Benchmarking Instance-Dependent Label Noise with Controlled Corruptions

In three linesCILN is a benchmark framework for instance-dependent label noise (IDN) generation through controlled input corruptions rather than imperfect annotators. 90 configurations tested on CIFAR-10, MNIST, and Adult show that noise structure, not just noise rate, affects benchmark difficulty and exposes failure modes in Co-Teaching and DivideMix.

Summary generated by Claude — human-verified