核心概念
Decompose-and-Compose (DaC) proposes a compositional approach to improve robustness to correlation shift by combining elements of images.
統計
"DFR retrains the last layer of a model previously trained by ERM with group-balanced data to make it robust to spurious correlation."
"MaskTune fine-tunes the model for one epoch using a masked version of the training data to focus on core features."
"Group DRO minimizes the worst-case loss across groups with strong regularization."
引用
"Models trained with ERM usually highly attend to either the causal components or the components having a high spurious correlation with the label."
"Our proposed method performs better than previous baselines on well-known benchmarks in the literature."