T-SNE visualization of large-scale neural recordings - Supplementary

George Dimitriadis1,*, Joana Neto1, Adam R. Kampff1

1Sainsbury Wellcome Centre, UCL, London, UK

Design of the 32 channel probe used to collect the PD32 data set.

Center of mass method for speeding up the t-sne process. We use two of our data sets (A: PD128, B: HD1) to check whether it was possible to pass through the t-sne algorithm only a percentage of the spikes and use the resulting 2D embedding as a template for positioning the rest of the spikes on the same plot. In both our experiments, the spikes that were placed post-t-sne using as a position the average of their 5 nearest neighbors in the high dimensional space, end up in the same place on the plot as would have been expected if they were part of the original t-sne.

Receiver Operating Characteristics (ROC) values for the different data sets used. Each point represents a unit with known spikes that appears on the t-sne plot with some spikes not belonging to it (False Positives). The blue line is the 50% True Positive Rate over False Positive Rate ratio. All the points had such small False Positive Rates that we used a semi-log plot to be able to show the relative distribution of the points on the plot.

Time (in hours) that different parts of the t-sne algorithm require to run for different combinations of features, spikes, perplexity and number of iterations. The machine was a gaming desktop with an i7 CPU, 76GB RAM and a Titan X GPU with 12GB of memory.
Per.: Perplexity, T.B. : Tree Building, L.E. : Learn Embedding
Features Spikes Seed Perp. Iter. T.B. L.E. L.E. / Iter. Total
384 128820