Core-Set Selection: Greedy k-Center Active Learning (2D)

Uncertainty sampling asks a model which points confuse it most; core-set selection asks a completely different question — which small labeled set best covers the whole feature space? This 2D companion renders the same unlabeled point cloud as the 3D version directly on its native feature plane — a pannable, zoomable scatter plot rather than an orbiting 3D stage — and runs the real greedy k-center algorithm live: each step it measures every unlabeled point's distance to the nearest already-labeled point, picks the point that is farthest from all of them, and adds it to the labeled batch. Coverage rings shrink round by round as the covering radius drops, live readouts track the exact radius and the size of the next improvement, and a random-sampling baseline — averaged over several trials at the same label budget — is computed alongside it so you can see numerically how much tighter greedy diversity sampling covers the space than picking labels blindly.