In the fourth category we lambed together all the cases that after splitting or merging of templates or moving spikes from one template to another ended up with acceptable SUs (based on their average spike shape, their probe heatmap, and their autocorrelogram). We arrived at 24 SUs that included a total of 38482 spikes. In these cases the combined information of where the spike lay in the embedding and in which template it belonged to made it fairly straightforward to either appropriately merge or split templates or move spikes along templates. In some cases a single SU would be represented by two separate t-sne groupings. An example of this are the two groupings embedded within blue ellipses in Figure 5. They represent a single unit after the merging of two kilosort templates (with spikes mixed over the two groupings). This unit happens to correspond to the juxtacellularly marked spikes. The unit has 4845 spikes in it. Out of those, the 4821 spikes’ timestamps correspond to the 4998 juxtacellular timestamps within a jitter of 1 ms. That translates to a False Possitive error of 0.5% and a False Negative error of 3.7%. In other cases, a single grouping would contain spikes assigned to two templates and a merging indicated a single SU (like the orange line at the bottom center of the embedding) while in others a single template would be represented by multiple groupings each defining an SU, resulting in the template’s splitting (like the two orange line groupings at the top center of the embedding). Finally the 5th category involved all the spikes that we were unable to assign to any of the previous divisions (SUs, MUs or noise). These were spikes that showed no obvious correlation between the embedding position and their template assignment. For example, the large red blob to the top and right of the large noise blob had 14184 spikes that kilosort had assigned to 95 separate templates (most of which were templates with fewer than 10 spikes each). There was no internal structure to the blob, i.e. the spikes of each template appeared randomly spread throughout. The embedding of all these spikes and templates in a single grouping made it straight forward to visuallize the situation and assign all the spikes to the unlabeld division.