Fig. 8: High-dimensional responses of real and artificial neurons to natural images and simulated two-dimensional activity.

a, In total, 5,000 natural images were shown to a mouse during a two-photon calcium imaging recording from V1 and higher visual areas. b, Activity from 69,957 neurons was sorted by Rastermap with splitting and binned into superneurons, plotted with the mouse’s running speed and the visual stimulus times. c, Linear receptive fields for the superneurons in b in the same order, in arbitrary units. d, AlexNet convolutional layer responses to the same 5,000 natural images sorted by Rastermap with splitting. Left: units in the convolutional layers colored by the Rastermap sorting. Right: unit activations sorted and binned into superneurons shown across stimuli. e–g, We simulated neural activity with an intrinsic dimensionality of 2 by randomly choosing an x and y value for each neuron in the range of 0 to 1 and modeling its activity as a place field. e, Left: simulated neurons are plotted at their ground truth (x,y) positions and colored by their position in the Rastermap sorting run with N_clusters = 100. Middle: same as the left panel, using Rastermap with splitting, resulting in N_clusters = 800. Right: same as the left panel, using t-SNE with multiple perplexities (P = (10, 100)) to sort the neurons. f, The k-nearest neighbor score for benchmarking embedding algorithms from ref. ²⁶: the percentage of k-nearest neighbors in the original space that are preserved as k-nearest neighbors in the embedding space—shown for Rastermap, Rastermap with splitting and t-SNE with various perplexities. g, Simulated activity sorted by Rastermap with splitting. RL, rostrolateral; AL, anterolateral; LM, lateromedial; V1, primary visual cortex; stim., stimulus; 2D, two-dimensional.