Separating corruption rate from relabelling structure

Identify the separate contributions of realised corruption rate and relabelling structure to the persistence contrast between random-relabeling and Patrini asymmetric-noise conditions.

Background

The exploratory ResNet-18 comparison contrasts a random-relabeling condition with a Patrini asymmetric-relabelling condition. These arms differ simultaneously in the realised fraction of changed labels and in how labels are reassigned, so the observed persistence difference cannot be attributed uniquely to either factor. The authors explicitly state that their individual contributions remain unresolved.

References

In the exploratory ResNet-18 relabelling comparison, noise structure and realised corruption rate vary between arms, leaving their individual contributions to the persistence contrast unresolved.

— A Topological Signature of Learning under Noisy Labels  (2610.23491 - Sun et al., 16 Oct 2026) in Section 8, Scope and Limitations, p. 8