prh-cknna-interpolates-between-cka-and-mnn

IN premise — summaries/2026/08/24/huh-2024-prh-sR-references-chunk-2.md

Created 2026-08-24T17:10:56+00:00

CKNNA (Centered Kernel Nearest-Neighbor Alignment) recovers CKA as k → dim(K) (full batch) and reduces to mutual k-NN when all nearby samples are treated as equidistant, forming a continuum between the two metrics.

Summary

CKNNA acts as a single tunable dial that smoothly blends full-batch CKA (looking at the entire dataset at once) with mutual nearest-neighbor matching (looking only at local neighborhoods), so you can choose how global or local the comparison should be by adjusting one parameter. In practice, this means one metric family covers a spectrum of use cases instead of requiring separate tools for global alignment versus local structure detection.