A Topological Signature of Learning under Noisy Labels.

arXiv:2610.23491v1 Announce Type: cross Abstract: How does noisy supervision change the representations learned by a neural network? Training accuracy measures agreement with training labels but provides little information about representation geometry. We use