Each tile is one configuration of the results grid in the video (dataset, number of training points, number of clusters k): 234 configurations over 16 datasets, from 100k to 100M points. The curves are test MSE on each benchmark’s official query set, averaged over seeds. The blue number is how many times fewer epochs e-means needs to reach the loss k-means has after 25 epochs: median 12.6×.
k-means: test MSE after each epoch (1 to 25)e-means trained for 1 to 5 epochs (each budget is its own run)k-means after 25 epochs◯ where e-means reaches it
pinch or ctrl + scroll to zoom · scroll or drag to pan · double-click to zoom in · hover a point to read itdrag to pan · pinch to zoom · double-tap to zoom in · tap a point to read it
e-means, clustered by e-means: Visualization of the Voronoi cells
