Scoring simulated data

Following previous reports^26,36, the spike times of the sorted units and the ground truth units were matched and compared using the confusion matrix algorithm from ref. ³⁶. We set the acceptable time error between sorted spikes and ground-truth spikes at 0.1 ms. Then we counted the number of spike pairs with matching spike times, n_match, the number of unmatched spikes in the ground-truth unit, n_miss, and the number of unmatched false-positive spikes in the sorted unit, n_fp.

To assess the quality of the match between ground-truth and sorted units we adopted the Accuracy definition in ref. ²⁶:

Figure 9 shows the accuracy distribution obtained for various degrees of pooling. Sorted units with accuracy >0.8 were counted as “recovered” from the pooled signal. For each parameter set we ran the simulation three times, randomizing the noise and the spike times. Results from the three runs are reported by mean ± SD (Fig. 6b).

Units with accuracy score> 0.8 were counted as recovered.