Simulated Dataset
The simulated dataset was created using 15 simulations with 4 to 6 neurons from \cite{Pedreira_2012}. This dataset has been made public \cite{Rey_2015}, and it has been used for testing the tracking ability of spike sorting algorithms \cite{Niediek_2016}. On each simulation we concatenated 70 repetitions of the simulated spikes and linearly increase/decrease the mean waveform amplitude of each class by a factor of 2.5 (i.e. increase from 1 to 2.5, or decrease from 2.5 to 1), but without changing the distance between each spike an its corresponding mean waveform (i.e. each spike gets the old mean waveform subtracted and then the new mean waveform added, with the latter being time-dependent). This way, we changed the centroid of the cluster associated with the class without affecting the high order statistics. Finally a Gaussian noise was added to each spike with a standard deviation of 10% of the maximum amplitude of the waveforms associated to the original classes (i.e. before drifting). A similar approach was used by Niediek et al. \cite{Niediek_2016}, but all the spikes were just scaled throughout the simulation to simulate the drifting, so higher order statistics were affected. An example of one of our simulations is shown in Fig. \ref{127787}.