Landmark Selection
Random per-class landmark selection was much faster and competitive with global k-means. Density-aware and random landmark selection methods worked well on CIFAR10, while k-center underperformed. The best CIFAR10 quality-runtime trade-off…
1 sources - 4 claims
Random per-class landmark selection was much faster and competitive with global k-means. Density-aware and random landmark selection methods worked well on CIFAR10, while k-center underperformed. The best CIFAR10 quality-runtime trade-off was reported for 128 to 512 landmarks per class. The paper identifies random per-class Nyström landmarks as a strong default.