Researchers propose Manifold-approximated Kernel Alignment to fix CKA scale issues
An arXiv paper introduces MKA, a representation-comparison metric that adds manifold geometry to kernel alignment and reports steadier behavior across data scales than CKA.
In plain words
Researchers often compare how two neural networks represent data using a metric called centered kernel alignment (CKA). The paper's authors say CKA ignores the shape of the data and behaves inconsistently at different scales, so they built MKA, which accounts for local neighborhood structure. In their tests, MKA was steadier, though not better everywhere.
What was announced
Centered kernel alignment is widely used to compare representations, judge whether networks are equivalent, and support neuroscience research, but the authors write that it does not account for the underlying manifold and depends on heuristics that make it behave differently at different data scales. Their proposed metric, Manifold-approximated Kernel Alignment, incorporates manifold geometry, meaning local neighborhood structure, into the alignment task. The paper derives a theoretical framework, removes representation-dependent degree weighting, produces a closed-form alignment score, and gives each row a probabilistic interpretation as a local Markov transition distribution. The authors also report kernel- and score-level stability with respect to neighborhood size.
Empirically, the authors evaluated MKA on synthetic datasets and real-world examples. In controlled synthetic experiments, they report that MKA was consistently less sensitive to hyperparameters and captured the selected topology-preserving relationships. On their large-scale benchmark, MKA showed its strongest aggregate performance in vision, was competitive on graphs, and was weaker than CKA in several NLP evaluations. The authors suggest manifold-aware kernel alignment could be useful when local neighborhood geometry matters, with possible applications in representation learning. These results come from the authors' own evaluations and have not been independently confirmed.
What it does not show
- The evidence is the paper's abstract only; full method details, experimental setup, and complete results were not available for verification.
- All performance findings, including MKA's advantages and its weaknesses in NLP, are the authors' own reported results and have not been independently confirmed.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.