Standard training accuracy measures how well a model agrees with its training labels, but it says little about the internal geometry of the representations the network has learned. The new paper, posted on arXiv, argues that this geometry matters, especially when supervision is noisy. The abstract frames the question directly: how does noisy supervision change the representations learned by a neural network?

The authors propose using persistent homology, a tool from topological data analysis, to study the shape of the representation space. By tracking how topological features appear and disappear across scales, they aim to identify a signature that distinguishes learning under clean labels from learning under noisy labels. This approach shifts the focus from prediction outcomes to the structure of the learned feature space itself.

Because the abstract is truncated, the specific results are not yet available in the summary. But the framing suggests that topology could serve as a diagnostic for label noise that is independent of training accuracy, potentially offering a new way to monitor and understand robustness in supervised learning.