Navigate Beyond Shortcuts: Debiased Learning through the Lens of Neural Collapse

arXiv:2405.05587v1 Announce Type: cross
Abstract: Recent studies have noted an intriguing phenomenon termed Neural Collapse, that is, when the neural networks establish the right correlation between feature spaces and the training targets, their last-layer features, together with the classifier weights, will collapse into a stable and symmetric structure. In this paper, we extend the investigation of Neural Collapse to the biased datasets with imbalanced attributes. We observe that models will easily fall into the pitfall of shortcut learning and form …

abstract arxiv beyond classifier correlation cs.cv cs.lg feature features layer lens networks neural collapse neural networks spaces studies targets through together training type will