Just figured out that #neuroscience has its own version of Moore's law: The number of simultaneously recorded #neurons doubles every ~7 years. This scaling has profound implications for #DataAnalysis and #modeling in #ComputationalNeuroscience. In this post, I review Stevenson & Kording's 2011 paper and reflect on its relevance today:
Figure 2 (panels a–i) from Marius Pachitariu et al. (2024) shows graph-based clustering strategies used in Kilosort4 to structure large-scale spike datasets. The figure illustrates how dense, high-dimensional spike features are iteratively reassigned and merged to obtain stable clusters from large neural populations. Panel a sketches the neighbor-based reassignment process that progressively reduces an initially large set of clusters. Panel b shows an example clustering overlaid on a t-SNE embedding of spike features. Panel c presents the hierarchical merging tree used to decide which clusters should be combined based on a modularity cost. Panel d summarizes the criteria for accepting or rejecting merges, combining feature-space bimodality with refractory-period constraints derived from spike timing. Panels e and f show the final clustering result, highlighting units that exhibit refractory periods. Panels g and h characterize the resulting units using average waveforms, autocorrelograms, cross-correlograms, and regression projections. Panel i visualizes the spatial distribution of clustered spikes along the probe. Together, the figure exemplifies how modern spike sorting algorithms impose structure on massive datasets by combining graph methods, statistical criteria, and biophysical constraints. Source: Pachitariu et al., Spike sorting with Kilosort4, 2024, Nature Methods, 914–921, DOI: 10.1038/s41592-024-02232-7ꜛ (license: CC BY 4.0)
One of its standout features is #Maxima (but also #Python, #R etc.) integration—you can create #notebooks that combine text, #LaTeX, Maxima commands, and plots, making it easy to produce scientific documents with live calculations and results.