Finite-Sample Sensitivity of Kernel Metrics for Generative Evaluation

The paper investigates how kernel-based metrics like KID and KAD perform as two‑sample tests on real embeddings of symbolic music, audio, and images. By applying mean shifts, isotropic contractions, and rotations to one sample, the authors find that the default kernels often fail to detect rotations at 780 samples per side, while they reliably detect mean shifts. Adjusting kernel bandwidths or using alternative kernels (e.g., narrower Gaussian, Laplace) restores sensitivity, and the authors recommend reporting the sample budget and detectable changes for any reported kernel distance.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

• rag