A topological signature of learning with noisy labels
In brief
- Noisy labels leave a measurable geometric signature in neural activations.
- Persistent homology reveals differences between clean and corrupted learning after 50 epochs.
- Agreement with corrupted labels stays low (7-11%) despite detectable signature.
This study examines how noisy labels affect representations learned by neural networks. Using persistent homology, researchers identify a measurable geometric signature in neural activations, even when agreement with corrupted labels remains low.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error