Online Spectral Classification for Long-Term Spike Sorting

In this paper, we explore online spike sorting for chronic extracellular recording. A challenge is to track drifts of clusters across time. In our proposed approach, time-series data of neural spikes are batched into sliding time frames, and a Laplacian eigenproblem corresponding to each time frame is efficiently solved using iterative methods with an initial guess provided by the previous time frame. The labels are obtained by comparing the eigenvectors of successive time frames. This approach is capable of classifying data points from time-varying but locally stationary distributions.