prh-cka-defined-as-normalized-hsic

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

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

CKA(K, L) = HSIC(K, L) / √(HSIC(K, K)·HSIC(L, L)) where HSIC(K, L) = (1/(n−1)²)·Trace(K̄·L̄); CKA is global, ordinal, invariant to isotropic scaling, and ranges in [0, 1].

Summary

This sets the rules for how the system scores similarity between two data representations, producing a number between 0 and 1 that reflects how aligned they are overall. The score is designed so that uniformly stretching one representation doesn't inflate the result, making it a fair, comparable metric you can trust when ranking or comparing models.