Manifold Approximation leads to Robust Kernel Alignment
arXiv:2510.22953v3 Announce Type: replace-cross Abstract: Centered kernel alignment (CKA) is a popular metric for comparing representations, determining equivalence of networks, and conducting neuroscience research. However, CKA does not account for the underlying manifold and relies on numerous heuristics that cause it to behave differently at different scales of data. In this work, we propose Manifold-approximated Kernel Alignment (MKA), which incorporates manifold geometry (i.e., local neighb
阅读原始来源来源记录
项目分析
暂无五维分析,可先查看上方来源资料。