The total similarity score for all pairs of units within days (blue squares) and across days (red squares), for an example pair of recordings, showing the first half of each recording (columns) versus the second half (rows). The data are sorted by shank, and then by depth on the shank. h, The distribution of total similarity scores in the two halves of a recording day, shown for the same units across the two halves (green) and for other neighboring units (centroid <50 μm away; blue). i, The same as h, but for units measured across days (red) after drift correction. The threshold (thresh) for putative matching (dashed line) depends on the number of units and of recordings. j, The probability densities of each similarity score, for putative matches (black) and for putative nonmatches (gray). k, Match probability P(match) computed by the naive Bayes classifier trained with the probability distributions in i. Format as in g. l, The distribution of match probabilities across two halves of the same day for same units (green) and neighbors (blue). Format as in h. m, The same as l, but for units recorded across days (red). If probability is >0.5, UnitMatch defines a pair as a match.