Core Concepts
Decompose-and-Compose (DaC) improves robustness to correlation shift by combining elements of images.
Stats
Models trained with ERM focus on causal or non-causal parts based on confidence levels.
DaC improves worst group accuracy compared to previous methods.
Quotes
"Models trained with ERM usually highly attend to either the causal components or the components having a high spurious correlation with the label."
"DaC improves robustness to correlation shift by a compositional approach based on combining elements of images."